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

Learn More →

Preoperative ultrasound radiomics analysis for expression of multiple molecular biomarkers in mass type of breast ductal carcinoma in situ

Preoperative ultrasound radiomics analysis for expression of multiple molecular biomarkers in... Background: The molecular biomarkers of breast ductal carcinoma in situ (DCIS) have important guiding signifi- cance for individualized precision treatment. This study was intended to explore the significance of radiomics based on ultrasound images to predict the expression of molecular biomarkers of mass type of DCIS. Methods: 116 patients with mass type of DCIS were included in this retrospective study. The radiomics features were extracted based on ultrasound images. According to the ratio of 7:3, the data sets of molecular biomarkers were split into training set and test set. The radiomics models were developed to predict the expression of estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), Ki67, p16, and p53 by using com- bination of multiple feature selection and classifiers. The predictive performance of the models were evaluated using the area under the curve (AUC) of the receiver operating curve. Results: The investigators extracted 5234 radiomics features from ultrasound images. 12, 23, 41, 51, 31 and 23 fea- tures were important for constructing the models. The radiomics scores were significantly (P < 0.05) in each molecular marker expression of mass type of DCIS. The radiomics models showed predictive performance with AUC greater than 0.7 in the training set and test set: ER (0.94 and 0.84), PR (0.90 and 0.78), HER2 (0.94 and 0.74), Ki67 (0.95 and 0.86), p16 (0.96 and 0.78), and p53 (0.95 and 0.74), respectively. Conclusion: Ultrasonic-based radiomics analysis provided a noninvasive preoperative method for predicting the expression of molecular markers of mass type of DCIS with good accuracy. Keywords: DCIS, Molecular biomarkers, Radiomics, Ultrasound Background potential to further develop into breast invasive can- Breast ductal carcinoma in  situ (DCIS) is a kind of cer [3]. The clinical treatments of patients with DCIS malignant tumor originated in the ductal epithelial tis- include surgical resection, radiotherapy, chemotherapy sue, limited to the basement membrane [1]. DCIS is and endocrine therapy, in which surgical resection the second most common breast tumor, and accounts includes simple focal resection and mastectomy, with for approximately 20–30% [2]. Some DCIS had the different therapeutic effects [4 ]. Although the progno- sis of DCIS is good, more than 14% of DCIS patients may develop invasive cancer without treatment within *Correspondence: heyun@stu.gxmu.edu.cn; yanghong@gxmu.edu.cn 10 years [5]. In the past 10 years, the incidence of DCIS Linyong Wu and Yujia Zhao have contributed equally to this work has gradually increased, highlighting the understand- Department of Medical Ultrasound, The First Affiliated Hospital ing the importance of DCIS pathology [6]. However, of Guangxi Medical University, Nanning, Guangxi Zhuang Autonomous Region 530021, People’s Republic of China the pathologic mechanism of the transition from DCIS Full list of author information is available at the end of the article © The Author(s) 2021. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons. org/ publi cdoma in/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Wu et al. BMC Med Imaging (2021) 21:84 Page 2 of 14 to invasive carcinoma is still unclear, which produces provide dynamic and accurate evaluation of biological clinical challenges of overdiagnosis and overtreatment behavior information for the clinic. in patients with DCIS [7]. Therefore, the investiga - With the breakthrough of imaging technology, mam- tors thought more studies were need to understand mography is an important examination method for the potential of the pathological process of DCIS, in DCIS, which is sensitive to the detection of calcification order to adapt to the current individualized, refined [14]. Ultrasound (US) has become the main examina- treatment. tion technology to detecting breast lesions [15], which Immunohistochemistry (IHC) can reflect the expres - is real-time, dynamic and non-invasive. There are two sion of molecular biomarkers in tumor tissue, which can types of DCIS: mass and non-mass. Some studies sug- further clarify the biological behaviors of tumors. The gested that the detection rate of US in 93 patients with expression of different molecular biomarkers can lead to mass type DCIS reached 77.4% [16]. The main character - different biological behaviors and treatments. Some stud - istics of DCIS in ultrasound were: uneven low or slightly ies have shown that some molecular biomarkers were low echo, irregular shape, unclear borders, parallel skin, important indicators for predicting biological behavior weakened posterior echo, calcification, and some blood and judging follow-up treatment in patients with DCIS, flow signals [17, 18]. However, mammography or US- such as estrogen receptor (ER), progesterone receptor assisted screening could increase overdiagnosis because (PR), human epidermal growth factor receptor 2 (HER2), both tests primarily detect low-grade invasive cancers Ki67, p16, and p53. ER and PR are the earliest molecular [19]. At the same time, radiologists are very time-con- biomarkers of breast cancer. They are predictors of breast suming to accumulate experience and have strong per- cancer prognosis and endocrine adjuvant therapy [8]. sonal subjectivity, which is another problem that needs HER2 is a proto-oncogene, which is mainly involved in to be solved. There is an urgent need for more advanced tumor signal transduction and cell proliferation. Its posi- imaging evaluation methods to guide the diagnosis and tive expression can lead to a high distant metastasis rate treatment of DCIS. and poor prognosis of breast cancer. Ki67 is an antigenic Breast lesions are diagnosed and screened by various nuclear protein that can be used as a proliferation marker. imaging methods, such as mammography, US, and mag- Its high expression is considered to be a biomarker of netic resonance imaging (MRI). All three examination tumor invasion [9]. Ki67 has a good application prospect have some limitations [20]. Radiomics is a hot subject of in predicting endocrine therapy response of breast can- artificial intelligence that is applied in the medical imag - cer [10]. Defined as a tumor suppressor gene, p16 is con - ing field, which is the cornerstone of precision science sidered to be an important cell cycle regulator [11]. P16 is in the future. Radiomics is defined as the extraction of closely related to abnormal methylation initiation. P53 is high-throughput features from single or multiple medi- a common tumor suppressor gene. Impaired function of cal image patterns to select features that are closely asso- p53, such as p53 mutation, can lead to uncontrolled pro- ciated with tumors,and the ultimate goal is to construct liferation of damaged cells [12]. Therefore, accurate iden - prediction models based on features to provide accurate tification of the expression of molecular biomarkers can tumor phenotypic analysis information and accurate help stratify tumor risk and facilitate the development of treatment decision-making [21]. Radiomics highlights personalized and accurate treatment plans. the image features that are not visible to the naked eye, Currently, the preoperative evaluation of the molecu- thus significantly enhancing the predictive power of lar biomarkers of DCIS mainly depends on IHC detec- medical imaging [22]. Radiomics has been developed tion after biopsy. However, because the progression of in a wide range of fields, such as disease diagnosis and tumors are dynamic process, there are differences in biological behavior judgment. For example, the US- spatio-temporal evolution. In addition, the evaluation radiomics model developed by Luo WQ et al. had better results of a few tissue biopsies do not necessarily repre- performance in distinguishing breast lesions than breast sent the expression of the molecular biomarkers of the imaging reporting and data system (BI-RADS) [23]. Lin whole tumor [13]. Invasive procedures and potential risks F et al. found that the radiomics score was more effective limit its multiple applications in monitoring tumor pro- than the clinical radiological model in benign and malig- gression and biological behavior. However, the preopera- nant breast lesions (< 1  cm) [24]. These series of studies tive monitoring of molecular biomarkers can dynamically showed