Automatic Histogram Specification for Glioma Grading Using Multicenter Data

Xi Chen; Yaping Wu; Guohua Zhao; Meiyun Wang; Wenyi Gao; Qian Zhang; Yusong Lin

doi:10.1155/2019/9414937

Automatic Histogram Specification for Glioma Grading Using Multicenter Data

Chen, Xi;Wu, Yaping;Zhao, Guohua;Wang, Meiyun;Gao, Wenyi;Zhang, Qian;Lin, Yusong 2019-12-19 00:00:00 Hindawi Journal of Healthcare Engineering Volume 2019, Article ID 9414937, 12 pages https://doi.org/10.1155/2019/9414937 Research Article Automatic Histogram Specification for Glioma Grading Using Multicenter Data 1,2 3 2 3 2 4 Xi Chen, Yaping Wu, Guohua Zhao, Meiyun Wang , Wenyi Gao, Qian Zhang, 1,2 and Yusong Lin School of Software, Zhengzhou University, Zhengzhou, Henan 450002, China Collaborative Innovation Center for Internet Healthcare, Zhengzhou University, Zhengzhou, Henan 450052, China Department of Radiology, Henan Provincial People’s Hospital, Zhengzhou, Henan 450003, China School of Computer Science, Zhongyuan University of Technology, Zhengzhou 450000, China Correspondence should be addressed to Yusong Lin; yslin@ha.edu.cn Received 22 July 2019; Revised 6 November 2019; Accepted 23 November 2019; Published 19 December 2019 Academic Editor: Emiliano Schena Copyright©2019XiChenetal.)isisanopenaccessarticledistributedundertheCreativeCommonsAttributionLicense,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Multicenter sharing is an eﬀective method to increase the data size for glioma research, but the data inconsistency among diﬀerent institutions hindered the eﬃciency. )is paper proposes a histogram speciﬁcation with automatic selection of reference frames for magnetic resonance images to alleviate this problem (HSASR). )e selection of reference frames is automatically performed by an optimized grid search strategy with coarse and ﬁne search. )e search range is ﬁrstly narrowed by coarse search of intraglioma samples, and then the suitable reference frame in histogram is selected by ﬁne search within the sample selected by coarse search. Validation experiments are conducted on two datasets GliomaHPPH2018 and BraTS2017 to perform glioma grading. )e results demonstrate the high performance of the proposed method. On the mixed dataset, the average AUC, accuracy, sensitivity, and speciﬁcity are 0.9786, 94.13%, 94.64%, and 93.00%, respectively. It is about 15% higher on all indicators compared with those without HSASR and has a slight advantage over the result of a manually selected reference frame by radiologists. Results show that our methods can eﬀectively alleviate multicenter data inconsistencies and lift the performance of the prediction model. accurately display the location and size, and it correlates well 1. Introduction with histological characteristics. In medical research, high quality data is diﬃcult to obtain in a single institution, so it Glioma is a prevalent fatal brain disease and the most malignant, which accounts for approximately 24.7% of all needs to be shared through multicenter. However, the primary brain and other central nervous system tumors and diﬀerence of multicentre data is a serious challenge. 74.6% of malignant tumors [1]. )e World Health Orga- In the acquisitionof multicenter gliomaMRI data,due to nization’s guidelines for glioma diagnosis and treatment are diﬀerences in acquisition equipment and parameters result divided into four levels, namely, I–II and III–IV for low- signiﬁcant diﬀerences in data samples in terms of speciﬁ- grade glioma (LGG) and high-grade glioma (HGG) [2]. In cation, size, contrast, and brightness. Diﬀerences among the clinical applications, biological behavior, treatment options, data lead to deviations in the grading eﬀect of glioma [4]. and prognoses of patients with glioma of diﬀerent grades are Figure 1 shows the diﬀerences among multicenter data. clearly diﬀerent. )erefore, the accurate preoperation Figure 1(a) shows that some of the data contain skulls and grading of Glioma is important. Magnetic Resonance Im- the others do not contain. Figure 1(b) shows that these data aging (MRI) is characterized by multidirectional tomogra- have diﬀerent scales and number of slices. A previous study phy and multiparameter high-resolution soft-tissue imaging indicated that pixel size and slice thickness remarkably aﬀect and is widely used to evaluate the tumor heterogeneity [3]. space and strength characteristics of the calculation [5, 6]. MRI is commonly used in glioma grading because it can )erefore, voxel size must be uniﬁed to reduce the error in 2 Journal of Healthcare Engineering )e main contributions of this study are as follows: (1) )is paper proposes a histogram speciﬁcation with automatic selection of reference frames (HSASR). (a) )is method can automatically select the suitable reference frame of histogram instead of radiologists during image enhancement, so as to enhance the consistency of brain tumors image contrast. (2) )is paper proposes a set of image standardization (b) algorithms to make the preprocessed multicenter data have better consistency, improve the adapt- 512 × 512 × 18 320 × 320 × 18 240 × 240 × 155 ability of data, and improve the accuracy of the glioma prediction model. 2. Related Works C1 C2 C3 (c) In recent years, researchers have gradually adopted multi- center data to replace single institution in clinical medical research. However, multicenter data also faced many C4 C5 C6 challenges. Berenguer et al. [14] carried out a test-retest phantom study of individual image acquisition parameters. Figure 1: Multicenter data. (a) Space of dataset is uneven. (b) )e impact of image parameters on the image cannot be Volume pixel of data element is not uniform. (c) Dataset contrast analyzed because of the diﬀerence of the model or machine brightness is not uniform, and C1–C6 are randomly selected data manufacturer. )erefore, variations because of the use of samples. images were eliminated by exclusively scanning phantoms. In addition, Hugo et al. [15] indicated that multicenter structure MRI studies have stronger statistical eﬃcacy than calculating characteristics. Figure 1(c) presents six images with diﬀerent contrasts. )ese diﬀerences are due to de- single-institution studies. However, central diﬀerences in viation in acquisition equipment and parameters. contrast sensitivity and spatial uniformity lead to diﬀerences In order to alleviate the inconsistency of contrast of in tissue classiﬁcation or image registration that may reduce multicenter data, it is necessary to enhance the image data or completely oﬀset the enhanced statistical eﬃcacy of appropriately. Histogram correction is a commonly used multicenter data. )erefore, maintaining data in a standard technique in image enhancement [7]. Among them, histo- environment is important. Nyul et al. [16] mentioned some gram equalization (HE) and normalization techniques [8, 9] problems in the original MRI scale preprocessing method are often used to adjust the overall brightness of images. )is and attempted to use the median value and other percentile type of methods mainly expands the range of gray value in numerical methods, such as landmark, to solve these the histogram without adjusting its intensity. )ese methods problems; they obtained robust results after preprocessing. )e validity of the standardized new landmark was also only enhance the overall image and improve its brightness, but the contrast between tumors and other tissues is not mentioned in the study, after which the image brightness considerably enhanced [10, 11]. Histogram speciﬁcation level and contrast consistency were signiﬁcantly improved. (HS) can change the frequency of grayscale values and Bakas et al. [17, 18] focused on the data preprocessing enhance any local brightness [12, 13], but it mainly adjusts method, which has a certain enlightening eﬀect on the the local brightness according to the reference frame. In preprocessing method of this study. general, the reference frame is selected by a radiologist. In preprocessing, contrast enhancement of multicenter However, this process consumes a considerable amount of data image is an important challenge. It was shown that a time and the ﬁnal reference frame may not be the best high-contrast medical image could lead to a better in- choice. )erefore, it is an urgent problem to automatically terpretation of the diﬀerent adjacent tissues in the imaged search for the best reference frame of histogram instead of body part [19, 20]. Accordingly, the resulting enhanced searching by a radiologist. image, which is in terms of signal intensities of diﬀerent )is paper presents a histogram speciﬁcation with au- tissues, can facilitate the automated segmentation, feature tomatic selection of reference frames for magnetic resonance extraction, and classiﬁcation of these tissues. Existing image images, called HSASR, speciﬁcally used to replace radiolo- enhancement techniques (empirical or heuristic) are re- gists manually selecting reference frames. )is method can markably related to a particular image and usually aimed at enhance the consistency of tumor intensity in the whole improving image contrast. However, no uniﬁed standard is medical image. In this work, we apply HSASR to the pre- available to measure the quality eﬀect of image enhancement processing of multicenter glioma data, and the eﬀectiveness [21–24]. of this method is proved by Glioma grading experiment. )e Image enhancement can be divided into two categories: results showed that this method had certain practical value spatial domain method and frequency domain method. in glioma grading research. Image enhancement in spatial domain usually corrects Journal of Healthcare Engineering 3 histogram. Histogram equalization is a commonly used N (1) P (l) � , ∀l ∈ L, global grayscale image enhancement technique, and the grayscale value is uniformly redistributed based on the where N � H × W and N is the number of pixels with a gray cumulative density function of the histogram [25]. However, level. )e cumulative distribution function (CDF) of P is equalization fails to consider the intensity of the grayscale r presented as follows: value. )e average brightness of the image can make the x x regions with large original intensity fade, whereas dark areas N s (x) � 􏽘 P (j) � 􏽘 , ∀x ∈ L. (2) r r will brighten after image equalization. Sengee et al. [26] j�0 j�0 suggested an extension method of BBHE based on the neighbourhood metric. )is method involved a few steps: Similarly, if the speciﬁed reference image is z, then the ﬁrst, a large histogram was divided into the subregion using gray cumulative distribution function is the neighbourhood metric and process independently. )is method results in boundary noise and possibly uneven v (y) � 􏽘 P (i) � 􏽘 , ∀y ∈ L, (3) z z histogram brightness in two parts. In addition, MedGA [24] M i�0 i�0 wasan enhancementmethodthatHEcombinedwithgenetic algorithm to directly improve the histogram frequency of where Mand N arethe numberofpixelswith graylevel L.HS attempts to obtain transformation function y � F(x) and images and has achieved obvious results. However, this method could only be limited to the presence of two maps gray level x in the original image to gray level y, such that the transformed image can have a histogram similar to grayscale regional tissues and not enhance complex brain tumors. the