that radiomics had better performance than tra- identify the progression of tumors and the changes in ditional imaging features in the diagnosis of breast dis- biological behavior, which has great significance for the eases to some extent. accurate formulation of treatment plans and the evalu- u Th s, this retrospective study intended to further clar - ation of curative effects. To avoid overdiagnosis and ify the relationship between US-radiomics and molecular overtreatment of patients with DCIS, it is necessary to markers of DCIS. Radiomics models had been developed Wu  et al. BMC Med Imaging (2021) 21:84 Page 3 of 14 to noninvasively evaluate the expression of molecular radiologists with five years of working experience manu - markers to help achieve accurate risk stratification and ally delineated the region of interest (ROI) of the lesions. treatment for patients with DCIS. The radiologists disregarded the diagnosis and patho - logical results of the patients [26]. After the discussion, when there was a big difference between the two radiolo - Materials and methods gists, the third radiologist with 10 years of experience re- Study cohort examined and confirmed the final boundary. This process Clinical data of 400 patients with DCIS who were path- provided reliable DCIS area contours and ensured the ologically confirmed by surgery were retrospectively accuracy of feature extraction. analyzed by the investigators. The data were based on the pathology reports of the first affiliated hospital of Guangxi medical university from January 2015 to July Image pre‑processing and feature extraction 2020. Further inclusion and exclusion criteria for this The Intelligence Foundry software (version 1.3, GE cohort study were as follows. Inclusion criteria: (1) pri- Healthcare, Shanghai, China) was applied for radiomics mary breast DCIS, (2) IHC results of molecular biomark- analysis. Figure  2 summarized the main flow of radiom - ers; and (3) preoperative US data within one month. ics analysis. The software relied on algorithms provided Exclusion criteria: (1) non-mass DCIS, including mani- by the Pyradiomics package that comply with the image festations of ductal dilation, diffuse calcification, and dif - biomarker standardization initiative (IBSI, version fuse distribution of lesions; (2) unclear image of target 2016) [27]. Features were automatically calculated and lesions; (3) secondary DCIS or postoperative recurrence extracted by the Pyradiomics extractor. The maximum of DCIS; (4) preoperative treatment history of radio- number of features extraction of the software was 5234, therapy, chemotherapy and traditional Chinese medicine; including: 122 original, 48 intraperinodular textural and (5) lack of clinical data. transition (ipris), 468 co-occurrence of local anisotropic This study finally enrolled a total of 116 patients with gradient orientations (CoLIAGe), 432 wavelets + local DCIS. The IHC (ER, PR, HER2, Ki67, p16, and p53) binary pattern (LBP), 2,944 shearlets, 1,080Gabors, 80 conformed to the diagnositic criteria of the department phased congruency-based local binary pattern (PLBP) pathology in this hospital, and were classified as positive and 60 wavelet-based improved local binary pattern or negative. The IHC results of Ki67 were positive or high (WILBP) features  (Additional file  1). Before feature expression (Ki67 > = 14%) and negative or low expres- extraction, the images were pre-processed: the gray sion (Ki67 < 14%) [25]. The number of patients enrolled value of the image was discretized with a bin size of 256, for each molecular biomarker were listed as follows: 112 and the original features were extracted. The features of cases (ER), 109 cases (PR), 94 cases (HER2), 107 cases wavelets + LBP, Shearlets, Gabors, PLBP and WILBP (Ki67), 74 cases (p16), and 116 cases (p53) (Fig. 1). were extracted by wavelet transform, shearlet transform and garber operator transform on the gray value matrix Image collection and tumor segmentation of the original images, respectively [28] (Fig. 3). Each US radiologist involved in image collection had over 5  years of experience in the field of breast. Before collecting US data, all radiologists were strictly trained. Data grouping and data cleaning GE Logiq E9 (GE Healthcare, United States), Aloka To balance the initial distribution of data, each sub-data EZU-MT28-S1 (Aloka, Japan) and MYLAB CLASS C set was randomly split into training set and test set in a (MYLAB, Italy) medical ultrasound diagnositic instru- ratio of 7:3. Based on the difference in the image extrac - ments were utilized for image collection. The breast tion feature quantization caused by different medical probe was selected, and the frequency was set to ultrasound diagnositic instruments and parameters, the 7–14  MHz. The patients took the supine position, put combat method was employed to solve this problem. The their hands on the head, and fully exposed the breast combat method could be used to coordinate and correct area and armpits on both sides. The lesions were scanned the differences between different machines and different from multiple angles, and the largest clear image of the center images. Some studies had applied this method to lesions were selected. The following ultrasonic charac - the MRI images [29]. In addition, the median value of the teristics of the lesions were recorded: BI-RADS classi- feature quantization value was applied to fill the miss - fication, location, size, shape, boundary, internal echo, ing sample. The min–max normalization method was calcification, posterior echo changes, ductal dilatation, employed to normalize the feature data to improve the blood flow signal distribution and axillary lymph nodes. comparability between features. It converted the original These images were imported into the ITKSNAP soft - data to the range of [0, 1] by linearization, which realized ware (version 3.8.0). To avoid subjective compliance, two the proportional scaling of the original data. Wu et al. BMC Med Imaging (2021) 21:84 Page 4 of 14 Fig. 1 Study cohort. a Workflow of study cohort inclusion. b Up-set plot of the expression of molecular markers shared between different samples Feature importance importance of features for subsequent model construct- The purpose of the study is difficult to explain with ing. Multiple combination techniques were applied to thousands of radiomics features of high-dimensional explain the importance of features: First, Spearman data. Feature importance analysis helps to explain the correlation coefficient test was used to eliminate high Wu  et al. BMC Med Imaging (2021) 21:84 Page 5 of 14 Fig. 2 Workflow of radiomics analysis correlation features with threshold values (0.75, 0.85, vector machine (SVM)) were employed to predict the 0.95). This test was a statistical index to measure the expression levels of the molecular biomarkers of DCIS correlation between two variables. Three dimension - [33, 34], and the score of each model was calculated. ality reduction methods (least absolute shrinkage and In addition, the fivefold cross-validation method was selection operator (LASSO) [30], random forests (RF) explored to improve the accuracy of the models. The test [31], and support vector machine-recursive feature set was used to evaluate the reliability of the models. elimination (SVM-RFE) [32]) separated or jointed sta- To accurately evaluate the predictive ability of radi- tistical tests for selecting the important features. In the omics models, the receiver operating curve (ROC), the statistical test, if the data accorded with the normal dis- area under the curve (AUC), accuracy (ACC), precision tribution, the t-test was adopted; otherwise, the Mann– (PREC), sensitivity (Sn) and specificity (Sp) were adopted Whitney U test was adopted. for the evaluation. The closer the AUC was to 1, the higher the diagnostic efficiency was. In this study, only Predictive radiomics models the best classification results of the classifier were shown. Machine learning algorithms were developed based on Python environment. Five machine-learning-based clas- sifiers (decision tree (DT), k-nearest neighbors (KNN), logistics regression (LR), naive Bayes (NB), and support Wu et al. BMC Med Imaging (2021) 21:84 Page 6 of 14 Fig. 3 The process of quantifying features. a Delineation of the ROIs. b Gray level co-occurrence matrix (GLCM), run length matrices (RLM), and histogram feature extraction. c The classification of 5234 features with p16-positives; 34 patients with p53-negatives and 82 Results patients with p53-positives. Patient characteristics and molecular biomarkers of