reference histogram. To preserve the inherent in- formation of the original image, function F should be a monotonically increasing function. )is function can be 3. Materials and Methods obtained by using the following equation: 3.1. Data Collection. Glioma data used in the study were v (y) � s (x). (4) z r obtained from two separate sources, BraTS2017 and Glio- )erefore, the gray level of the map can be obtained by maHPPH2018 datasets. )e BraTS2017 dataset came from a variety of scanning instruments from 19 medical in- using stitutions. BraTS2017 dataset includes 210 HGG data and 75 − 1 (5) y � v 􏼂s (x)􏼃, z r LGG data; the data was already segmented when it was − 1 acquired. GliomaHPPH2018 dataset was obtained from where v is the inverse of v . multiple equipment of diﬀerent models from multiple In the discrete case, the inverse function usually does not manufacturers, including 4 Siemens equipment (1 1.5t exist. )e inverse function is usually replaced by the best equipment, 3 3T equipment) and 4GE equipment (2 1.5t objective function to approximate the y of a particular gray equipment, 2 3T equipment), as shown in Table 1. )e level x as follows: 􏼌 􏼌 magnetic resonance data collected on these devices are 􏼌 􏼌 􏼌 􏼌 y � argmin 􏼌 s (x) − v (k)􏼌. r z (6) diﬀerent from each other. After data inspection, there are up to 186 default inspection protocols for the devices, and Equation (6) represents the absolute value of the dif- contrast-enhanced T1-weighted imaging (CET1) was used in ference between the original image’s cumulative histogram this study, including 35 CET1 sequences (see Table 2), layer and gray level functions of the speciﬁed reference histogram, thickness of 5mm to 6.5mm, repetition time of 220– and then the minimum value is selected as the y value. With 1970ms, echo time of 2.46–28.60ms, and resolution of this transformation rule, each gray level x can be mapped to 256 ×256 ×18, 320 ×290 ×18, 320 ×320 ×18, 384 ×384 ×18, y. )erefore, the mapped image will be similar to the desired 448 ×408 ×18, and 512 ×512 ×18. )e GliomaHPPH2018 histogram. dataset includes 161HGG data and 77 LGG data. A three- )e reference frame in the histogram is usually selected dimensional region of interest (ROI) for all the tumors in by the doctor. However, the selected reference image may GliomaHPPH2018 was depicted by two senior radiologists not be suitable, thereby consuming a considerable amount of in Henan provincial people’s hospital. )e ROIs were seg- the radiologist’s time. To solve this problem, an optimized mented slice-by-slice on the axial plane using the CET1 grid search strategy with coarse and ﬁne search is proposed sequences. After discussion between the two radiologists, the to select the suitable reference frame automatically. ﬁnal decision is made to select the optimal segmentation ﬁle. 3.3. HSASR. In this study, an improved grid search method 3.2. Histogram Speciﬁcation. HS, an eﬀective image en- is proposed to solve the problem of selecting a speciﬁed hancement technique, is an extension of histogram equal- reference image. )e grid search is generally used to divide ization that can eﬀectively alleviate the problems of grids of the same length in a certain spatial range based on histogram equalization. Let r � 􏽮r 􏽯be an H × W discrete ij the proposed coordinate system. Each coordinate point input digital image with L gray levels, and let L �{0, 1, . . ., represents a parameter, and these searched parameters are L − 1}. )e histogram or gray level probability density of an calculated and analysed based on the step size. Finally, image is deﬁned as follows: 4 Journal of Healthcare Engineering Table 1: GliomaHPPH2018MR acquisition equipment list. Number Manufacturer Model Magnetic ﬁeld strength Amount 1 GE OPTIMA MR360 1.5T 1 2 GE Signa HDxt 1.5T 1 3 GE DISCOVERY MR750 3T 2 4 SIEMENS Sempra 1.5T 1 5 SIEMENS Prisma 3T 1 6 SIEMENS TrioTim 3T 1 7 SIEMENS Verio 3T 1 Note: amount is for number of machines. Table 2: CET1 sequence collection parameter statistics. Number Manufacturer Magnetic ﬁeld Model D-spacing Layer thickness Resolution Voxel size 1 GE 1.5 OPTIMA MR360 6.5 5.5 512 ×512 ×18 0.4688\0.4688 2 GE 1.5 OPTIMA MR360 7 6 512 ×512 ×18 0.4688\0.4688 3 GE 1.5 OPTIMA MR360 7.5 6 256 ×256 ×18 0.9375\0.9375 4 GE 1.5 OPTIMA MR360 7.5 6 512 ×512 ×18 0.4688\0.4688 5 GE 1.5 OPTIMA MR360 8 6 512 ×512 ×18 0.4688\0.4688 6 GE 1.5 Signa HDxt 7 6 512 ×512 ×18 0.4688\0.4688 7 GE 1.5 Signa HDxt 7 6 512 ×512 ×18 0.5273\0.5273 8 GE 1.5 Signa HDxt 7.5 6 512 ×512 ×18 0.4688\0.4688 9 GE 1.5 Signa HDxt 7.5 6.5 512 ×512 ×18 0.4688\0.4688 10 GE 1.5 Signa HDxt 8 6 512 ×512 ×18 0.4688\0.4688 11 GE 3 DISCOVERY MR750 6.5 5 512 ×512 ×18 0.4688\0.4688 12 GE 3 DISCOVERY MR750 7 6 512 ×512 ×18 0.4688\0.4688 13 GE 3 DISCOVERY MR750 7.5 6 512 ×512 ×18 0.4688\0.4688 14 SIEMENS 1.5 Sempra 7.2 6 448 ×408 ×18 0.5134\0.5134 15 SIEMENS 3 Prisma 7.2 6 320 ×320 ×18 0.7188\0.7188 16 SIEMENS 3 Prisma 7.2 6 384 ×384 ×18 0.5989\0.5989 17 SIEMENS 3 Prisma 7.8 6 320 ×320 ×18 0.7188\0.7188 18 SIEMENS 3 TrioTim 6.6 6 320 ×320 ×18 0.7188\0.7188 19 SIEMENS 3 TrioTim 6.6 6 320 ×320 ×18 0.75\0.75 20 SIEMENS 3 TrioTim 7.2 6 256 ×256 ×18 0.8984\0.8984 21 SIEMENS 3 TrioTim 7.2 6 256 ×256 ×18 0.9375\0.9375 22 SIEMENS 3 TrioTim 7.2 6 320 ×290 ×18 0.7188\0.7188 23 SIEMENS 3 TrioTim 7.2 6 320 ×320 ×18 0.7188\0.7188 24 SIEMENS 3 TrioTim 7.2 6 320 ×320 ×18 0.75\0.75 optimal parameters are considered the output. Given that randomly selected for histogram speciﬁcation (the this method must traverse all the corresponding points in same sample is selected each time), and then the the grid, many unnecessary invalid calculations are gener- performance of data after histogram speciﬁcation is tested. ated, resulting in an exponential increase in time. To reduce time consumption, this study improves the (3) Select the sets of all the best area under the curve directional grid optimization search based on data charac- (AUC) values according to threshold Tand perform teristics. On the basis of the characteristics of glioma itself, the next ﬁne search. the method is improved by using coarse segmentation and (4) Fine search: in the optimal cluster, the search is subdivision. Horizontal and vertical searches are used for initiated from the middle section (two pointers are coarse segmentation and subdivision, respectively. )e set to point to the middle section), and pointer 1 speciﬁc steps of the improved method are as follows: searches upward successively. If reference frames (1) )e number of clusters N and the number of slices K do not contain tumors, then the search is stopped. of each tumor are determined. Each cluster is set to i, If the diﬀerence between the current and previous and the initialization step size is K/2. A 2D grid is values is greater than T, then the search is stopped, established by using N and K, and the grid nodes are and the optimal value is selected. At the same time, the corresponding reference slices of N and K. similar to the previous step, pointer 2 is searched downward. (2) In the coarse search, each time it searched by step size from the beginning of each set(horizontally), the (5) )e optimal reference slice map is the ﬁnal output. If median slice in each cluster was selected as the multiple optimal values exist, then selection is made reference frame. )en, 30% of data in all datasets are based on accuracy and speciﬁcity indicators. If Journal of Healthcare Engineering 5 (i) Input: Mixed dataset F, data sample number N, number of each sample slice K (ii) Output: Optimal reference slice F (1) Step1: Coarse search (2) Function HSBM (n, F1, l )//Histogram speciﬁcation based on brain MRI (3) { (4) F1 �Random (range (0, N), F∗30%)//Random selection of 30% of dataset (5) For each n ∈ F do (6) For each n1 ∈ F1 do (7) cdf � hist · cumsum( ) F F 1 n1 n (8) cdf � hist · cumsum( ) F F n n k/2 min (9) Calculate: hist � arg |(cdf /shape(F )) − ((cdf /shape(F ))(m))| F n Fn1 n1 n k/2 k/2 (10) End for (11) Calculate: Auc(n ) � roc auc score(Lable − Lable ) K/2 original predict (12) End for (13) } (14) F2 � Sort max(Auc ) k/2 (15) F2 �Max(F2) (16) For each c ∈ F2 do (17) if |F2 − Auc(c)|≤ T then (18) Select c ⟶ F3 (19) end (20) End for (21) Select the optimal set F3 (22) Step 2 Fine search (23) Initializes Pointers l , l 1 2 (24) For each n ∈ F3//Search both ends (25) l � k/2 (26) For l ≤ kandexit(tumor ); l + +do 1 l ++ 1 (27) HSBM (n, F1, l ) (28) End for (29) End for (30) For each n ∈ F3 (31) l � k/2 (32) For l ≤ kandexit(tumor ); l + +do 2 l ++ 2 (33) HSBM (n, F1, l ) (34) End for (35) End for (36) Output: select the ﬁrst optimal slice F ALGORITHM 1: HSASR algorithm pseudocode description. several optimal reference frames existed, then the 3.4. Experiments. Figure 3 shows the overall process of ﬁrst optimal reference frame is selected as the ﬁnal multicenter data grading prediction. First, data are im- reference frame. ported, data preprocessing is performed, features are extracted by using the feature calculation method, features Algorithm 1 describes the HSASR algorithm. )e pa- are selected, and model training is ﬁnally conducted. rameter T in Step 1 prevents the collection of additional Contribution points of this study are in the preprocessing optimal values from being missed. )e T in Step 2 saves stage. )e HSASR method proposed is the most important search time. Figure 2 presents the 2D diagram of the im- part of preprocessing. proved optimization grid search. In this study, N represents the number of data samples, K denotes the number of slices of each sample, and K/2 is the middle slice of the sample. 3.5. Preprocessing. Storage formats of multicenter data are First, a horizontal coarse search is performed. )e pale blue often inconsistent. Data formats of this study include NIfTI circles indicate the performance values of intermediate slices and DICOM format ﬁles. )erefore, to standardize multi- in each cluster. )e optimal clusters are selected for lon- center data formats, this study adopts the format conversion gitudinal ﬁne search, the middle slice is considered the initial method based on the Convert3D tool to unify all data sample value, and both ends are searched simultaneously. )e formats into the NIfTI format. )e majority of patients in vertical circles and dark blue squares represent the perfor- the GliomaHPPH2018 dataset have 18 DICOM ﬁles with mance values in the optimal set and the selected ﬁnal ref- CET1 sequence, and a small number of patients have 36 ﬁles. erence frame, respectively. After format conversion, all slices of each patient’s CET1 6 Journal of Healthcare Engineering k/2 Intermediate slice as reference Optimal slice as reference Single sample slice as reference Figure 2: Diagram of the optimized grid search (note: k/2 represents the middle slice of the cluster). Multicenter standardization Feature engineering Feature calculation Feature selection Grading prediction Preprocessing Raw data BraTS2017 Feature Set1 Format conversion Feature selection diagram PCA Brain tissues extraction Pyradiomics Feature Set2 Feature samples Resampling x x GliomaHPPH2018 Selection feature Predictive analysis Figure 3: Flowchart of multicenter data grading prediction. sequence are converted into the NIfTI format ﬁle, thereby In addition, spatial consistency of the data image is maintained.Multiple voxelrangesoccur inboth datasets.On providing convenience in subsequently unifying data processing. the basis of Convert3D, we adopt the shortest distance in- Furthermore, to unify the multicenter data in pre- terpolation resampling technology to carry out scale processing, maintaining the consistency of brain tissue resampling in each data sample. )e size of each data sample structureis alsoimportant.)e GliomaHPPH2018 datasetin is uniﬁed to 240 ×240 ×155, and the label sample is also this study contained skull images, which are removed at the converted into a ﬁle of the same size. In addition, the dis- time of acquisitionin the BraTS2017dataset. )issituation is tance between uniﬁed slices is 1mm, and the original po- not only a certain impact on HS but also introduced diﬃ- sition was [0, − 239, 0]. culties in combining the two datasets due to the gap between To solve the problem of inconsistent brightness and them. In this study, the FSL tool is ﬁrstly used in location contrast of multicenter data, this paper proposes a histogram registration, and then the skulls are removed based on the speciﬁcation withautomatic selection ofreferenceframes for bet script. In this study, 238 data skulls are removed, and magnetic resonance images. On the basis of the character- brain tissues and regions of interest are relatively intact. istics of glioma itself, this paper puts forward an optimized Journal of Healthcare Engineering 7 MRI without HSASR MRI with HSASR L1 L2 L3 L1 L2 L3 L4 L5 L6 L4 L5 L6 H1 H2 H3 H1 H2 H3 H4 H5 H6 H4 H5 H6 Figure 4: MRI comparison map before and after HSASR. Table 3: Results of the ﬁrst three glioma grading experiments. Without HSASR With HSASR Training Testing AUC ACC (%) SEN (%) SPE (%) AUC ACC (%) SEN (%) SPE (%) GH BT 0.6850 59.93 58.45 64.00 0.9507 90.88 90.18 82.67 BT GH 0.7605 72.68 81.98 53.25 0.9146 85.90 89.87 77.63 (80%) (GH+BT)+(20%) 0.8252 78.17 85.42 61.67 0.9786 94.13 94.64 93.00 (GH+BT) (80%)GH+(20%)GH 0.8394 81.25 82.06 76.32 0.9556 89.36 87.20 94.71 (80%)BT+(20%)BT 0.8512 79.12 80.19 74.99 0.9934 95.61 97.35 92.65 Note: GH is for GliomaHPPH2018, and BTstands for BraTS2017. BTrepresents 80% of the data as training and 20% as testing. GH alone represents 80% of GH data as training and 20% of GH as testing. BT alone represents 80% of BT dataset as training and 20% of BT as testing. grid search strategy with coarse and ﬁne search. First, in search, the step size is changed to 1, data are searched from coarse search, all data samples are selected as the object of the middle number upward and downward, and the search is reference selection in this study. Each ﬁle is considered a over according to the end condition. collection, and the step size is set to the middle number of the collection (78), starting from the beginning of every collection each time. )e middle slice (78) of each collection 3.6. Feature Engineering. All the data being processed are imported into feature calculation. For each region of in- is used as the reference slice of HS, 30% of image samples are randomly selected from all data, and image enhancement is terest, 557 radiomics features are calculated through Pyr- adiomics (Pyradiomics is a tool for the computational carried out via HS and subsequently imported into the model veriﬁcation process of this study. T is set to 1%, and characteristics of medical images). In this study, 9 spatial geometric features are included, which are only calculated in the set with the highest AUC value and AUC value error less the original space. Eighteen ﬁrst-order statistical features than 1% is selected. Furthermore, the study carries out ﬁne 8 Journal of Healthcare Engineering Receiver operating characteristic curve Receiver operating characteristic curve 1.0 1.0 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 False positive rate False positive rate Train ROC AUC = 0.98571 Train ROC AUC = 0.95225 Test ROC AUC = 0.91456 Test ROC AUC = 0.95073 (a) (b) Figure 5: ROC plot. Table 4: Results of the contrast experiment. Training Testing Method AUC ACC (%) SEN (%) SPE (%) (80%) (GH+BT)+(20%) (GH+BT) HE 0.9176 86.19 90.44 76.05 (80%) (GH+BT)+(20%) (GH+BT) Handpicked 0.9447 89.90 90.35 88.88 (80%) (GH+BT)+(20%) (GH+BT) Our method 0.9786 94.13 94.64 93.00 Note: GH is for GliomaHPPH2018, and BT is for BraTS2017. GH+BT represents 80% of the data as training and 20% as testing. and 43 texture features are calculated in the original data and 3.7. Experimental Comparison Designs. To verify the eﬀec- eight wavelet decomposition spaces, respectively. For all tiveness of processed multicenter data, this study proposed image histological features, model training is conducted the following process designs: directly after the processing of missing values. (1) Take the processed GliomaHPPH2018 and )e calculated features are imported into the model BraTS2017 datasets as the training and testing sets, scriptof this studyfor prediction.)e modelscript isdivided respectively. )en, use the two unpreprocessed into two parts, namely, feature selection and classiﬁer. Five datasets as contrast experiments. selection methods and nine classiﬁers are combined to set up (2) Mix together 521 data samples from Glio- the classiﬁer, and the training and testing sets are imported maHPPH2018 and BraTS2017, preprocess the data, into the model for prediction. For the designs of experiment carry out feature selection and model training, and 2 and 3 in the next section, 80% and 20% of data are selected analyze the results. from the data table as training and testing sets, respectively. )e prediction is then repeated 10 times. )e average value (3) Use single sequences from the processed and un- of ﬁnal results is obtained, and the classiﬁer and selection processed GliomaHPPH2018 and BraTS2017 as methods with the best results are ﬁnally selected with the comparison experiment. rating device. )e selection methods used in the study are (4) Make comparisonexperiments between this research SelectKBest (f_classif) for classifying the tag features be- method and common image enhancement methods tween tasks for ANOVA f values, principal component on the mixed dataset. Finally, compare these results analysis (PCA), kernel principal component analysis with the untreated data. In addition, to verify the (KPCA), independent component correlation algorithm eﬀectiveness of HSASR, compare the research results (ICA), and factor analysis (FA). )e classiﬁers used in this with the reference diagram results selected by study are decision tree (DT), random forest (RF), bagging radiologists. (BAG), binary search tree (BSA), naive Bayes (NB), mul- tilayer perception (MLP), support vector machine (SVM), 4. Results and Discussion logistic (LR), and k-nearest neighbour (KNN). For the two separate datasets in experiment 1, the stable training model 4.1. Analysis of Image Enhancement. )e purpose of weak- is selected by using a tenfold cross-validation scheme, and ening the multicenter data is to relieve the diﬀerences between then the grading prediction is ﬁnally performed. groups and retain individual characteristics. )at is to say, True positive rate True positive rate Journal of Healthcare Engineering 9 Table 5: Performance comparison with similar works. Author and reference Architecture/method Output class Data Maximum ACC (%) Khawaldeh et al. [27] AlexNet+Preprocessing 3 130 91.16 Chen et al. [28] CAD+Preprocessing 2 274 91.27 Our method HSASR 2 523 94.13 80 80 22 9 2 26 60 15 3 58 73 LGG HGG LGG HGG Predicted lable Predicted lable (a) (b) Figure 6: Confusion matrix of mixed dataset. 0.84 0.83 0.77 0.82 FILT 0.72 0.82 0.82 0.81 0.77 0.79 0.80 PCA 0.81 0.81 0.71 0.83 0.77 0.78 0.69 0.8 0.8 0.76 KPCA 0.67 0.78 0.76 0.81 0.81 0.73 0.84 0.74 0.81 0.72 ICA 0.69 0.8 0.79 0.82 0.69 0.79 0.76 0.75 0.8 FA 0.79 0.79 0.81 0.72 0.81 0.69 0.83 0.76 0.8 0.68 DT RF BAG BST NB MLP SVM LR KNN Classifier Figure 7: Heat map without HSASR. Feature reduction True lable HGG LGG True lable HGG LGG AUC score 10 Journal of Healthcare Engineering FILT 0.9 0.97 0.97 0.97 0.95 0.98 0.98 0.96 0.97 0.96 0.85 0.94 0.9 0.98 0.95 0.97 PCA 0.96 0.97 0.95 0.93 0.87 0.96 0.93 0.97 0.97 0.96 0.97 KPCA 0.94 0.96 0.90 ICA 0.84 0.96 0.95 0.96 0.9 0.97 0.97 0.96 0.96 0.87 FA 0.85 0.95 0.94 0.96 0.88 0.97 0.96 0.97 0.95 0.84 DT RF BAG BST NB MLP SVM LR KNN Classiﬁer Figure 8: Heat map with HSASR. Experiment 1 in Table 3 presents the results of Glio- under the premise of retaining tumor morphological char- acteristics,theimagedataofdiﬀerentmedicalinstitutionshave maHPPH2018and BraTS2017 datasets as the training and similar contrast and brightness. In this study, contrast and testing set, respectively. )e results show that the data with brightness of multicenter glioma data images are signiﬁcantly HSASR is better predicted, and the indexes are signiﬁcantly improved after image enhancement in preprocessing. Figure 4 improved compared with the unprocessed data, as shown in shows the MRI results of the CET1 sequence in 12 patients. Figure 5. Figure 5(a) shows the ROC curve in which )eleft-handsideofFigure4presents thatthe ﬁrstandsecond BraTS2017 is a training set and GliomaHPPH2018 is a rows are LGG without HSASR, and the third and fourth rows testing set, and Figure 5(b) shows the ROC curve in which are HGG without HSASR. All processed glioma data are the testing set and the training set interchange. It can be shown on the right side of the ﬁgure. )e examples show on observed from the ﬁgure that when testing set and training the left and right sides had a one-to-one correspondence. set swap, the better results can also be obtained. )e results show that the data processed by the HRASR method have From the perspective of image, the range of contrast and brightness of data without HSASR in Figure 4 are relatively good adaptability and can alleviate the diﬀerence among complex and the tumor area is fuzzy and diﬃcult to distin- multicenter data images. guish. )e data after HSASR are consistent in contrast and Experiment 2 in Table 3 presents the results of the mixed brightness, the tumor area is signiﬁcantly enhanced, and basic dataset. )e calculated characteristics are imported into the features (tumor morphology and lesion range) of the original proposed grading model for 10 iterations of random pre- image can be retained. diction, and the average value is subsequently obtained. Finally, the model with the best average performance is selected as the ﬁnal model. )e results show that AUC after 4.2. Performance of Glioma Grading. )ispaperaimstoreﬂect processing generally increased by more than 15%, and other the eﬀectiveness of these methods indirectly by using model indicators also improve signiﬁcantly. Experiment 3 in Ta- prediction indicators. To validate the eﬀect of the processed ble 3 shows that the data processed by HSASR on a single multicenterdatafurther,thissectionconductstheveriﬁcationof dataset also achieve a good grading eﬀect. grading results based on the following indicators: AUC, ac- Experiment 4 in Table 4 shows the comparison exper- curacy (ACC), sensitivity (SEN), and speciﬁcity (SPE). iment conducted by three diﬀerent methods on the mixed Feature reduction AUC score Journal of Healthcare Engineering 11 Coarse search 0.980 Fine search 0.98 0.978 0.96 0.976 0.94 0.92 0.974 0.90 0.972 0.88 0.970 0.86 6(78) 12(78) 18(78) 24(78) 30(78) 36(78) 42(78) 48(78) 69(10) 71(10) 73(10) 