interest Radiomics analysis The mean age of all patients was 48.8 ± 11.1 years, and the The correlation clustering heatmaps among 5234 fea - age range was 29–84  years. The characteristics parame - tures of each molecular biomarkers (ER, PR, HER2, ters were shown in Table 1. The ultrasonographic features Ki67, p16, and p53) were shown in Fig.  4. A list of of the patients were similar to those reported in the lit- 18 feature importance methods were obtained, and eratures. The expression of ER, PR, HER2, Ki67, p16, and the combination feature selection methods for opti- p53 were as follows: 49 patients with ER-negatives and mal modeling results of each molecular biomark- 63 patients with ER-positives; 53 patients with PR-neg- ers were as follows: Spearman0.75 + Statistical atives and 59 patients with PR-positives; 36 patients with Test + RF, Spearman0.75 + Statistical Test + RF, Spear- HER2-negatives and 58 patients with HER2-positives; 45 man0.75 + L ASSO, Spearman0.75 + Statistical patients with Ki67-negatives and 62 patients with Ki67- Test + RF, Spearman0.75 + Statistical Test + SVM-RFE , positives; 29 patients with p16-negatives and 45 patients Wu  et al. BMC Med Imaging (2021) 21:84 Page 7 of 14 Table 1 Patient characteristics and molecular biomarkers of interest Parameters N = 116 Parameters N = 116 Median age (years) 48.8 ± 11.1 Shape rule (yes/no) 26/96 Immunohistochemistry Clear boundary (yes/no) 50/66 ER (−/+/NA) 49/63/4 Aspect ratio (< 1/ > = 1) 6/110 PR (−/+/NA) 53/56/7 Echo uniformity (yes/no) 19/97 HER2 (−/+/NA) 36/58/22 Calcification (yes/no) 69/47 Ki67 (−/+/NA) 45/62/9 Intrafocal blood flow (yes/no) 79/37 P16 (−/+/NA) 29/45/42 Peripheral blood flow (yes/no) 32/84 P53(−/+/NA) 34/82/0 Catheter dilatation (yes/no) 9/107 Ultrasonic characteristics lymph nodes (< 1/ > = 1) 93/23 Median size (cm) 2.6 ± 1.6 BI-RADS classification (3/4a/4b/4c/5/6) 12/31/30/23/9/11 Fig. 4 Correlation cluster analysis of 5234 radiomics features. The Pearson correlation test was used to analyze the correlation between features, and the "pheatmap" R software package was applied to draw heat maps. a ER; b PR; c HER2; d Ki67; e p16; f p53 Ninety models were obtained by constructing predic- and Spearman0.85 + SVM-RFE. 12 features, 23 fea- tion models with five classifiers, and the performance tures, 41 features, 20 features, 31 features and 23 fea- of the models were presented in Fig.  6  (Additional tures were important for constructing prediction file  2). The optimal radiomics models were constructed models. The heatmaps of the model features were pre - by DT, SVM, KNN, SVM, KNN and KNN classifiers, sented in Fig. 5. Wu et al. BMC Med Imaging (2021) 21:84 Page 8 of 14 Fig. 5 Important features for each molecular biomarkers. a ER; b PR; c HER2; d Ki67; e p16; f p53 respectively, and showed above moderate predictive Discussion performance in predicting the expression of molecular This study was the first non-invasive comprehensive markers of DCIS (Table  2). Radiomics scores of train- analysis based on US-radiomics to predict the expres- ing set and test set were significantly different in each sion of molecular markers of DCIS. The investigators molecular marker expression (training set, P < 0.001, recruited only 116 patients with DCIS for this study, but test set, P < 0.05). The predictive performance of the it was exciting to see that the radiomics models showed radiomics models of each molecular biomarker in the more than moderate predictive performance in predict- training set: ER (AUC, 0.94, 95% confidence interval ing molecular biomarker expression of DCIS. (CI) 0.89–0.99), PR (AUC, 0.90, 95% CI 0.83–0.97), DCIS is a malignant tumor with good prognosis, but HER2 (AUC, 0.94, 95% CI 0.89–0.99), Ki67 (AUC, it is heterogeneous in morphology and genetics. Before 0.95, 95% CI 0.90–0.99), p16 (AUC, 0.96, 95% CI 0.91– the imaging examination was performed, the diagnosis of 1.00), p53 (AUC, 0.95, 95% CI 0.90–0.99), respectively DCIS was only due to the appearance of nipple discharge (Fig.  7). The calibration curve of the prediction models and/or palpable mass symptoms, which accounted for in the training set confirmed the better consistency of only 2% of DCIS detected. It showed that DCIS with hid- the models (Fig.  8). The radiomics models showed pre - den symptoms were easily missed [35]. With the screen- dictive performance with AUC greater than 0.7 in the ing of imaging technology (mammography, US and MRI), test set: ER (AUC, 0.84, 95% CI 0.68–0.99), PR (AUC, the detection rate of DCIS had gradually increased. This 0.78, 95% CI, 0.60–0.96), HER2 (AUC, 0.74, 95% CI detection rate included symptomatic DCIS, and whether 0.74–0.99), Ki67 (AUC, 0.86, 95% CI 0.67–0.97), p16 there was overdiagnosis in the detection of insidious (AUC, 0.78, 95% CI 0.59–0.97), p53 (AUC, 0.74, 95% CI DCIS was also a hot topic of controversy [36], Unfortu- 0.55–0.93), respectively (Fig. 9). nately, the diagnosis of DCIS marked women as at risk of invasive breast cancer, so women diagnosed with DCIS may suffer serious psychological distress, leading to the progression of DCIS [37]. In addition, the current Wu  et al. BMC Med Imaging (2021) 21:84 Page 9 of 14 Fig. 6 Heat maps of evaluation indicators for ninety radiomics prediction models. a ER; b PR; c HER2; d Ki67; e p16; f p53 Table 2 Evaluation of radiomics models in each DCIS molecular biomarkers Training set Test set AUC ACC PREC Sn Sp AUC ACC PREC Sn Sp ER 0.94 0.90 0.93 0.89 0.91 0.84 0.82 0.81 0.90 0.73 PR 0.90 0.84 0.89 0.80 0.89 0.78 0.76 0.80 0.71 0.8 HER2 0.94 0.88 0.90 0.90 0.84 0.74 0.72 0.78 0.78 0.64 Ki67 0.95 0.88 0.84 0.98 0.74 0.86 0.76 0.79 0.79 0.71 p16 0.96 0.90 0.90 0.94 0.85 0.78 0.70 0.77 0.71 0.67 p53 0.95 0.89 0.91 0.93 0.79 0.74 0.74 0.83 0.80 0.60 treatment methods were also facing the controversy with the progression of DCIS to invasive cancer [40]. For over the treatment of some patients [38]. Therefore, the example, Zhang GJ et  al. [41] found that 79% of DCIS main clinical challenge in DCIS has been to distinguish patients were positive for P53 when studying the occur- between patients who have a better chance of develop- rence and development of breast cancer. Davis et al. [42] ing invasive cancer and require more treatment and demonstrated that high Ki67 expression was an inde- those who are less likely to develop DCIS and need less pendent predictor of postoperative recurrence in patients or no treatment [39]. Immunohistochemical markers can with DCIS. Cornfield DB had found a higher recurrence explain the changes in the biological behaviors of tumors rate with PR > 3.5% using tree structure survival [43]. The on the molecular level. More and more studies have results showed that the changes in the biological behavior pointed out the changes in molecular markers associated Wu et al. BMC Med Imaging (2021) 21:84 Page 10 of 14 Fig. 7 Performance of the radiomics models in the training set. a ER; b PR; c HER2; d Ki67; e p16; f p53 Fig. 8 Calibration curves of the radiomics models in the training set. The oblique dashed line represents the perfect prediction of the ideal model. The solid line represents the performance of the radiomics model, and the dotted line near the diagonal indicates a better prediction. a ER; b PR; c HER2; d Ki67; e p16; f p53 Wu  et al. BMC Med Imaging (2021) 21:84 Page 11 of 14 Fig. 9 Performance of the radiomics models in the test set. a ER; b PR; c HER2; d Ki67; e p16; f p53 of DCIS were closely related to the expression of molecu- and quantitative analysis of the imaging features of the lar biomarkers. whole imaging mass. Therefore, radiomics is a precision About various imaging technologies, they also have medical method for non-invasive diagnosis, evaluation application limitations [44]. Mammography is the main of efficacy, biological behavior [50]. Currently, radiom - method of early breast cancer detection, but it is closely ics mainly relies machine learning algorithms to identify related to the density of the lesion and the possibility of meaningful features of image training data set, and for covering the lesion [45]. However, Chinese women have further interpretation of the information and the opti- dense breasts, so they had certain limitations in finding mization, so as to accurately predict the content of the suspicious lesions in the dense tissues of breasts through research. An independent data set is applied to test the mammography [46, 47]. The traditional mammography universality of the model and provides feedback for fur- diagnosis method will cause trauma to the patient to a ther optimization of the model [51]. To a certain