75(10) 77(10) 79(10) 81(10) N K (a) (b) Figure 9: Diagram of the proposed method (note: the red circle and blue square represent the AUC of machine and manually selected reference images, respectively. )e black square represents optimal reference frame). dataset. )ese methods include the HE method, reference Figure 9 illustrates the experimental results of the grid frame method of manual selection by radiologists, and the search in this study. In this study, ﬁve reference slices with the method proposed in this paper. )e results show that highest performance indexes were selected from two datasets via rough search, and the reference slices with the best per- compared withthe othertwo methods, the methodproposed in this paper signiﬁcantly improves the grading eﬀect of formance were screened out. )e N axis, K, circle, and square denoted the number of samples, the number of slices in each glioma. )e reference frame selected by radiologists based on experience shows signiﬁcant improvement in glioma sampleset,the valueof AUCof thereferenceimageselectedby grading, but it can be seen from the results that the per- the HSASR, and the predicted performance value of the formance of the selected reference frame is worser than that reference image selected by the doctor, respectively. )e ob- of the automatically selected. tained performance of the reference frame selected by the Table 5 gives the performance comparison of similar method is both low and high, indicating that the diﬀerence in work in the references using glioma dataset. In [27], after reference objects had a signiﬁcant impact on the result of data preprocessing, they used a modiﬁed version of AlexNet grading. )e proposed method can replace radiologists in for classifying MR brain images into three classes like choosing the best reference frame and save a considerable amount of time. Hence, the proposed method has certain healthy brain, LGG, and HGG. However, the amount of data of this work is so small that cannot verify the robustness of application value in clinical trials. the proposed method. Reference [28] also involved data preprocessing, but our data is twice as large, and the result 5. Conclusions veriﬁed by the grading model is better than this method. Figure 6 is the confusion matrix corresponding to the Inconsistencies among data prevent multicenter data from ﬁnal model of mixed glioma data in experiment 2, and playing to its shared advantage. )is paper proposes a LGG and HGG represent the low-level label and high-level histogram speciﬁcation method with automatic selection of label, respectively. Figure 6(a) shows the confusion matrix reference frames for magnetic resonance images to alleviate before processing, and Figure 6(b) shows the predicted the problem of contrast inconsistencies among multicenter data distribution after processing. Figure 6 shows a sig- data. )e core of histogram speciﬁcation is to change the niﬁcant decrease in the number of glioma predicted in- local brightness of the image according to the reference correctly after preprocessing. Meanwhile, in order to frame, but the reference frame of traditional histogram intuitively evaluate the performance diﬀerences of the speciﬁcation is usually manually selected by radiologists. model before and after preprocessing, Figures 7 and 8 list )is method not only increases the workload of radiologists, the performance heat maps of grading prediction before but also cannot guarantee the optimal reference frame. In and after preprocessing, respectively. )e horizontal and the method of this paper, the search range is ﬁrstly narrowed vertical coordinates and corresponding results represent by coarse search intraglioma samples, then the suitable the classiﬁer, feature selection or dimension reduction reference frame in histogram is selected by ﬁne search within method, and the maximum average AUC value, re- sample selected by coarse search. Finally, the eﬀectiveness spectively. )ese data indicate that the overall perfor- and feasibility of the proposed method are veriﬁed by a mance of the data after preprocessing is about 15% higher grading experiment based on two datasets. )e results show than that without HSASR processing. that multicenter data processed by this method have good AUC AUC 12 Journal of Healthcare Engineering International Journal of Pattern Recognition and Artiﬁcial adaptability, which improves the grading results and has Intelligence, vol. 31, no. 4, Article ID 1754006, 2017. certain practical value for clinical prediction. [13] M. F. A. Hassan, A. S. A. Ghani, D. Ramachandram, A. Radman, and S. A. Suandi, “Enhancement of under-ex- Data Availability posed image for object tracking algorithm through homo- morphic ﬁltering and mean histogram matching,” Advanced )e datasets used in this paper are public dataset Science Letters, vol. 23, no. 11, pp. 11257–11261, 2017. (BraTS2017) and Henan Provincial People’s Hospital [14] R. Berenguer, P. J. Mdr, z. J. Canales-Vazquez ´ et al., (GliomaHPPH2018) dataset; BraTS2017 can be obtained “Radiomics of CT features may be nonreproducible and re- through the following URL: https://www.med.upenn.edu/ dundant: inﬂuence of CTacquisition parameters,” Radiology, sbia/brats2017/data.html. vol. 288, no. 2, Article ID 172361, 2018. [15] H. Schnack, N. van Haren, M. Picchioni et al., “A multicenter mri schizophrenia study: reliability of voxel-based mor- Conflicts of Interest phometry,” Schizophrenia Research, vol. 102, no. 1–3, p. 75, )e authors declare that they have no conﬂicts of interest. [16] L. G. Nyul, J. K. Udupa, and X. Xuan Zhang, “New variants of a method of MRI scale standardization,” IEEE Transactions on Acknowledgments Medical Imaging, vol. 19, no. 2, pp. 143–150, 2000. [17] S. Bakas, H. Akbari, A. Sotiras et al., “Advancing the Cancer )is work was supported by the National Natural Science Genome Atlas glioma MRI collections with expert segmen- tation labels and radiomic features,” Scientiﬁc Data, vol. 4, Foundation of China under grant 81772009 and Scientiﬁc Article ID 170117, 2017. and Technological Research Project of Henan Province [18] B. H. Menze, A. Jakab, S. Bauer et al., “)e multimodal brain under grant 182102310162. tumor image segmentation benchmark (BRATS),” IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993– References 2024, 2015. [19] G. Chen, L. Li, W. Jin, M. Liu, and F. Shi, “Image contrast [1] Q. T. Ostrom, H. Gittleman, J. Xu et al., “CBTRUS statistical enhancement method based on display and human visual report: primary brain and other central nervous system tu- system characteristics,” Applied Optics, vol. 58, no. 7, mors diagnosed in the United States in 2009-2013,” Neuro- pp. 1813–1823, 2019. Oncology, vol. 18, no. 5, pp. v1–v75, 2016. [20] C.-M. Chen, C.-C.Chen,M.-C. Wu etal., “Automatic contrast [2] D. N. Louis, H. Ohgaki, O. D. Wiestler et al., “)e 2007 WHO enhancement of brain MR images using hierarchical corre- classiﬁcation of tumours of the central nervous system,” Acta lation histogram analysis,” Journal of Medical and Biological Neuropathologica, vol. 114, no. 5, p. 547, 2007. Engineering, vol. 35, no. 6, pp. 724–734, 2015. [3] W. Shan and X.-L. Wang, “Clinical application value of 3.0T [21] X. Wang and L. Chen, “An eﬀective histogram modiﬁcation MR diﬀusion tensor imaging in grade diagnosis of gliomas,” scheme for image contrast enhancement,” Signal Processing: Oncology Letters, vol. 14, no. 2, pp. 2009–2014, 2017. Image Communication, vol. 58, pp. 187–198, 2017. [4] V. Kumar, Y. Gu, S. Basu et al., “Radiomics: the process and [22] M. Paulinas and A. Usinskas, “A survey of genetic algorithms the challenges,” Magnetic Resonance Imaging, vol. 30, no. 9, applications for image enhancement and segmentation,” In- pp. 1234–1248, 2012. formation Technology and Control, vol. 36, no. 3, pp. 278–284, [5] R. Rohan, H. Adam, B. Jason, R. Robert, and S. Daniel, “Outcomes in reoperated low-grade gliomas,” Neurosurgery, [23] A. Draa and A. Bouaziz, “An artiﬁcial bee colony algorithm vol. 77, no. 2, p. 175, 2015. for image contrast enhancement,” Swarm and Evolutionary [6] E. Ferrante and N. Paragios, “Slice-to-volume medical image Computation, vol. 16, pp. 69–84, 2014. registration: a survey,” Medical Image Analysis, vol. 39, [24] L. Rundo, A. Tangherloni, M. S. Nobile, C. Militello, and pp. 101–123, 2017. P. Cazzaniga, “MedGA: a novel evolutionary method for [7] Y. Zhu, “Review of image enhancement algorithms,” China image enhancement in medical imaging systems,” Expert Computer and Communication, vol. 10, no. 4, pp. 438–448, Systems with Applications, vol. 119, 2018. [25] S.-D. Chen and A. Rahman Ramli, “Preserving brightness in [8] L. Zhuang and Y. Guan, “Image enhancement via subimage histogram equalization based contrast enhancement tech- histogram equalization based on mean and variance,” Com- niques,” Digital Signal Processing, vol. 14, no. 5, pp. 413–428, putational Intelligence and Neuroscience, vol. 2017, no. 2, Article ID 6029892, 2017. [26] N. Sengee, A. Sengee, and H. K. Choi, “Image contrast en- [9] M. F. Khan, E. Khan, and Z. A. Abbasi, “Image contrast hancement using Bi-histogram equalization with neighbor- enhancement using normalized histogram equalization,” hood metrics,” IEEE Transactions on Consumer Electronics, Optik, vol. 126, no. 24, pp. 4868–4875, 2015. vol. 56, no. 4, pp. 2727–2734, 2011. [10] M. Ramkumar and B. Karthikeyan, “A survey on image [27] S. Khawaldeh, U. Pervaiz, A. Raﬁq, and R. Alkhawaldeh, enhancement methods,” International Journal of Engineering “Noninvasive grading of glioma tumor using magnetic res- and Technology, vol. 5, no. 2, pp. 960–962, 2013. onance imaging with convolutional neural networks,” Applied [11] M. Kanmani and V. Narsimhan, “An image contrast en- Sciences, vol. 8, no. 1, p. 27, 2017. hancement algorithm for grayscale images using particle [28] W. Chen, B. Liu, S. Peng, J. Sun, and X. Qiao, “Computer- swarm optimization,” Multimedia Tools and Applications, aided grading of gliomas combining automatic segmentation vol. 77, no. 3, pp. 1–17, 2018. and radiomics,” International Journal of Biomedical Imaging, [12] Z. Jing, G. Chen, and Z. Jia, “An image stitching algorithm vol. 2018, Article ID 2512037, 11 pages, 2018. based on histogram matching and SIFT algorithm,” International Journal of Advances in Rotating Machinery Multimedia Journal of The Scientific Journal of Engineering World Journal Sensors Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Control Science and Engineering Advances in Civil Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com Journal of Journal of Electrical and Computer Robotics Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 VLSI Design Advances in OptoElectronics International Journal of Modelling & Aerospace International Journal of Simulation Navigation and in Engineering Engineering Observation Hindawi Hindawi Hindawi Hindawi Volume 2018 Volume 2018 Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com www.hindawi.com www.hindawi.com Volume 2018 International Journal of Active and Passive International Journal of Antennas and Advances in Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Healthcare Engineering Hindawi Publishing Corporation http://www.deepdyve.com/lp/hindawi-publishing-corporation/automatic-histogram-specification-for-glioma-grading-using-multicenter-vfz7pGtPrb