extent, certain extent and reduce the patient’s treatment com- it improves the utilization of image information and ena- pliance. Due to its high sensitivity to soft tissues, US can bles differential diagnosis of diseases on more subtle lev - better show lesions in dense glands and has become the els that cannot be recognized by the naked eye. primary imaging method for Chinese women to screen Breast radiomics studies are mostly applied to the and diagnose breast diseases. For non-mass DCIS, US is prediction of the molecular classification, lymph node difficult to recognize [48]. Therefore, this retrospective metastasis and molecular markers of invasive ductal study only examined mass DCIS, which is a limitation of carcinoma. For example, Demircioglu A et al. [52] con- the study. MRI has considerable advantages in detecting structed radiomics models for predicting Ki67 expres- breast lesions, but its specificity is limited by several fac - sion in invasive breast cancer based on eight features tors that affect image quality, such as magnetic field and extracted from MRI images, with an AUC of 0.81. gradient strength, coil performance, contrast agent effi - Zhou et al. [53] explored the significance of MRI-radi- cacy and menstrual cycle [49]. omics models for predicting the expression of HER2 Radiomics mainly studies the quantitative features that in patients with invasive breast cancer before surgery; are related to biology in medical images. Radiomics fea- the validation set AUC reached 0.81. There are few tures are considered the invisible tissue infrastructure reports on DICS with radiomics. However, there are components of the object to be imaged, which can serve clinical challenges in the diagnosis and treatment of as a valuable method for studying cancer by imaging, patients with DCIS. Tumor progression and treatment such as MRI. Radiomics can provide in vivo visualization decisions are affected by multiple tumor molecular Wu et al. BMC Med Imaging (2021) 21:84 Page 12 of 14 biomarkers, which require to comprehensively ana- Conclusion lyze and evaluate the molecular biomarkers of DCIS. The application of machine learning-based radiomics To expand the application of radiomics in DCIS, the analysis provided a non-invasive method for predict- investigators carried out this study to assess the fea- ing the expression of multiple molecular biomarkers in sibility of molecular biomarkers of DCIS. The investi- DCIS, with good prediction performance. This study gators believe that information obtained from multiple also demonstrated the potential of radiomics in patho- molecular biomarkers can help explain the underlying logic assessment and individualized precision therapy. pathological process of DCIS. In this study, the first highlight, the first comprehen- Abbreviations sive analysis of molecular markers of DCIS was con- DCIS: Ductal carcinoma in situ; ER: Estrogen receptor; PR: Progesterone recep- ducted based on radiomics. Second highlight, there tor; HER2: Human epidermal growth factor receptor-2; AUC : Area under the curve; CI: Confidence interval; IHC: Immunohistochemistry; US: Ultrasound; were thousands of radiomics features, including eight MRI: Magnetic resonance imaging; BI-RADS: Breast imaging reporting and data classifications: original feature can reflect the number system; ROI: Region of interest; CoLIAGe: Co-occurrence of local anisotropic of voxels in the images, intensity distribution, pixel gradient orientations; LBP: Local binary pattern; PLBP: Phased congruency- based local binary pattern; WILBP: Wavelet-based improved local binary pair frequency, image average gray value, size and pattern; GLCM: Gray level co-occurrence matrix; RLM: Run length matrices; shape of ROI (https:// pyrad iomics. readt hedocs. io/ LASSO: Least absolute shrinkage and selection operator; RF: Random forests; en/ latest/ featu res. html); Ipris features capture nodu- SVM-RFE: Support vector machine-recursive feature elimination; DT: Decision tree; KNN: K-nearest neighbors; LR: Logistics regression; NB: Naive Bayes; SVM: lar heterogeneity and differential growth patterns; Support vector machine; ROC: Receiver operating curve; ACC : Accuracy; PREC: CoLIAGe features can distinguish disease phenotypes Precision; Sn: Sensitivity; Sp: Specificity. that have similar morphologic appearances [54]; wave- lets features represent most of the edge information Supplementary Information in images; Shearlet features are better for processing The online version contains supplementary material available at https:// doi. high-dimensional signals; Gabors features extract the org/ 10. 1186/ s12880- 021- 00610-7. edge and gradient information of image and reflect the spatial frequency feature; PLBP, and WILBP: PLBP Additional file 1. 5234 radiomics features matrix file. features are an oriented local texture descriptor that Additional file 2. Modeling matrix file for molecular biomarkers. combines the phase congruency approach with the LBP. The third highlight of this study was to construct Acknowledgements dozens of prediction models by combining multiple Not applicable. classifiers with multiple feature selection to select the Authors’ contributions optimal prediction results. RF and SVM-RFE had sig- Conceptualization: HQ and YL; Methodology: PL, DW and XL; Formal analysis nificant performance in feature selection of multiple and investigation: LW, YZ, PL, HQ, YL, DW, XL, YH, HY; Writing—original draft preparation: LW and YZ; Writing—review and editing: LW and YZ; Resources: molecular markers, KNN and SVM classification per- HY and YH; Supervision: HY and YH. All authors read and approved the final formed well too. Finally, through the verification of the manuscript. test set, the prediction models all showed moderate Funding performance. Not applicable. There were also some shortcomings in our study. First, this retrospective study had the problem of small Availability of data and materials The datasets supporting the conclusions of this article were included within sample size. It was necessary to increase the sample the article and its additional files. size or multi-center cooperation to construct univer- sal models. Second, when the radiologists manually Declarations delineated the ROIs, there were a certain degree of subjectivity to the contours of the lesions, which may Ethics approval and consent to participate This retrospective breast DCIS study was approved by the ethics committee lead to poor robustness of the models. In addition, the of the First Affiliated Hospital of Guangxi Medical University. Informed consent delineation process was done by only one radiologist. was waived. This study on the implementation of all procedures are in line Third, the investigators only investigated the features with the National Research Council of moral standards. extracted from the largest section, which could not Consent for publication represent the whole tumor. Due to the limitations of Not applicable. US, it was not possible to conduct three-dimensional Competing interests studies similar to other imaging studies. The authors declare that they have no competing interests. Wu  et al. BMC Med Imaging (2021) 21:84 Page 13 of 14 Author details 17. Li W, Zhou Q, Xia S, Wu Y, Fei X, Wang Y, Tao L, Fan J, Zhou W. Application Department of Medical Ultrasound, The First Affiliated Hospital of Guangxi of contrast-enhanced ultrasound in the diagnosis of ductal carcinoma Medical University, Nanning, Guangxi Zhuang Autonomous Region 530021, in situ: analysis of 127 cases. J Ultrasound Med. 2020;39(1):39–50. People’s Republic of China. GE Healthcare, Shanghai, People’s Republic 18. Moon HJ, Kim EK, Kim MJ, Yoon JH, Park VY. Comparison of clinical and of China. pathologic characteristics of ductal carcinoma in situ detected on mam- mography versus ultrasound only in asymptomatic patients. Ultrasound Received: 4 December 2020 Accepted: 21 April 2021 Med Biol. 2019;45(1):68–77. 19. Evans A, Vinnicombe S. Overdiagnosis in breast imaging. Breast. 2017;31:270–3. 20. Conti A, Duggento A, Indovina I, Guerrisi M, Toschi N. Radiomics in breast cancer classification and prediction. Semin Cancer Biol. 2020;1:323–4. 21. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, References they are data. Radiology. 2016;278(2):563–77. 1. Wellings S, Jensen HJJI. On the origin and progression of ductal carci- 22. Pinker K, Chin J, Melsaether AN, Morris EA, Moy L. Precision medicine and noma in the human breast. Origin Breast Carcinoma. 1973;50(5):1111–8. radiogenomics in breast cancer: new approaches toward diagnosis and 2. Liu Y, West R, Weber JD, Colditz GA. Race and risk of subsequent treatment. Radiology. 