Loading next page...

References (38)

Y. Zhu (2017)
Review of image enhancement algorithms,
China Computer and Communication, 10
R. Berenguer, M. Pastor-Juan, J. Canales‐Vázquez, M. Castro-García, M. Villas, F. Legorburo, S. Sabater (2018)
Radiomics of CT Features May Be Nonreproducible and Redundant: Influence of CT Acquisition Parameters.
Radiology, 288 2
Enzo Ferrante, N. Paragios (2017)
Slice-to-volume medical image registration: a survey
Medical image analysis, 39
Soong-Der Chen, A. Ramli (2004)
Preserving brightness in histogram equalization based contrast enhancement techniques
Digit. Signal Process., 14
Make comparison experiments between this research method and common image enhancement methods on the mixed dataset
Xuewen Wang, Lixia Chen (2017)
An effective histogram modification scheme for image contrast enhancement
Signal Process. Image Commun., 58
W. Shan, Xing-long Wang (2017)
Clinical application value of 3.0T MR diffusion tensor imaging in grade diagnosis of gliomas
Oncology Letters, 14
Wei Chen, Boqiang Liu, Suting Peng, Jiawei Sun, Xu Qiao (2018)
Computer-Aided Grading of Gliomas Combining Automatic Segmentation and Radiomics
International Journal of Biomedical Imaging, 2018
M. Kanmani, V. Narsimhan (2018)
An image contrast enhancement algorithm for grayscale images using particle swarm optimization
Multimedia Tools and Applications, 77
Q. Ostrom, H. Gittleman, Jordan Xu, Courtney Kromer, Yingli Wolinsky, C. Kruchko, J. Barnholtz-Sloan (2016)
CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2009-2013.
Neuro-oncology, 18 suppl_5
The optimal reference slice map is the final output. If multiple optimal values exist, then selection is made based on accuracy and specificity indicators
Ramkumar M, K. B. (2013)
A Survey on Image Enhancement Methods
H. Schnack, N. Haren, M. Picchioni, M. Weisbrod, H. Sauer, T. Cannon, M. Huttunen, R. Murray, R. Kahn, H. Pol (2008)
A MULTICENTER MRI SCHIZOPHRENIA STUDY: RELIABILITY OF VOXEL-BASED MORPHOMETRY
Schizophrenia Research, 102
Bjoern Menze, A. Jakab, S. Bauer, Jayashree Kalpathy-Cramer, K. Farahani, J. Kirby, Y. Burren, N. Porz, J. Slotboom, R. Wiest, L. Lanczi, E. Gerstner, M. Weber, T. Arbel, B. Avants, N. Ayache, Patricia Buendia, D. Collins, Nicolas Cordier, Jason Corso, A. Criminisi, T. Das, H. Delingette, Çağatay Demiralp, C. Durst, M. Dojat, Senan Doyle, Joana Festa, F. Forbes, Ezequiel Geremia, B. Glocker, P. Golland, Xiaotao Guo, A. Hamamci, K. Iftekharuddin, R. Jena, N. John, E. Konukoglu, D. Lashkari, J. Mariz, Raphael Meier, Sérgio Pereira, Doina Precup, S. Price, Tammy Riklin-Raviv, Syed Reza, Michael Ryan, Duygu Sarikaya, L. Schwartz, Hoo-Chang Shin, J. Shotton, Carlos Silva, N. Sousa, N. Subbanna, G. Székely, Thomas Taylor, O. Thomas, N. Tustison, Gözde Ünal, F. Vasseur, M. Wintermark, Dong Ye, Liang Zhao, Binsheng Zhao, D. Zikic, M. Prastawa, M. Reyes, K. Leemput (2015)
The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)
IEEE Transactions on Medical Imaging, 34
Liyun Zhuang, Ye-peng Guan (2017)
Image Enhancement via Subimage Histogram Equalization Based on Mean and Variance
Computational Intelligence and Neuroscience, 2017
Abu Fauzi, A. Shahrizan, Ramachandram Dhanesh, Radman Abduljalil, SuandiShahrel Azmin (2017)
Enhancement of Under-Exposed Image for Object Tracking Algorithm Through Homomorphic Filtering and Mean Histogram Matching
Advanced Science Letters, 23
L. Nyúl, J. Udupa, Xuan Zhang (2000)
New variants of a method of MRI scale standardization
IEEE Transactions on Medical Imaging, 19
(2017)
International Journal of Pattern Recognition and Artificial Intelligence , vol. 31, no. 4, Article ID 1754006
Virendra Kumar, Yuhua Gu, Satrajit Basu, A. Berglund, S. Eschrich, M. Schabath, K. Forster, H. Aerts, A. Dekker, D. Fenstermacher, Dmitry Goldgof, L. Hall, P. Lambin, Y. Balagurunathan, R. Gatenby, R. Gillies (2012)
Radiomics: the process and the challenges.
Magnetic resonance imaging, 30 9
Saed Khawaldeh, Usama Pervaiz, A. Rafiq, Rami Alkhawaldeh (2017)
Noninvasive Grading of Glioma Tumor Using Magnetic Resonance Imaging with Convolutional Neural Networks
Applied Sciences, 8
N. Sengee, Altansukh Sengee, Heung-Kook Choi (2010)
Image contrast enhancement using bi-histogram equalization with neighborhood metrics
IEEE Transactions on Consumer Electronics, 56
Louis N, Ohgaki H, Cavenee K (1982)
The classification of tumours of the central nervous system.
Neuropathology and applied neurobiology, 8 1
R. Ramakrishna, A. Hebb, Jason Barber, R. Rostomily, D. Silbergeld (2015)
Outcomes in Reoperated Low-Grade Gliomas.
Neurosurgery, 77 2
Mantas Paulinas, A. Ušinskas (2007)
A SURVEY OF GENETIC ALGORITHMS APPLICATIONS FOR IMAGE ENHANCEMENT AND SEGMENTATION
Information Technology and Control, 36
A. Draa, Amira Bouaziz (2014)
An artificial bee colony algorithm for image contrast enhancement
Swarm Evol. Comput., 16
The main contributions of this study
Jing Zhang, G. Chen, Zhaoyang Jia (2017)
An Image Stitching Algorithm Based on Histogram Matching and SIFT Algorithm
Int. J. Pattern Recognit. Artif. Intell., 31
L. Rundo, A. Tangherloni, Marco Nobile, C. Militello, D. Besozzi, G. Mauri, P. Cazzaniga (2019)
MedGA: A novel evolutionary method for image enhancement in medical imaging systems
Expert Syst. Appl., 119
Select the sets of all the best area under the curve (AUC) values according to threshold T and perform the next fine search
Guo Chen, Li Li, Weiqi Jin, Mingcong Liu, Feng Shi (2019)
Image contrast enhancement method based on display and human visual system characteristics.
Applied optics, 58 7
王. Hao, 张. Ye, 沈宏海 Hong-hai, 张景忠 Jing-zhong (2017)
Review of image enhancement algorithms
Chinese Optics, 10
S. Bakas, H. Akbari, Aristeidis Sotiras, M. Bilello, Martin Rozycki, J. Kirby, J. Freymann, K. Farahani, C. Davatzikos (2017)
Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features
Scientific Data, 4
D. Louis, H. Ohgaki, O. Wiestler, W. Cavenee, P. Burger, A. Jouvet, B. Scheithauer, P. Kleihues (2007)
The 2007 WHO classification of tumours of the central nervous system
Acta Neuropathologica, 114
Chiao-Min Chen, Chih-Cheng Chen, Ming-Chi Wu, G. Horng, Hsien-Chu Wu, Shih-Hua Hsueh, His-Yun Ho (2015)
Automatic Contrast Enhancement of Brain MR Images Using Hierarchical Correlation Histogram Analysis
Journal of Medical and Biological Engineering, 35
Take the processed GliomaHPPH2018 and BraTS2017 datasets as the training and testing sets, respectively. Then
Mohammad Khan, E. Khan, Z. Abbasi (2015)
Image contrast enhancement using normalized histogram equalization
Optik, 126
Q. Ostrom, H. Gittleman, Peter Liao, Chaturia Rouse, Yanwen Chen, J. Dowling, Yingli Wolinsky, C. Kruchko, J. Barnholtz-Sloan (2012)
CBTRUS statistical report: primary brain and central nervous system tumors diagnosed in the United States in 2007-2011.
Neuro-oncology, 16 Suppl 4
Use single sequences from the processed and un-processed GliomaHPPH2018 and BraTS2017 as comparison experiment

Publisher: Hindawi Publishing Corporation
Copyright: Copyright © 2019 Xi Chen et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN: 2040-2295
eISSN: 2040-2309
DOI: 10.1155/2019/9414937
Publisher site: See Article on Publisher Site