2018;287(3):732–47. aggressive breast cancer following ductal carcinoma in situ. Cancer. 23. Luo WQ, Huang QX, Huang XW, Hu HT, Zeng FQ, Wang W. Predicting 2019;125(18):3225–33. breast cancer in breast imaging reporting and data system (BI-RADS) 3. Villanueva H, Grimm S, Dhamne S, Rajapakshe K, Visbal A, Davis CM, ultrasound category 4 or 5 lesions: a nomogram combining radiomics Ehli EA, Hartig SM, Coarfa C, Edwards DP. The emerging roles of steroid and BI-RADS. Sci Rep. 2019;9(1):11921. hormone receptors in ductal carcinoma in situ (DCIS) of the breast. J 24. Lin F, Wang Z, Zhang K, Yang P, Ma H, Shi Y, Liu M, Wang Q, Cui J, Mao Mammary Gland Biol Neoplasia. 2018;23(4):237–48. N, et al. Contrast-enhanced spectral mammography-based radiomics 4. Kuerer HM, Albarracin CT, Yang WT, Cardiff RD, Brewster AM, Symmans nomogram for identifying benign and malignant breast lesions of sub-1 WF, Hylton NM, Middleton LP, Krishnamurthy S, Perkins GH, et al. Ductal cm. Front Oncol. 2020;10:573630. carcinoma in situ: state of the science and roadmap to advance the field. 25. Cardoso F, Bartlett JMS, Slaets L, van Deurzen CHM, van Leeuwen- J Clin Oncol. 2009;27(2):279–88. Stok E, Porter P, Linderholm B, Hedenfalk I, Schröder C, Martens J, et al. 5. Schnitt SJ. Diagnosis of ductal carcinoma in situ in an era of de-escalation Characterization of male breast cancer: results of the EORTC 10085/ of therapy. Modern. 2020;34:1–7. TBCRC/BIG/NABCG International Male Breast Cancer Program. Ann Oncol. 6. Martínez-Pérez C, Turnbull AK, Ekatah GE, Arthur LM, Sims AH, Thomas 2018;29(2):405–17. JS, Dixon JM. Current treatment trends and the need for better predic- 26. Hu HT, Wang Z, Huang XW, Chen SL, Zheng X, Ruan SM, Xie XY, Lu MD, Yu tive tools in the management of ductal carcinoma in situ of the breast. J, Tian J, et al. Ultrasound-based radiomics score: a potential biomarker for Cancer Treat Rev. 2017;55:163–72. the prediction of microvascular invasion in hepatocellular carcinoma. Eur 7. Shah C, Wobb J, Manyam B, Kundu N, Arthur D, Wazer D, Fernandez E, Radiol. 2019;29(6):2890–901. Vicini F. Management of ductal carcinoma in situ of the breast: a review. 27. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan JAMA Oncol. 2016;2(8):1083–8. V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts H. Computational 8. Giuliano M, Schettini F, Rognoni C, Milani M, Jerusalem G, Bachelot T, De radiomics system to decode the radiographic phenotype. Cancer Res. Laurentiis M, Thomas G, De Placido P, Arpino G, et al. Endocrine treat- 2017;77(21):e104–7. ment versus chemotherapy in postmenopausal women with hormone 28. Peng Y, Lin P, Wu L, Wan D, Zhao Y, Liang L, Ma X, Qin H, Liu Y, Li X, et al. receptor-positive, HER2-negative, metastatic breast cancer: a systematic Ultrasound-based radiomics analysis for preoperatively predicting dif- review and network meta-analysis. Lancet Oncol. 2019;20(10):1360–9. ferent histopathological subtypes of primary liver cancer. Front Oncol. 9. Hida AI, Omanovic D, Pedersen L, Oshiro Y, Ogura T, Nomura T, Kure- 2020;10:1646. bayashi J, Kanomata N, Moriya T. Automated assessment of Ki-67 in breast 29. Lucia F, Visvikis D, Vallières M, Desseroit MC, Miranda O, Robin P, Bonaffini cancer: the utility of digital image analysis using virtual triple staining and PA, Alfieri J, Masson I, Mervoyer A, et al. External validation of a combined whole slide imaging. Histopathology. 2020;77(3):471–80. PET and MRI radiomics model for prediction of recurrence in cervical 10. Guarneri V, Dieci MV, Bisagni G, Frassoldati A, Bianchi GV, De Salvo GL, cancer patients treated with chemoradiotherapy. Eur J Nucl Med Mol Orvieto E, Urso L, Pascual T, Paré L, et al. De-escalated therapy for HR+/ Imaging. 2019;46(4):864–77. HER2+ breast cancer patients with Ki67 response after 2-week letrozole: 30. Atabaki-Pasdar N, Ohlsson M, Viñuela A, Frau F, Pomares-Millan H, results of the PerELISA neoadjuvant study. Ann Oncol. 2019;30(6):921–6. Haid M, Jones AG, Thomas EL, Koivula RW, Kurbasic A, et al. Predicting 11. Kim M, Katayose Y, Rojanala L, Shah S, Sgagias M, Jang L, Jung YJ, Lee SH, and elucidating the etiology of fatty liver disease: a machine learning Hwang SG, Cowan KH. Induction of apoptosis in p16INK4A mutant cell modeling and validation study in the IMI DIRECT cohorts. PLoS Med. lines by adenovirus-mediated overexpression of p16INK4A protein. Cell 2020;17(6):e1003149. Death Differ. 2000;7(8):706–11. 31. Kapwata T, Gebreslasie MT. Random forest variable selection in spatial 12. Shan M, Zhang X, Liu X, Qin Y, Liu T, Liu Y, Wang J, Zhong Z, Zhang Y, Geng malaria transmission modelling in Mpumalanga Province, South Africa. J, et al. P16 and p53 play distinct roles in different subtypes of breast Geospat Health. 2016;11(3):434. cancer. PLoS ONE. 2013;8(10):e76408. 32. Tian XP, Su N, Wang L, Huang WJ, Liu YH, Zhang X, Huang HQ, Lin TY, 13. Juan MW, Yu J, Peng GX, Jun LJ, Feng SP, Fang LP. Correlation between Ma SY, Rao HL, et al. A CpG Methylation Classifier to Predict Relapse DCE-MRI radiomics features and Ki-67 expression in invasive breast in Adults with T-Cell Lymphoblastic Lymphoma. Clin Cancer Res. cancer. Oncol Lett. 2018;16(4):5084–90. 2020;26(14):3760–70. 14. Holmberg L, Wong YN, Tabár L, Ringberg A, Karlsson P, Arnesson LG, Sand- 33. Naeem SM, Mabrouk MS, Marzouk SY, Eldosoky MA. A diagnostic elin K, Anderson H, Garmo H, Emdin S. Mammography casting-type calci- genomic signal processing (GSP)-based system for automatic feature fication and risk of local recurrence in DCIS: analyses from a randomised analysis and detection of COVID-19. Brief Bioinform. 2020;22:1197–206. study. Br J Cancer. 2013;108(4):812–9. 34. Dey N, Rajinikanth V, Fong SJ, Kaiser MS, Mahmud M. Social group 15. Won SY, Park HS, Kim EK, Kim SI, Moon HJ, Yoon JH, Park VY, Park S, Kim MJ, optimization-assisted Kapur’s entropy and morphological segmentation Cho YU, et al. Survival rates of breast cancer patients aged 40 to 49 years for automated detection of COVID-19 infection from computed tomogra- according to detection modality in Korea: screening ultrasound versus phy images. Cognit Comput. 2020;54:1–13. mammography. Korean J Radiol. 2020;22:159–67. 35. Rosner D, Bedwani RN, Vana J, Baker HW, Murphy GP. Noninvasive breast 16. Watanabe T, Yamaguchi T, Tsunoda H, Kaoku S, Tohno E, Yasuda H, Ban K, carcinoma: results of a national survey by the American College of Sur- Hirokaga K, Tanaka K, Umemoto T, et al. Ultrasound image classification of geons. Ann Surg. 1980;192(2):139–47. ductal carcinoma in situ (DCIS) of the breast: analysis of 705 DCIS lesions. 36. Groen EJ, Elshof LE, Visser LL, Rutgers EJT, Winter-Warnars HAO, Lips EH, Ultrasound Med Biol. 2017;43(5):918–25. Wesseling J. Finding the balance between over- and under-treatment of ductal carcinoma in situ (DCIS). Breast. 2017;31:274–83. Wu et al. BMC Med Imaging (2021) 21:84 Page 14 of 14 37. Davey C, White V, Warne C, Kitchen P, Villanueva E, Erbas B. Understanding with breast ultrasonography versus mammography for breast cancer a ductal carcinoma in situ diagnosis: patient views and surgeon descrip- screening in women at average risk. Cochrane Database Syst Rev. tions. Eur J Cancer Care. 2011;20(6):776–84. 2013;2013(4):Cd009632. 38. Hunter NB, Kilgore MR, Davidson NE. The long and winding road for 47. Pu H, Peng J, Xu F, Liu N, Wang F, Huang X, Jia Y. Ultrasound and clinical breast cancer biomarkers to reach clinical utility. Clin Cancer Res. characteristics of false-negative results in mammography screening of 2020;26:5543–5. dense breasts. Clin Breast Cancer. 2020;20(4):317–25. 39. van Seijen M, Lips EH, Thompson AM, Nik-Zainal S, Futreal A, Hwang ES, 48. Gunawardena DS, Burrows S, Taylor DB. Non-mass versus mass-like ultra- Verschuur E, Lane J, Jonkers J, Rea DW, et al. Ductal carcinoma in situ: to sound patterns in ductal carcinoma in situ: is there an association with treat or not to treat, that is the question. Br J Cancer. 2019;121(4):285–92. high-risk histology? Clin Radiol. 2020;75(2):140–7. 40. Visser LL, Elshof LE, Van de Vijver K, Groen EJ, Almekinders MM, Sanders 49. Partridge SC, Nissan N, Rahbar H, Kitsch AE, Sigmund EE. Diffusion- J, Bierman C, Peters D, Hofland I, Broeks A, et al. Discordant marker expres- weighted breast MRI: clinical applications and emerging techniques. J sion between invasive breast carcinoma and corresponding synchronous Magn Reson Imaging. 2017;45(2):337–55. and preceding DCIS. Am J Surg Pathol. 2019;43(11):1574–82. 50. Alderson PO, Summers RM. The evolving status of radiomics. J Natl Can- 41. Zhang GJ, Kimijima I, Abe R, Kanno M, Katagata N, Hara K, Watanabe T, cer Inst. 2020;112:869–70. Tsuchiya A. Correlation between the expression of apoptosis-related 51. Lundervold AS, Lundervold A. An overview of deep learning in medical bcl-2 and p53 oncoproteins and the carcinogenesis and progression of imaging focusing on MRI. Z Med Phys. 2019;29(2):102–27. breast carcinomas. Clin Cancer Res. 1997;3(12 Pt 1):2329–35. 52. Demircioglu A, Grueneisen J, Ingenwerth M, Hoffmann O, Pinker- 42. Davis JE, Nemesure B, Mehmood S, Nayi V, Burke S, Brzostek SR, Singh M. Domenig K, Morris E, Haubold J, Forsting M, Nensa F, Umutlu L. A rapid Her2 and Ki67 biomarkers predict recurrence of ductal carinoma in situ. volume of interest-based approach of radiomics analysis of breast MRI Appl Immunohistochem Mol Morphol. 2016;24(1):20–5. for tumor decoding and phenotyping of breast cancer. PLoS ONE. 43. Cornfield DB, Palazzo JP, Schwartz GF, Goonewardene SA, Kovatich AJ, 2020;15(6):e0234871. Chervoneva I, Hyslop T, Schwarting R. The prognostic significance of 53. Zhou J, Tan H, Li W, Liu Z, Wu Y, Bai Y, Fu F, Jia X, Feng A, Liu H, et al. multiple morphologic features and biologic markers in ductal carcinoma Radiomics signatures based on multiparametric MRI for the preoperative in situ of the breast: a study of a large cohort of patients treated with prediction of the her2 status of patients with breast cancer. Acad Radiol. surgery alone. Cancer. 2004;100(11):2317–27. 2020;23:568–98. 44. Lee CH, Dershaw DD, Kopans D, Evans P, Monsees B, Monticciolo D, 54. Prasanna P, Tiwari P, Madabhushi A. Co-occurrence of local anisotropic Brenner RJ, Bassett L, Berg W, Feig S, et al. Breast cancer screening with gradient orientations (CoLlAGe): a new radiomics descriptor. Sci Rep. imaging: recommendations from the Society of Breast Imaging and the 2016;6:37241. ACR on the use of mammography, breast MRI, breast ultrasound, and other technologies for the detection of clinically occult breast cancer. J Publisher’s Note Am Coll Radiol. 2010;7(1):18–27. Springer Nature remains neutral with regard to jurisdictional claims in pub- 45. Vourtsis A, Berg WA. Breast density implications and supplemental lished maps and institutional affiliations. screening. Eur Radiol. 2019;29(4):1762–77. 46. Gartlehner G, Thaler K, Chapman A, Kaminski-Hartenthaler A, Ber- zaczy D, Van Noord MG, Helbich TH. Mammography in combination Re Read ady y to to submit y submit your our re researc search h ? Choose BMC and benefit fr ? Choose BMC and benefit from om: : fast, convenient online submission thorough peer review by experienced researchers in your field rapid publication on acceptance support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations maximum visibility for your research: over 100M website views per year At BMC, research is always in progress. Learn more biomedcentral.com/submissions http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BMC Medical Imaging Springer Journals

Preoperative ultrasound radiomics analysis for expression of multiple molecular biomarkers in mass type of breast ductal carcinoma in situ

Download PDF
14 pages

Loading next page...
 
/lp/springer-journals/preoperative-ultrasound-radiomics-analysis-for-expression-of-multiple-qci41iBg0v

References (58)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2021
eISSN
1471-2342
DOI
10.1186/s12880-021-00610-7
Publisher site
See Article on Publisher Site

Abstract

Background: The molecular biomarkers of breast ductal carcinoma in situ (DCIS) have important guiding signifi- cance for individualized precision treatment. This study was intended to explore the significance of radiomics based on ultrasound images to predict the expression of molecular biomarkers of mass type of DCIS. Methods: 116 patients with mass type of DCIS were included in this retrospective study. The radiomics features were extracted based on ultrasound images. According to the ratio of 7:3, the data sets of molecular biomarkers were split into training set and test set. The radiomics models were developed to predict the expression of estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), Ki67, p16, and p53 by using com- bination of multiple feature selection and classifiers. The predictive performance of the models were evaluated using the area under the curve (AUC) of the receiver operating curve. Results: The investigators extracted 5234 radiomics features from ultrasound images. 12, 23, 41, 51, 31 and 23 fea- tures were important for constructing the models. The radiomics scores were significantly (P < 0.05) in each molecular marker expression of mass type of DCIS. The radiomics models showed predictive performance with AUC greater than 0.7 in the training set and test set: ER (0.94 and 0.84), PR (0.90 and 0.78), HER2 (0.94 and 0.74), Ki67 (0.95 and 0.86), p16 (0.96 and 0.78), and p53 (0.95 and 0.74), respectively. Conclusion: Ultrasonic-based radiomics analysis provided a noninvasive preoperative method for predicting the expression of molecular markers of mass type of DCIS with good accuracy. Keywords: DCIS, Molecular biomarkers, Radiomics, Ultrasound Background potential to further develop into breast invasive can- Breast ductal carcinoma in  situ (DCIS) is a kind of cer [3]. The clinical treatments of patients with DCIS malignant tumor originated in the ductal epithelial tis- include surgical resection, radiotherapy, chemotherapy sue, limited to the basement membrane [1]. DCIS is and endocrine therapy, in which surgical resection the second most common breast tumor, and accounts includes simple focal resection and mastectomy, with for approximately 20–30% [2]. Some DCIS had the different therapeutic effects [4 ]. Although the progno- sis of DCIS is good, more than 14% of DCIS patients may develop invasive cancer without treatment within *Correspondence: heyun@stu.gxmu.edu.cn; yanghong@gxmu.edu.cn 10 years [5]. In the past 10 years, the incidence of DCIS Linyong Wu and Yujia Zhao have contributed equally to this work has gradually increased, highlighting the understand- Department of Medical Ultrasound, The First Affiliated Hospital ing the importance of DCIS pathology [6]. However, of Guangxi Medical University, Nanning, Guangxi Zhuang Autonomous Region 530021, People’s Republic of China the pathologic mechanism of the transition from DCIS Full list of author information is available at the end of the article © The Author(s) 2021. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons. org/ publi cdoma in/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Wu et al. BMC Med Imaging (2021) 21:84 Page 2 of 14 to invasive carcinoma is still unclear, which produces provide dynamic and accurate evaluation of biological clinical challenges of overdiagnosis and overtreatment behavior information for the clinic. in patients with DCIS [7]. Therefore, the investiga - With the breakthrough of imaging technology, mam- tors thought more studies were need to understand mography is an important examination method for the potential of the pathological process of DCIS, in DCIS, which is sensitive to the detection of calcification order to adapt to the current individualized, refined [14]. Ultrasound (US) has become the main examina- treatment. tion technology to detecting breast lesions [15], which Immunohistochemistry (IHC) can reflect the expres - is real-time, dynamic and non-invasive. There are two sion of molecular biomarkers in tumor tissue, which can types of DCIS: mass and non-mass. Some studies sug- further clarify the biological behaviors of tumors. The gested that the detection rate of US in 93 patients with expression of different molecular biomarkers can lead to mass type DCIS reached 77.4% [16]. The main character - different biological behaviors and treatments. Some stud - istics of DCIS in ultrasound were: uneven low or slightly ies have shown that some molecular biomarkers were low echo, irregular shape, unclear borders, parallel skin, important indicators for predicting biological behavior weakened posterior echo, calcification, and some blood and judging follow-up treatment in patients with DCIS, flow signals [17, 18]. However, mammography or US- such as estrogen receptor (ER), progesterone receptor assisted screening could increase overdiagnosis because (PR), human epidermal growth factor receptor 2 (HER2), both tests primarily detect low-grade invasive cancers Ki67, p16, and p53. ER and PR are the earliest molecular [19]. At the same time, radiologists are very time-con- biomarkers of breast cancer. They are predictors of breast suming to accumulate experience and have strong per- cancer prognosis and endocrine adjuvant