Abstract

Hindawi Journal of Healthcare Engineering Volume 2019, Article ID 9414937, 12 pages https://doi.org/10.1155/2019/9414937 Research Article Automatic Histogram Specification for Glioma Grading Using Multicenter Data 1,2 3 2 3 2 4 Xi Chen, Yaping Wu, Guohua Zhao, Meiyun Wang , Wenyi Gao, Qian Zhang, 1,2 and Yusong Lin School of Software, Zhengzhou University, Zhengzhou, Henan 450002, China Collaborative Innovation Center for Internet Healthcare, Zhengzhou University, Zhengzhou, Henan 450052, China Department of Radiology, Henan Provincial People’s Hospital, Zhengzhou, Henan 450003, China School of Computer Science, Zhongyuan University of Technology, Zhengzhou 450000, China Correspondence should be addressed to Yusong Lin; yslin@ha.edu.cn Received 22 July 2019; Revised 6 November 2019; Accepted 23 November 2019; Published 19 December 2019 Academic Editor: Emiliano Schena Copyright©2019XiChenetal.)isisanopenaccessarticledistributedundertheCreativeCommonsAttributionLicense,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Multicenter sharing is an eﬀective method to increase the data size for glioma research, but the data inconsistency among diﬀerent institutions hindered the eﬃciency. )is paper proposes a histogram speciﬁcation with automatic selection of reference frames for magnetic resonance images to alleviate this problem (HSASR). )e selection of reference frames is automatically performed by an optimized grid search strategy with coarse and ﬁne search. )e search range is ﬁrstly narrowed by coarse search of intraglioma samples, and then the suitable reference frame in histogram is selected by ﬁne search within the sample selected by coarse search. Validation experiments are conducted on two datasets GliomaHPPH2018 and BraTS2017 to perform glioma grading. )e results demonstrate the high performance of the proposed method. On the mixed dataset, the average AUC, accuracy, sensitivity, and speciﬁcity are 0.9786, 94.13%, 94.64%, and 93.00%, respectively. It is about 15% higher on all indicators compared with those without HSASR and has a slight advantage over the result of a manually selected reference frame by radiologists. Results show that our methods can eﬀectively alleviate multicenter data inconsistencies and lift the performance of the prediction model. accurately display the location and size, and it correlates well 1. Introduction with histological characteristics. In medical research, high quality data is diﬃcult to obtain in a single institution, so it Glioma is a prevalent fatal brain disease and the most malignant, which accounts for approximately 24.7% of all needs to be shared through multicenter. However, the primary brain and other central nervous system tumors and diﬀerence of multicentre data is a serious challenge. 74.6% of malignant tumors [1]. )e World Health Orga- In the acquisitionof multicenter gliomaMRI data,due to nization’s guidelines for glioma diagnosis and treatment are diﬀerences in acquisition equipment and parameters result divided into four levels, namely, I–II and III–IV for low- signiﬁcant diﬀerences in data samples in terms of speciﬁ- grade glioma (LGG) and high-grade glioma (HGG) [2]. In cation, size, contrast, and brightness. Diﬀerences among the clinical applications, biological behavior, treatment options, data lead to deviations in the grading eﬀect of glioma [4]. and prognoses of patients with glioma of diﬀerent grades are Figure 1 shows the diﬀerences among multicenter data. clearly diﬀerent. )erefore, the accurate preoperation Figure 1(a) shows that some of the data contain skulls and grading of Glioma is important. Magnetic Resonance Im- the others do not contain. Figure 1(b) shows that these data aging (MRI) is characterized by multidirectional tomogra- have diﬀerent scales and number of slices. A previous study phy and multiparameter high-resolution soft-tissue imaging indicated that pixel size and slice thickness remarkably aﬀect and is widely used to evaluate the tumor heterogeneity [3]. space and strength characteristics of the calculation [5, 6]. MRI is commonly used in glioma grading because it can )erefore, voxel size must be uniﬁed to reduce the error in 2 Journal of Healthcare Engineering )e main contributions of this study are as follows: (1) )is paper proposes a histogram speciﬁcation with automatic selection of reference frames (HSASR). (a) )is method can automatically select the suitable reference frame of histogram instead of radiologists during image enhancement, so as to enhance the consistency of brain tumors image contrast. (2) )is paper proposes a set of image standardization (b) algorithms to make the preprocessed multicenter data have better consistency, improve the adapt- 512 × 512 × 18 320 × 320 × 18 240 × 240 × 155 ability of data, and improve the accuracy of the glioma prediction model. 2. Related Works C1 C2 C3 (c) In recent years, researchers have gradually adopted multi- center data to replace single institution in clinical medical research. However, multicenter data also faced many C4 C5 C6 challenges. Berenguer et al. [14] carried out a test-retest phantom study of individual image acquisition parameters. Figure 1: Multicenter data. (a) Space of dataset is uneven. (b) )e impact of image parameters on the image cannot be Volume pixel of data element is not uniform. (c) Dataset contrast analyzed because of the diﬀerence of the model or machine brightness is not uniform, and C1–C6 are randomly selected data manufacturer. )erefore, variations because of the use of samples. images were eliminated by exclusively scanning phantoms. In addition, Hugo et al. [15] indicated that multicenter structure MRI studies have stronger statistical eﬃcacy than calculating characteristics. Figure 1(c) presents six images with diﬀerent contrasts. )ese diﬀerences are due to de- single-institution studies. However, central diﬀerences in viation in acquisition equipment and parameters. contrast sensitivity and spatial uniformity lead to diﬀerences In order to alleviate the inconsistency of contrast of in tissue classiﬁcation or image registration that may reduce multicenter data, it is necessary to enhance the image data or completely oﬀset the enhanced statistical eﬃcacy of appropriately. Histogram correction is a commonly used multicenter data. )erefore, maintaining data in a standard technique in image enhancement [7]. Among them, histo- environment is important. Nyul et al. [16] mentioned some gram equalization (HE) and normalization techniques [8, 9] problems in the original MRI scale preprocessing method are often used to adjust the overall brightness of images. )is and attempted to use the median value and other percentile type of methods mainly expands the range of gray value in numerical methods, such as landmark, to solve these the histogram without adjusting its intensity. )ese methods problems; they obtained robust results after preprocessing. )e validity of the standardized new landmark was also only enhance the overall image and improve its brightness, but the contrast between tumors and other tissues is not mentioned in the study, after which the image brightness considerably enhanced [10, 11]. Histogram speciﬁcation level and contrast consistency were signiﬁcantly improved. (HS) can change the frequency of grayscale values and Bakas et al. [17, 18] focused on the data preprocessing enhance any local brightness [12, 13], but it mainly adjusts method, which has a certain enlightening eﬀect on the the local brightness according to the reference frame. In preprocessing method of this study. general, the reference frame is selected by a radiologist. In preprocessing, contrast enhancement of multicenter However, this process consumes a considerable amount of data image is an important challenge. It was shown that a time and the ﬁnal reference frame may not be the best high-contrast medical image could lead to a better in- choice. )erefore, it is an urgent problem to automatically terpretation of the diﬀerent adjacent tissues in the imaged search for the best reference frame of histogram instead of body part [19, 20]. Accordingly, the resulting enhanced searching by a radiologist. image, which is in terms of signal intensities of diﬀerent )is paper presents a histogram speciﬁcation with au- tissues, can facilitate the automated segmentation, feature tomatic selection of reference frames for magnetic resonance extraction, and classiﬁcation of these tissues. Existing image images, called HSASR, speciﬁcally used to replace radiolo- enhancement techniques (empirical or heuristic) are re- gists manually selecting reference frames. )is method can markably related to a particular image and usually aimed at enhance the consistency of tumor intensity in the whole improving image contrast. However, no uniﬁed standard is medical image. In this work, we apply HSASR to the pre- available to measure the quality eﬀect of image enhancement processing of multicenter glioma data, and the eﬀectiveness [21–24]. of this method is proved by Glioma grading experiment. )e Image enhancement can be divided into two categories: results showed that this method had certain practical value spatial domain method and frequency domain method. in glioma grading research. Image enhancement in spatial domain usually corrects Journal of Healthcare Engineering 3 histogram. Histogram equalization is a commonly used N (1) P (l) � , ∀l ∈ L, global grayscale image enhancement technique, and the grayscale value is uniformly redistributed based on the where N � H × W and N is the number of pixels with a gray cumulative density function of the histogram [25]. However, level. )e cumulative distribution function (CDF) of P is equalization fails to consider the intensity of the grayscale r presented as follows: value. )e average brightness of the image can make the x x regions with large original intensity fade, whereas dark areas N s (x) � 􏽘 P (j) � 􏽘 , ∀x ∈ L. (2) r r will brighten after image equalization. Sengee et al. [26] j�0 j�0 suggested an extension method of BBHE based on the neighbourhood metric. )is method involved a few steps: Similarly, if the speciﬁed reference image is z, then the ﬁrst, a large histogram was divided into the subregion using gray cumulative distribution function is the neighbourhood metric and process independently. )is method results in boundary noise and possibly uneven v (y) � 􏽘 P (i) � 􏽘 , ∀y ∈ L, (3) z z histogram brightness in two parts. In addition, MedGA [24] M i�0 i�0 wasan enhancementmethodthatHEcombinedwithgenetic algorithm to directly improve the histogram frequency of where Mand N arethe numberofpixelswith graylevel L.HS attempts to obtain transformation function y � F(x) and images and has achieved obvious results. However, this method could only be limited to the presence of two maps gray level x in the original image to gray level y, such that the transformed image can have a histogram similar to grayscale regional tissues and not enhance complex brain tumors. the reference histogram. To preserve the inherent in- formation of the original image, function F should be a monotonically increasing function. )is function can be 3. Materials and Methods obtained by using the following equation: 3.1. Data Collection. Glioma data used in the study were v (y) � s (x). (4) z r obtained from two separate sources, BraTS2017 and Glio- )erefore, the gray level of the map can be obtained by maHPPH2018 datasets. )e BraTS2017 dataset came from a variety of scanning instruments from 19 medical in- using stitutions. BraTS2017 dataset includes 210 HGG data and 75 − 1 (5) y � v 􏼂s (x)􏼃, z r LGG data; the data was already segmented when it was − 1 acquired. GliomaHPPH2018 dataset was obtained from where v is the inverse of v . multiple equipment of diﬀerent models from multiple In the discrete case, the inverse function usually does not manufacturers, including 4 Siemens equipment (1 1.5t exist. )e inverse function is usually replaced by the best equipment, 3 3T equipment) and 4GE equipment (2 1.5t objective function to approximate the y of a particular gray equipment, 2 3T equipment), as shown in Table 1. )e level x as follows: 􏼌 􏼌 magnetic resonance data collected on these devices are 􏼌 􏼌 􏼌 􏼌 y � argmin 􏼌 s (x) − v (k)􏼌. r z (6) diﬀerent from each other. After data inspection, there are up to 186 default inspection protocols for the devices, and Equation (6) represents the absolute value of the dif- contrast-enhanced T1-weighted imaging (CET1) was used in ference between the original image’s cumulative histogram this study, including 35 CET1 sequences (see Table 2), layer and gray level functions of the speciﬁed reference histogram, thickness of 5mm to 6.5mm, repetition time of 220– and then the minimum value is selected as the y value. With 1970ms, echo time of 2.46–28.60ms, and resolution of this transformation rule, each gray level x can be mapped to 256 ×256 ×18, 320 ×290 ×18, 320 ×320 ×18, 384 ×384 ×18, y. )erefore, the mapped image will be similar to the desired 448 ×408 ×18, and 512 ×512 ×18. )e GliomaHPPH2018 histogram. dataset includes 161HGG data and 77 LGG data. A three- )e reference frame in the histogram is usually selected dimensional region of interest (ROI) for all the tumors in by the doctor. However, the selected reference image may GliomaHPPH2018 was depicted by two senior radiologists not be suitable, thereby consuming a considerable amount of in Henan provincial people’s hospital. )e ROIs were seg- the radiologist’s time. To solve this problem, an optimized mented slice-by-slice on the axial plane using the CET1 grid search strategy with coarse and ﬁne search is proposed sequences. After discussion between the two radiologists, the to select the suitable reference frame automatically. ﬁnal decision is made to select the optimal segmentation ﬁle. 3.3. HSASR. In this study, an improved grid search method 3.2. Histogram Speciﬁcation. HS, an eﬀective image en- is proposed to solve the problem of selecting a speciﬁed hancement technique, is an extension of histogram equal- reference image. )e grid search is generally used to divide ization that can eﬀectively alleviate the problems of grids of the same length in a certain spatial range based on histogram equalization. Let r � 􏽮r 􏽯be an H × W discrete ij the proposed coordinate system. Each coordinate point input digital image with L gray levels, and let L �{0, 1, . . ., represents a parameter, and these searched parameters are L − 1}. )e histogram or gray level probability density of an calculated and analysed based on the step size. Finally, image is deﬁned as follows: 4 Journal of Healthcare Engineering Table 1: GliomaHPPH2018MR acquisition equipment list. Number Manufacturer Model Magnetic ﬁeld strength Amount 1 GE OPTIMA MR360 1.5T 1 2 GE Signa HDxt 1.5T 1 3 GE DISCOVERY MR750 3T 2 4 SIEMENS Sempra 1.5T 1 5 SIEMENS Prisma 3T 1 6 SIEMENS TrioTim 3T 1 7 SIEMENS Verio 3T 1 Note: amount is for number of machines. Table 2: CET1 sequence collection parameter statistics. Number Manufacturer Magnetic ﬁeld Model D-spacing Layer thickness Resolution Voxel size 1 GE 1.5 OPTIMA MR360 6.5 5.5 512 ×512 ×18 0.4688\0.4688 2 GE 1.5 OPTIMA