therapy [8]. sonal subjectivity, which is another problem that needs HER2 is a proto-oncogene, which is mainly involved in to be solved. There is an urgent need for more advanced tumor signal transduction and cell proliferation. Its posi- imaging evaluation methods to guide the diagnosis and tive expression can lead to a high distant metastasis rate treatment of DCIS. and poor prognosis of breast cancer. Ki67 is an antigenic Breast lesions are diagnosed and screened by various nuclear protein that can be used as a proliferation marker. imaging methods, such as mammography, US, and mag- Its high expression is considered to be a biomarker of netic resonance imaging (MRI). All three examination tumor invasion [9]. Ki67 has a good application prospect have some limitations [20]. Radiomics is a hot subject of in predicting endocrine therapy response of breast can- artificial intelligence that is applied in the medical imag - cer [10]. Defined as a tumor suppressor gene, p16 is con - ing field, which is the cornerstone of precision science sidered to be an important cell cycle regulator [11]. P16 is in the future. Radiomics is defined as the extraction of closely related to abnormal methylation initiation. P53 is high-throughput features from single or multiple medi- a common tumor suppressor gene. Impaired function of cal image patterns to select features that are closely asso- p53, such as p53 mutation, can lead to uncontrolled pro- ciated with tumors,and the ultimate goal is to construct liferation of damaged cells [12]. Therefore, accurate iden - prediction models based on features to provide accurate tification of the expression of molecular biomarkers can tumor phenotypic analysis information and accurate help stratify tumor risk and facilitate the development of treatment decision-making [21]. Radiomics highlights personalized and accurate treatment plans. the image features that are not visible to the naked eye, Currently, the preoperative evaluation of the molecu- thus significantly enhancing the predictive power of lar biomarkers of DCIS mainly depends on IHC detec- medical imaging [22]. Radiomics has been developed tion after biopsy. However, because the progression of in a wide range of fields, such as disease diagnosis and tumors are dynamic process, there are differences in biological behavior judgment. For example, the US- spatio-temporal evolution. In addition, the evaluation radiomics model developed by Luo WQ et al. had better results of a few tissue biopsies do not necessarily repre- performance in distinguishing breast lesions than breast sent the expression of the molecular biomarkers of the imaging reporting and data system (BI-RADS) [23]. Lin whole tumor [13]. Invasive procedures and potential risks F et al. found that the radiomics score was more effective limit its multiple applications in monitoring tumor pro- than the clinical radiological model in benign and malig- gression and biological behavior. However, the preopera- nant breast lesions (< 1  cm) [24]. These series of studies tive monitoring of molecular biomarkers can dynamically showed that radiomics had better performance than tra- identify the progression of tumors and the changes in ditional imaging features in the diagnosis of breast dis- biological behavior, which has great significance for the eases to some extent. accurate formulation of treatment plans and the evalu- u Th s, this retrospective study intended to further clar - ation of curative effects. To avoid overdiagnosis and ify the relationship between US-radiomics and molecular overtreatment of patients with DCIS, it is necessary to markers of DCIS. Radiomics models had been developed Wu  et al. BMC Med Imaging (2021) 21:84 Page 3 of 14 to noninvasively evaluate the expression of molecular radiologists with five years of working experience manu - markers to help achieve accurate risk stratification and ally delineated the region of interest (ROI) of the lesions. treatment for patients with DCIS. The radiologists disregarded the diagnosis and patho - logical results of the patients [26]. After the discussion, when there was a big difference between the two radiolo - Materials and methods gists, the third radiologist with 10 years of experience re- Study cohort examined and confirmed the final boundary. This process Clinical data of 400 patients with DCIS who were path- provided reliable DCIS area contours and ensured the ologically confirmed by surgery were retrospectively accuracy of feature extraction. analyzed by the investigators. The data were based on the pathology reports of the first affiliated hospital of Guangxi medical university from January 2015 to July Image pre‑processing and feature extraction 2020. Further inclusion and exclusion criteria for this The Intelligence Foundry software (version 1.3, GE cohort study were as follows. Inclusion criteria: (1) pri- Healthcare, Shanghai, China) was applied for radiomics mary breast DCIS, (2) IHC results of molecular biomark- analysis. Figure  2 summarized the main flow of radiom - ers; and (3) preoperative US data within one month. ics analysis. The software relied on algorithms provided Exclusion criteria: (1) non-mass DCIS, including mani- by the Pyradiomics package that comply with the image festations of ductal dilation, diffuse calcification, and dif - biomarker standardization initiative (IBSI, version fuse distribution of lesions; (2) unclear image of target 2016) [27]. Features were automatically calculated and lesions; (3) secondary DCIS or postoperative recurrence extracted by the Pyradiomics extractor. The maximum of DCIS; (4) preoperative treatment history of radio- number of features extraction of the software was 5234, therapy, chemotherapy and traditional Chinese medicine; including: 122 original, 48 intraperinodular textural and (5) lack of clinical data. transition (ipris), 468 co-occurrence of local anisotropic This study finally enrolled a total of 116 patients with gradient orientations (CoLIAGe), 432 wavelets + local DCIS. The IHC (ER, PR, HER2, Ki67, p16, and p53) binary pattern (LBP), 2,944 shearlets, 1,080Gabors, 80 conformed to the diagnositic criteria of the department phased congruency-based local binary pattern (PLBP) pathology in this hospital, and were classified as positive and 60 wavelet-based improved local binary pattern or negative. The IHC results of Ki67 were positive or high (WILBP) features  (Additional file  1). Before feature expression (Ki67 > = 14%) and negative or low expres- extraction, the images were pre-processed: the gray sion (Ki67 < 14%) [25]. The number of patients enrolled value of the image was discretized with a bin size of 256, for each molecular biomarker were listed as follows: 112 and the original features were extracted. The features of cases (ER), 109 cases (PR), 94 cases (HER2), 107 cases wavelets + LBP, Shearlets, Gabors, PLBP and WILBP (Ki67), 74 cases (p16), and 116 cases (p53) (Fig. 1). were extracted by wavelet transform, shearlet transform and garber operator transform on the gray value matrix Image collection and tumor segmentation of the original images, respectively [28] (Fig. 3). Each US radiologist involved in image collection had over 5  years of experience in the field of breast. Before collecting US data, all radiologists were strictly trained. Data grouping and data cleaning GE Logiq E9 (GE Healthcare, United States), Aloka To balance the initial distribution of data, each sub-data EZU-MT28-S1 (Aloka, Japan) and MYLAB CLASS C set was randomly split into training set and test set in a (MYLAB, Italy) medical ultrasound diagnositic instru- ratio of 7:3. Based on the difference in the image extrac - ments were utilized for image collection. The breast tion feature quantization caused by different medical probe was selected, and the frequency was set to ultrasound diagnositic instruments and parameters, the 7–14  MHz. The patients took the supine position, put combat method was employed to solve this problem. The their hands on the head, and fully exposed the breast combat method could be used to coordinate and correct area and armpits on both sides. The lesions were scanned the differences between different machines and different from multiple angles, and the largest clear image of the center images. Some studies had applied this method to lesions were selected. The following ultrasonic charac - the MRI images [29]. In addition, the median value of the teristics of the lesions were recorded: BI-RADS classi- feature quantization value was applied to fill the miss - fication, location, size, shape, boundary, internal echo, ing sample. The