MR360 7 6 512 ×512 ×18 0.4688\0.4688 3 GE 1.5 OPTIMA MR360 7.5 6 256 ×256 ×18 0.9375\0.9375 4 GE 1.5 OPTIMA MR360 7.5 6 512 ×512 ×18 0.4688\0.4688 5 GE 1.5 OPTIMA MR360 8 6 512 ×512 ×18 0.4688\0.4688 6 GE 1.5 Signa HDxt 7 6 512 ×512 ×18 0.4688\0.4688 7 GE 1.5 Signa HDxt 7 6 512 ×512 ×18 0.5273\0.5273 8 GE 1.5 Signa HDxt 7.5 6 512 ×512 ×18 0.4688\0.4688 9 GE 1.5 Signa HDxt 7.5 6.5 512 ×512 ×18 0.4688\0.4688 10 GE 1.5 Signa HDxt 8 6 512 ×512 ×18 0.4688\0.4688 11 GE 3 DISCOVERY MR750 6.5 5 512 ×512 ×18 0.4688\0.4688 12 GE 3 DISCOVERY MR750 7 6 512 ×512 ×18 0.4688\0.4688 13 GE 3 DISCOVERY MR750 7.5 6 512 ×512 ×18 0.4688\0.4688 14 SIEMENS 1.5 Sempra 7.2 6 448 ×408 ×18 0.5134\0.5134 15 SIEMENS 3 Prisma 7.2 6 320 ×320 ×18 0.7188\0.7188 16 SIEMENS 3 Prisma 7.2 6 384 ×384 ×18 0.5989\0.5989 17 SIEMENS 3 Prisma 7.8 6 320 ×320 ×18 0.7188\0.7188 18 SIEMENS 3 TrioTim 6.6 6 320 ×320 ×18 0.7188\0.7188 19 SIEMENS 3 TrioTim 6.6 6 320 ×320 ×18 0.75\0.75 20 SIEMENS 3 TrioTim 7.2 6 256 ×256 ×18 0.8984\0.8984 21 SIEMENS 3 TrioTim 7.2 6 256 ×256 ×18 0.9375\0.9375 22 SIEMENS 3 TrioTim 7.2 6 320 ×290 ×18 0.7188\0.7188 23 SIEMENS 3 TrioTim 7.2 6 320 ×320 ×18 0.7188\0.7188 24 SIEMENS 3 TrioTim 7.2 6 320 ×320 ×18 0.75\0.75 optimal parameters are considered the output. Given that randomly selected for histogram speciﬁcation (the this method must traverse all the corresponding points in same sample is selected each time), and then the the grid, many unnecessary invalid calculations are gener- performance of data after histogram speciﬁcation is tested. ated, resulting in an exponential increase in time. To reduce time consumption, this study improves the (3) Select the sets of all the best area under the curve directional grid optimization search based on data charac- (AUC) values according to threshold Tand perform teristics. On the basis of the characteristics of glioma itself, the next ﬁne search. the method is improved by using coarse segmentation and (4) Fine search: in the optimal cluster, the search is subdivision. Horizontal and vertical searches are used for initiated from the middle section (two pointers are coarse segmentation and subdivision, respectively. )e set to point to the middle section), and pointer 1 speciﬁc steps of the improved method are as follows: searches upward successively. If reference frames (1) )e number of clusters N and the number of slices K do not contain tumors, then the search is stopped. of each tumor are determined. Each cluster is set to i, If the diﬀerence between the current and previous and the initialization step size is K/2. A 2D grid is values is greater than T, then the search is stopped, established by using N and K, and the grid nodes are and the optimal value is selected. At the same time, the corresponding reference slices of N and K. similar to the previous step, pointer 2 is searched downward. (2) In the coarse search, each time it searched by step size from the beginning of each set(horizontally), the (5) )e optimal reference slice map is the ﬁnal output. If median slice in each cluster was selected as the multiple optimal values exist, then selection is made reference frame. )en, 30% of data in all datasets are based on accuracy and speciﬁcity indicators. If Journal of Healthcare Engineering 5 (i) Input: Mixed dataset F, data sample number N, number of each sample slice K (ii) Output: Optimal reference slice F (1) Step1: Coarse search (2) Function HSBM (n, F1, l )//Histogram speciﬁcation based on brain MRI (3) { (4) F1 �Random (range (0, N), F∗30%)//Random selection of 30% of dataset (5) For each n ∈ F do (6) For each n1 ∈ F1 do (7) cdf � hist · cumsum( ) F F 1 n1 n (8) cdf � hist · cumsum( ) F F n n k/2 min (9) Calculate: hist � arg |(cdf /shape(F )) − ((cdf /shape(F ))(m))| F n Fn1 n1 n k/2 k/2 (10) End for (11) Calculate: Auc(n ) � roc auc score(Lable − Lable ) K/2 original predict (12) End for (13) } (14) F2 � Sort max(Auc ) k/2 (15) F2 �Max(F2) (16) For each c ∈ F2 do (17) if |F2 − Auc(c)|≤ T then (18) Select c ⟶ F3 (19) end (20) End for (21) Select the optimal set F3 (22) Step 2 Fine search (23) Initializes Pointers l , l 1 2 (24) For each n ∈ F3//Search both ends (25) l � k/2 (26) For l ≤ kandexit(tumor ); l + +do 1 l ++ 1 (27) HSBM (n, F1, l ) (28) End for (29) End for (30) For each n ∈ F3 (31) l � k/2 (32) For l ≤ kandexit(tumor ); l + +do 2 l ++ 2 (33) HSBM (n, F1, l ) (34) End for (35) End for (36) Output: select the ﬁrst optimal slice F ALGORITHM 1: HSASR algorithm pseudocode description. several optimal reference frames existed, then the 3.4. Experiments. Figure 3 shows the overall process of ﬁrst optimal reference frame is selected as the ﬁnal multicenter data grading prediction. First, data are im- reference frame. ported, data preprocessing is performed, features are extracted by using the feature calculation method, features Algorithm 1 describes the HSASR algorithm. )e pa- are selected, and model training is ﬁnally conducted. rameter T in Step 1 prevents the collection of additional Contribution points of this study are in the preprocessing optimal values from being missed. )e T in Step 2 saves stage. )e HSASR method proposed is the most important search time. Figure 2 presents the 2D diagram of the im- part of preprocessing. proved optimization grid search. In this study, N represents the number of data samples, K denotes the number of slices of each sample, and K/2 is the middle slice of the sample. 3.5. Preprocessing. Storage formats of multicenter data are First, a horizontal coarse search is performed. )e pale blue often inconsistent. Data formats of this study include NIfTI circles indicate the performance values of intermediate slices and DICOM format ﬁles. )erefore, to standardize multi- in each cluster. )e optimal clusters are selected for lon- center data formats, this study adopts the format conversion gitudinal ﬁne search, the middle slice is considered the initial method based on the Convert3D tool to unify all data sample value, and both ends are searched simultaneously. )e formats into the NIfTI format. )e majority of patients in vertical circles and dark blue squares represent the perfor- the GliomaHPPH2018 dataset have 18 DICOM ﬁles with mance values in the optimal set and the selected ﬁnal ref- CET1 sequence, and a small number of patients have 36 ﬁles. erence frame, respectively. After format conversion, all slices of each patient’s CET1 6 Journal of Healthcare Engineering k/2 Intermediate slice as reference Optimal slice as reference Single sample slice as reference Figure 2: Diagram of the optimized grid search (note: k/2 represents the middle slice of the cluster). Multicenter standardization Feature engineering Feature calculation Feature selection Grading prediction Preprocessing Raw data BraTS2017 Feature Set1 Format conversion Feature selection diagram PCA Brain tissues extraction Pyradiomics Feature Set2 Feature samples Resampling x x GliomaHPPH2018 Selection feature Predictive analysis Figure 3: Flowchart of multicenter data grading prediction. sequence are converted into the NIfTI format ﬁle, thereby In addition, spatial consistency of the data image is maintained.Multiple voxelrangesoccur inboth datasets.On providing convenience in subsequently unifying data processing. the basis of Convert3D, we adopt the shortest distance in- Furthermore, to unify the multicenter data in pre- terpolation resampling technology to carry out scale processing, maintaining the consistency of brain tissue resampling in each data sample. )e size of each data sample structureis alsoimportant.)e GliomaHPPH2018 datasetin is uniﬁed to 240 ×240 ×155, and the label sample is also this study contained skull images, which are removed at the converted into a ﬁle of the same size. In addition, the dis- time of acquisitionin the BraTS2017dataset. )issituation is tance between uniﬁed slices is 1mm, and the original po- not only a certain impact on HS but also introduced diﬃ- sition was [0, − 239, 0]. culties in combining the two datasets due to the gap between To solve the problem of inconsistent brightness and them. In this study, the FSL tool is ﬁrstly used in location contrast of multicenter data, this paper proposes a histogram registration, and then the skulls are removed based on the speciﬁcation withautomatic selection ofreferenceframes for bet script. In this study, 238 data skulls are removed, and magnetic resonance images. On the basis of the character- brain tissues and regions of interest are relatively intact. istics of glioma itself, this paper puts forward an optimized Journal of Healthcare Engineering 7 MRI without HSASR MRI with HSASR L1 L2 L3 L1 L2 L3 L4 L5 L6 L4 L5 L6 H1 H2 H3 H1 H2 H3 H4 H5 H6 H4 H5 H6 Figure 4: MRI comparison map before and after HSASR. Table 3: Results of the ﬁrst three glioma grading experiments. Without HSASR With HSASR Training Testing AUC ACC (%) SEN (%) SPE (%) AUC ACC (%) SEN (%) SPE (%) GH BT 0.6850 59.93 58.45 64.00 0.9507 90.88 90.18 82.67 BT GH 0.7605 72.68 81.98 53.25 0.9146 85.90 89.87 77.63 (80%) (GH+BT)+(20%) 0.8252 78.17 85.42 61.67 0.9786 94.13 94.64 93.00 (GH+BT) (80%)GH+(20%)GH 0.8394 81.25 82.06 76.32 0.9556 89.36 87.20 94.71 (80%)BT+(20%)BT 0.8512 79.12 80.19 74.99 0.9934 95.61 97.35 92.65 Note: GH is for GliomaHPPH2018, and BTstands for BraTS2017. BTrepresents 80% of the data as training and 20% as testing. GH alone represents 80% of GH data as training and 20% of GH as testing. BT alone represents 80% of BT dataset as training and 20% of BT as testing. grid search strategy with coarse and ﬁne search. First, in search, the step size is changed to 1, data are searched from coarse search, all data samples are selected as the object of the middle number upward and downward, and the search is reference selection in this study. Each ﬁle is considered a over according to the end condition. collection, and the step size is set to the middle number of the collection (78), starting from the beginning of every collection each time. )e middle slice (78) of each collection 3.6. Feature Engineering. All the data being processed are imported into feature calculation. For each region of in- is used as the reference slice of HS, 30% of image samples are randomly selected from all data, and image enhancement is terest, 557 radiomics features are calculated through Pyr- adiomics (Pyradiomics is a tool for the computational carried out via HS and subsequently imported into the model veriﬁcation process of this study. T is set to 1%, and characteristics of medical images). In this study, 9 spatial geometric features are included, which are only calculated in the set with the highest AUC value and AUC value error less the original space. Eighteen ﬁrst-order statistical features than 1% is selected. Furthermore, the study carries out ﬁne 8 Journal of Healthcare Engineering Receiver operating characteristic curve Receiver operating characteristic curve 1.0 1.0 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 False positive rate False positive rate Train ROC AUC = 0.98571 Train ROC AUC = 0.95225 Test ROC AUC = 0.91456 Test ROC AUC = 0.95073 (a) (b) Figure 5: ROC plot. Table 4: Results of the contrast experiment. Training Testing Method AUC ACC (%) SEN (%) SPE (%) (80%) (GH+BT)+(20%) (GH+BT) HE 0.9176 86.19 90.44 76.05 (80%) (GH+BT)+(20%) (GH+BT) Handpicked 0.9447 89.90 90.35 88.88 (80%) (GH+BT)+(20%) (GH+BT) Our method 0.9786 94.13 94.64 93.00 Note: GH is for GliomaHPPH2018, and BT is for BraTS2017. GH+BT represents 80% of the data as training and 20% as testing. and 43 texture features are calculated in the original data and 3.7. Experimental Comparison Designs. To verify the eﬀec- eight wavelet decomposition spaces, respectively. For all tiveness of processed multicenter data, this study proposed image histological features, model training is conducted the following process designs: directly after the processing of missing values. (1) Take the processed GliomaHPPH2018 and )e calculated features are imported into the model BraTS2017 datasets as the training and testing sets, scriptof this studyfor prediction.)e modelscript isdivided respectively. )en, use the two unpreprocessed into two parts, namely, feature selection and classiﬁer. Five datasets as contrast experiments. selection methods and nine classiﬁers are combined to set up (2) Mix together 521 data samples from Glio- the classiﬁer, and the training and testing sets are imported maHPPH2018 and BraTS2017, preprocess the data, into the model for prediction. For the designs of experiment carry out feature selection and model training, and 2 and 3 in the next section, 80% and 20% of data are selected analyze the results. from the data table as training and testing sets, respectively. )e prediction is then repeated 10 times. )e average value (3) Use single sequences from the processed and un- of ﬁnal results is obtained, and the classiﬁer and selection processed GliomaHPPH2018 and BraTS2017 as methods with the best results are ﬁnally selected with the comparison experiment. rating device. )e selection methods used in the study are (4) Make comparisonexperiments between this research SelectKBest (f_classif) for classifying the tag features be- method and common image enhancement methods tween tasks for ANOVA f values, principal component on the mixed dataset. Finally, compare these results analysis (PCA), kernel principal component analysis with the untreated data. In addition, to verify the (KPCA), independent component correlation algorithm eﬀectiveness of HSASR, compare the research results (ICA), and factor analysis (FA). )e classiﬁers used in this with the reference diagram results selected by study are decision tree (DT), random forest (RF), bagging radiologists. (BAG), binary search tree (BSA), naive Bayes (NB), mul- tilayer perception (MLP), support vector machine (SVM), 4. Results and Discussion logistic (LR), and k-nearest neighbour (KNN). For the two separate datasets in experiment 1, the stable training model 4.1. Analysis of Image Enhancement. )e purpose of weak- is selected by using a tenfold cross-validation scheme, and ening the multicenter data is to relieve the diﬀerences between then the grading prediction is ﬁnally performed. groups and retain individual characteristics. )at is to say, True positive rate True positive rate Journal of Healthcare Engineering 9 Table 5: Performance comparison with similar works. Author and reference Architecture/method Output class Data Maximum ACC (%) Khawaldeh et al. [27] AlexNet+Preprocessing 3 130 91.16 Chen et al. [28] CAD+Preprocessing 2 274 91.27 Our method HSASR 2 523 94.13 80 80 22 9 2 26 60 15 3 58 73 LGG HGG LGG HGG Predicted lable Predicted lable (a) (b) Figure 6: Confusion matrix of mixed dataset. 0.84 0.83 0.77 0.82 FILT 0.72 0.82 0.82 0.81 0.77 0.79 0.80 PCA 0.81 0.81 0.71 0.83 0.77 0.78 0.69 0.8 0.8 0.76 KPCA 0.67 0.78 0.76 0.81 0.81 0.73 0.84 0.74 0.81 0.72 ICA 0.69 0.8 0.79 0.82 0.69 0.79 0.76 0.75 0.8 FA 0.79 0.79 0.81 0.72 0.81 0.69 0.83 0.76 0.8 0.68 DT RF BAG BST NB MLP SVM LR KNN Classifier Figure 7: Heat map without HSASR. Feature reduction True lable HGG LGG True lable HGG LGG AUC score 10 Journal of Healthcare Engineering FILT 0.9 0.97 0.97 0.97 0.95 0.98 0.98 0.96 0.97 0.96 0.85 0.94 0.9 0.98 0.95 0.97 PCA 0.96 0.97 0.95 0.93 0.87 0.96 0.93 0.97 0.97 0.96 0.97 KPCA 0.94 0.96 0.90 ICA 0.84 0.96 0.95 0.96 0.9 0.97 0.97 0.96 0.96 0.87 FA 0.85 0.95 0.94 0.96 0.88 0.97 0.96 0.97 0.95 0.84 DT RF BAG BST NB MLP SVM LR KNN Classiﬁer Figure 8: Heat map with HSASR. Experiment 1 in Table 3 presents the results of Glio- under the premise of retaining tumor morphological char- acteristics,theimagedataofdiﬀerentmedicalinstitutionshave maHPPH2018and BraTS2017 datasets as the training and similar contrast and brightness. In this study, contrast and testing set, respectively. )e results show that the data with brightness of multicenter glioma data images are signiﬁcantly HSASR is better predicted, and the indexes are signiﬁcantly improved after image enhancement in preprocessing. Figure 4 improved compared with the unprocessed data, as shown in shows the MRI results of the CET1 sequence in 12 patients. Figure 5. Figure 5(a) shows the ROC curve in which )eleft-handsideofFigure4presents thatthe ﬁrstandsecond BraTS2017 is a training set and GliomaHPPH2018 is a rows are LGG without HSASR, and the third and fourth rows testing set, and Figure 5(b) shows the ROC curve in which are HGG without HSASR. All processed glioma data are the testing set and the training set interchange. It can be shown on the right side of the ﬁgure. )e examples show on observed from the ﬁgure that when testing set and training the left and right sides had a one-to-one correspondence. set swap, the better results can also be obtained. )e results show that the data processed by the HRASR method have From the perspective of image, the range of contrast and brightness of data without HSASR in Figure 4 are relatively good adaptability and can alleviate the diﬀerence among complex and the tumor area is fuzzy and diﬃcult to distin- multicenter data images. guish. )e data after HSASR are consistent in contrast and Experiment 2 in Table 3 presents the results of the mixed brightness, the tumor area is signiﬁcantly enhanced, and basic dataset. )e calculated characteristics are imported into the features (tumor morphology and lesion range) of the original proposed grading model for 10 iterations of random pre- image can be retained. diction, and the average value is subsequently obtained. Finally, the model with the best average performance is selected as the ﬁnal model. )e results show that AUC after 4.2. Performance of Glioma Grading. )ispaperaimstoreﬂect processing generally increased by more than 15%, and other the eﬀectiveness of these methods indirectly by using model indicators also improve signiﬁcantly. Experiment 3 in Ta- prediction indicators. To validate the eﬀect of the processed ble 3 shows that the data processed by HSASR on a single multicenterdatafurther,thissectionconductstheveriﬁcationof dataset also achieve a good grading eﬀect. grading results based on the following indicators: AUC, ac- Experiment 4 in Table 4 shows the comparison exper- curacy (ACC), sensitivity (SEN), and speciﬁcity (SPE). iment conducted by three diﬀerent methods on the mixed Feature reduction AUC score Journal of Healthcare Engineering 11 Coarse search 0.980 Fine search 0.98 0.978 0.96 0.976 0.94 0.92 0.974 0.90 0.972 0.88 0.970 0.86 6(78) 12(78) 18(78) 24(78) 30(78) 36(78) 42(78) 48(78) 69(10) 71(10) 73(10) 75(10) 77(10) 79(10) 81(10) N K (a) (b) Figure 9: Diagram of the proposed method (note: the red circle and blue square represent the AUC of machine and manually selected reference images, respectively. )e black square represents optimal reference frame). dataset. )ese methods include the HE method, reference Figure 9 illustrates the experimental results of the grid frame method of manual selection by radiologists, and the search in this study. In this study, ﬁve reference slices with the method proposed in this paper. )e results show that highest performance indexes were selected from two datasets via rough search, and the reference slices with the best per- compared withthe othertwo methods, the methodproposed in this paper signiﬁcantly improves the grading eﬀect of formance were screened out. )e N axis, K, circle, and square denoted the number of samples, the number of slices in each glioma. )e reference frame selected by radiologists based on experience shows signiﬁcant improvement in glioma sampleset,the valueof AUCof thereferenceimageselectedby grading, but it can be seen from the results that the per- the HSASR, and the predicted performance value of the formance of the selected reference frame is worser than that reference image selected by the doctor, respectively. )e ob- of the automatically selected. tained performance of the reference frame selected by the Table 5 gives the performance comparison of similar method is both low and high, indicating that the diﬀerence in work in the references using glioma dataset. In [27], after reference objects had a signiﬁcant impact on the result of data preprocessing, they used a modiﬁed version of AlexNet grading. )e proposed method can replace radiologists in for classifying MR brain images into three classes like choosing the best reference frame and save a considerable amount of time. Hence, the proposed method has certain healthy brain, LGG, and HGG. However, the amount of data of this work is so small that cannot verify the robustness of application value in clinical trials. the proposed method. Reference [28] also involved data preprocessing, but our data is twice as large, and the result 5. Conclusions veriﬁed by the grading model is better than this method. Figure 6 is the confusion matrix corresponding to the Inconsistencies among data prevent multicenter data from ﬁnal model of mixed glioma data in experiment 2, and playing to its shared advantage. )is paper proposes a LGG and HGG represent the low-level label and high-level histogram speciﬁcation method with automatic selection of label, respectively. Figure 6(a) shows the confusion matrix reference frames for magnetic resonance images to alleviate before processing, and Figure 6(b) shows the predicted the problem of contrast inconsistencies among multicenter data distribution after processing. Figure 6 shows a sig- data. )e core of histogram speciﬁcation is to change the niﬁcant decrease in the number of glioma predicted in- local brightness of the image according to the reference correctly after preprocessing. Meanwhile, in order to frame, but the reference frame of traditional histogram intuitively evaluate the performance diﬀerences of the speciﬁcation is usually manually selected by radiologists. model before and after preprocessing, Figures 7 and 8 list )is method not only increases the workload of radiologists, the performance heat maps of grading prediction before but also cannot guarantee the optimal reference frame. In and after preprocessing, respectively. )e horizontal and the method of this paper, the search range is ﬁrstly narrowed vertical coordinates and corresponding results represent by coarse search intraglioma samples, then the suitable the classiﬁer, feature selection or dimension reduction reference frame in histogram is selected by ﬁne search within method, and the maximum average AUC value, re- sample selected by coarse search. Finally, the eﬀectiveness spectively. )ese data indicate that the overall perfor- and feasibility of the proposed method are veriﬁed by a mance of the data after preprocessing is about 15% higher grading experiment based on two datasets. )e results show than that without HSASR processing. that multicenter data processed by this method have good AUC AUC 12 Journal of Healthcare Engineering International Journal of Pattern Recognition and Artiﬁcial adaptability, which improves the grading results and has Intelligence, vol. 31, no. 4, Article ID 1754006, 2017. certain practical value for clinical prediction. [13] M. F. A. Hassan, A. S. A. Ghani, D. Ramachandram, A. Radman, and S. A. Suandi, “Enhancement of under-ex- Data Availability posed image for object tracking algorithm through homo- morphic ﬁltering and mean histogram matching,” Advanced )e datasets used in this paper are public dataset Science Letters, vol. 23, no. 11, pp. 11257–11261, 2017. (BraTS2017) and Henan Provincial People’s Hospital [14] R. Berenguer, P. J. Mdr, z. J. Canales-Vazquez ´ et al., (GliomaHPPH2018) dataset; BraTS2017 can be obtained “Radiomics of CT features may be nonreproducible and re- through the following URL: https://www.med.upenn.edu/ dundant: inﬂuence of CTacquisition parameters,” Radiology, sbia/brats2017/data.html. vol. 288, no. 2, Article ID 172361, 2018. [15] H. Schnack, N. van Haren, M. Picchioni et al., “A multicenter mri schizophrenia study: reliability of voxel-based mor- Conflicts of Interest phometry,” Schizophrenia Research, vol. 102, no. 1–3, p. 75, )e authors declare that they have no conﬂicts of interest. [16] L. G. Nyul, J. K. Udupa, and X. Xuan Zhang, “New variants of a method of MRI scale standardization,” IEEE Transactions on Acknowledgments Medical Imaging, vol. 19, no. 2, pp. 143–150, 2000. [17] S. Bakas, H. Akbari, A. Sotiras et al., “Advancing the Cancer )is work was supported by the National Natural Science Genome Atlas glioma MRI collections with expert segmen- tation labels and radiomic features,” Scientiﬁc Data, vol. 4, Foundation of China under grant 81772009 and Scientiﬁc Article ID 170117, 2017. and Technological Research Project of Henan Province [18] B. H. Menze, A. Jakab, S. Bauer et al., “)e multimodal brain under grant 182102310162. tumor image segmentation benchmark (BRATS),” IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993– References 2024, 2015. [19] G. Chen, L. Li, W. Jin, M. Liu, and F. Shi, “Image contrast [1] Q. T. Ostrom, H. Gittleman, J. Xu et al., “CBTRUS statistical enhancement method based on display and human visual report: primary brain and other central nervous system tu- system characteristics,” Applied Optics, vol. 58, no. 7, mors diagnosed in the United States in 2009-2013,” Neuro- pp. 1813–1823, 2019. Oncology, vol. 18, no. 5, pp. v1–v75, 2016. [20] C.-M. Chen, C.-C.Chen,M.-C. Wu etal., “Automatic contrast [2] D. N. Louis, H. Ohgaki, O. D. Wiestler et al., “)e 2007 WHO enhancement of brain MR images using hierarchical corre- classiﬁcation of tumours of the central nervous system,” Acta lation histogram analysis,” Journal of Medical and Biological Neuropathologica, vol. 114, no. 5, p. 547, 2007. Engineering, vol. 35, no. 6, pp. 724–734, 2015. [3] W. Shan and X.-L. Wang, “Clinical application value of 3.0T [21] X. Wang and L. Chen, “An eﬀective histogram modiﬁcation MR diﬀusion tensor imaging in grade diagnosis of gliomas,” scheme for image contrast enhancement,” Signal Processing: Oncology Letters, vol. 14, no. 2, pp. 2009–2014, 2017. Image Communication, vol. 58, pp. 187–198, 2017. [4] V. Kumar, Y. Gu, S. Basu et al., “Radiomics: the process and [22] M. Paulinas and A. Usinskas, “A survey of genetic algorithms the challenges,” Magnetic Resonance Imaging, vol. 30, no. 9, applications for image enhancement and segmentation,” In- pp. 1234–1248, 2012. formation Technology and Control, vol. 36, no. 3, pp. 278–284, [5] R. Rohan, H. Adam, B. Jason, R. Robert, and S. Daniel, “Outcomes in reoperated low-grade gliomas,” Neurosurgery, [23] A. Draa and A. Bouaziz, “An artiﬁcial bee colony algorithm vol. 77, no. 2, p. 175, 2015. for image contrast enhancement,” Swarm and Evolutionary [6] E. Ferrante and N. Paragios, “Slice-to-volume medical image Computation, vol. 16, pp. 69–84, 2014. registration: a survey,” Medical Image Analysis, vol. 39, [24] L. Rundo, A. Tangherloni, M. S. Nobile, C. Militello, and pp. 101–123, 2017. P. Cazzaniga, “MedGA: a novel evolutionary method for [7] Y. Zhu, “Review of image enhancement algorithms,” China image enhancement in medical imaging systems,” Expert Computer and Communication, vol. 10, no. 4, pp. 438–448, Systems with Applications, vol. 119, 2018. [25] S.-D. Chen and A. Rahman Ramli, “Preserving brightness in [8] L. Zhuang and Y. Guan, “Image enhancement via subimage histogram equalization based contrast enhancement tech- histogram equalization based on mean and variance,” Com- niques,” Digital Signal Processing, vol. 14, no. 5, pp. 413–428, putational Intelligence and Neuroscience, vol. 2017, no. 2, Article ID 6029892, 2017. [26] N. Sengee, A. Sengee, and H. K. Choi, “Image contrast en- [9] M. F. Khan, E. Khan, and Z. A. Abbasi, “Image contrast hancement using Bi-histogram equalization with neighbor- enhancement using normalized histogram equalization,” hood metrics,” IEEE Transactions on Consumer Electronics, Optik, vol. 126, no. 24, pp. 4868–4875, 2015. vol. 56, no. 4, pp. 2727–2734, 2011. [10] M. Ramkumar and B. Karthikeyan, “A survey on image [27] S. Khawaldeh, U. Pervaiz, A. Raﬁq, and R. Alkhawaldeh, enhancement methods,” International Journal of Engineering “Noninvasive grading of glioma tumor using magnetic res- and Technology, vol. 5, no. 2, pp. 960–962, 2013. onance imaging with convolutional neural networks,” Applied [11] M. Kanmani and V. Narsimhan, “An image contrast en- Sciences, vol. 8, no. 1, p. 27, 2017. hancement algorithm for grayscale images using particle [28] W. Chen, B. Liu, S. Peng, J. Sun, and X. Qiao, “Computer- swarm optimization,” Multimedia Tools and Applications, aided grading of gliomas combining automatic segmentation vol. 77, no. 3, pp. 1–17, 2018. and radiomics,” International Journal of Biomedical Imaging, [12] Z. Jing, G. Chen, and Z. Jia, “An image stitching algorithm vol. 2018, Article ID 2512037, 11 pages, 2018. based on histogram matching and SIFT algorithm,” International Journal of Advances in Rotating Machinery Multimedia Journal of The Scientific Journal of Engineering World Journal Sensors Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Control Science and Engineering Advances in Civil Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com Journal of Journal of Electrical and Computer Robotics Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 VLSI Design Advances in OptoElectronics International Journal of Modelling & Aerospace International Journal of Simulation Navigation and in Engineering Engineering Observation Hindawi Hindawi Hindawi Hindawi Volume 2018 Volume 2018 Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com www.hindawi.com www.hindawi.com Volume 2018 International Journal of Active and Passive International Journal of Antennas and Advances in Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal

Journal of Healthcare Engineering – Hindawi Publishing Corporation

Published: Dec 19, 2019

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Automatic Histogram Specification for Glioma Grading Using Multicenter Data

Automatic Histogram Specification for Glioma Grading Using Multicenter Data

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Automatic Histogram Specification for Glioma Grading Using Multicenter Data

Automatic Histogram Specification for Glioma Grading Using Multicenter Data

References (38)

Abstract

Journal

Recommended Articles

There are no references for this article.

Our policy towards the use of cookies