min–max normalization method was calcification, posterior echo changes, ductal dilatation, employed to normalize the feature data to improve the blood flow signal distribution and axillary lymph nodes. comparability between features. It converted the original These images were imported into the ITKSNAP soft - data to the range of [0, 1] by linearization, which realized ware (version 3.8.0). To avoid subjective compliance, two the proportional scaling of the original data. Wu et al. BMC Med Imaging (2021) 21:84 Page 4 of 14 Fig. 1 Study cohort. a Workflow of study cohort inclusion. b Up-set plot of the expression of molecular markers shared between different samples Feature importance importance of features for subsequent model construct- The purpose of the study is difficult to explain with ing. Multiple combination techniques were applied to thousands of radiomics features of high-dimensional explain the importance of features: First, Spearman data. Feature importance analysis helps to explain the correlation coefficient test was used to eliminate high Wu  et al. BMC Med Imaging (2021) 21:84 Page 5 of 14 Fig. 2 Workflow of radiomics analysis correlation features with threshold values (0.75, 0.85, vector machine (SVM)) were employed to predict the 0.95). This test was a statistical index to measure the expression levels of the molecular biomarkers of DCIS correlation between two variables. Three dimension - [33, 34], and the score of each model was calculated. ality reduction methods (least absolute shrinkage and In addition, the fivefold cross-validation method was selection operator (LASSO) [30], random forests (RF) explored to improve the accuracy of the models. The test [31], and support vector machine-recursive feature set was used to evaluate the reliability of the models. elimination (SVM-RFE) [32]) separated or jointed sta- To accurately evaluate the predictive ability of radi- tistical tests for selecting the important features. In the omics models, the receiver operating curve (ROC), the statistical test, if the data accorded with the normal dis- area under the curve (AUC), accuracy (ACC), precision tribution, the t-test was adopted; otherwise, the Mann– (PREC), sensitivity (Sn) and specificity (Sp) were adopted Whitney U test was adopted. for the evaluation. The closer the AUC was to 1, the higher the diagnostic efficiency was. In this study, only Predictive radiomics models the best classification results of the classifier were shown. Machine learning algorithms were developed based on Python environment. Five machine-learning-based clas- sifiers (decision tree (DT), k-nearest neighbors (KNN), logistics regression (LR), naive Bayes (NB), and support Wu et al. BMC Med Imaging (2021) 21:84 Page 6 of 14 Fig. 3 The process of quantifying features. a Delineation of the ROIs. b Gray level co-occurrence matrix (GLCM), run length matrices (RLM), and histogram feature extraction. c The classification of 5234 features with p16-positives; 34 patients with p53-negatives and 82 Results patients with p53-positives. Patient characteristics and molecular biomarkers of interest Radiomics analysis The mean age of all patients was 48.8 ± 11.1 years, and the The correlation clustering heatmaps among 5234 fea - age range was 29–84  years. The characteristics parame - tures of each molecular biomarkers (ER, PR, HER2, ters were shown in Table 1. The ultrasonographic features Ki67, p16, and p53) were shown in Fig.  4. A list of of the patients were similar to those reported in the lit- 18 feature importance methods were obtained, and eratures. The expression of ER, PR, HER2, Ki67, p16, and the combination feature selection methods for opti- p53 were as follows: 49 patients with ER-negatives and mal modeling results of each molecular biomark- 63 patients with ER-positives; 53 patients with PR-neg- ers were as follows: Spearman0.75 + Statistical atives and 59 patients with PR-positives; 36 patients with Test + RF, Spearman0.75 + Statistical Test + RF, Spear- HER2-negatives and 58 patients with HER2-positives; 45 man0.75 + L ASSO, Spearman0.75 + Statistical patients with Ki67-negatives and 62 patients with Ki67- Test + RF, Spearman0.75 + Statistical Test + SVM-RFE , positives; 29 patients with p16-negatives and 45 patients Wu  et al. BMC Med Imaging (2021) 21:84 Page 7 of 14 Table 1 Patient characteristics and molecular biomarkers of interest Parameters N = 116 Parameters N = 116 Median age (years) 48.8 ± 11.1 Shape rule (yes/no) 26/96 Immunohistochemistry Clear boundary (yes/no) 50/66 ER (−/+/NA) 49/63/4 Aspect ratio (< 1/ > = 1) 6/110 PR (−/+/NA) 53/56/7 Echo uniformity (yes/no) 19/97 HER2 (−/+/NA) 36/58/22 Calcification (yes/no) 69/47 Ki67 (−/+/NA) 45/62/9 Intrafocal blood flow (yes/no) 79/37 P16 (−/+/NA) 29/45/42 Peripheral blood flow (yes/no) 32/84 P53(−/+/NA) 34/82/0 Catheter dilatation (yes/no) 9/107 Ultrasonic characteristics lymph nodes (< 1/ > = 1) 93/23 Median size (cm) 2.6 ± 1.6 BI-RADS classification (3/4a/4b/4c/5/6) 12/31/30/23/9/11 Fig. 4 Correlation cluster analysis of 5234 radiomics features. The Pearson correlation test was used to analyze the correlation between features, and the "pheatmap" R software package was applied to draw heat maps. a ER; b PR; c HER2; d Ki67; e p16; f p53 Ninety models were obtained by constructing predic- and Spearman0.85 + SVM-RFE. 12 features, 23 fea- tion models with five classifiers, and the performance tures, 41 features, 20 features, 31 features and 23 fea- of the models were presented in Fig.  6  (Additional tures were important for constructing prediction file  2). The optimal radiomics models were constructed models. The heatmaps of the model features were pre - by DT, SVM, KNN, SVM, KNN and KNN classifiers, sented in Fig. 5. Wu et al. BMC Med Imaging (2021) 21:84 Page 8 of 14 Fig. 5 Important features for each molecular biomarkers. a ER; b PR; c HER2; d Ki67; e p16; f p53 respectively, and showed above moderate predictive Discussion performance in predicting the expression of molecular This study was the first non-invasive comprehensive markers of DCIS (Table  2). Radiomics scores of train- analysis based on US-radiomics to predict the expres- ing set and test set were significantly different in each sion of molecular markers of DCIS. The investigators molecular marker expression (training set, P < 0.001, recruited only 116 patients with DCIS for this study, but test set, P < 0.05). The predictive performance of the it was exciting to see that the radiomics models showed radiomics models of each molecular biomarker in the more than moderate predictive performance in predict- training set: ER (AUC, 0.94, 95% confidence interval ing molecular biomarker expression of DCIS. (CI) 0.89–0.99), PR (AUC, 0.90, 95% CI 0.83–0.97), DCIS is a malignant tumor with good prognosis, but HER2 (AUC, 0.94, 95% CI 0.89–0.99), Ki67 (AUC, it is heterogeneous in morphology and genetics. Before 0.95, 95% CI 0.90–0.99), p16 (AUC, 0.96, 95% CI 0.91– the imaging examination was performed, the diagnosis of 1.00), p53 (AUC, 0.95, 95% CI 0.90–0.99), respectively DCIS was only due to the appearance of nipple discharge (Fig.  7). The calibration curve of the prediction models and/or palpable mass symptoms, which accounted for in the training set confirmed the better consistency of only 2% of DCIS detected. It showed that DCIS with hid- the models (Fig.  8). The radiomics models showed pre - den symptoms were easily missed [35]. With the screen- dictive performance with AUC greater than 0.7 in the ing of imaging technology (mammography, US and MRI), test set: ER (AUC, 0.84, 95% CI 0.68–0.99), PR (AUC, the detection rate of DCIS had gradually increased. This 0.78, 95% CI, 0.60–0.96), HER2 (AUC, 0.74, 95% CI detection rate included symptomatic DCIS, and whether 0.74–0.99), Ki67 (AUC, 0.86, 95% CI 0.67–0.97), p16 there was overdiagnosis in the detection of insidious (AUC, 0.78, 95% CI 0.59–0.97), p53 (AUC, 0.74, 95% CI DCIS was also a hot topic of controversy [36], Unfortu- 0.55–0.93), respectively (Fig. 9). nately, the diagnosis of DCIS marked women as at risk of invasive breast cancer, so women diagnosed with DCIS may suffer serious psychological distress, leading to the progression of DCIS [37]. In addition, the current Wu  et al. BMC Med Imaging (2021) 21:84 Page 9 of 14 Fig. 6 Heat maps of evaluation indicators for ninety radiomics prediction models. a ER; b PR; c HER2; d Ki67; e p16; f p53 Table 2 Evaluation of radiomics models in each DCIS molecular biomarkers Training set Test set AUC ACC PREC Sn Sp AUC ACC PREC Sn Sp ER 0.94 0.90 0.93 0.89 0.91 0.84 0.82 0.81 0.90 0.73 PR 0.90 0.84 0.89 0.80 0.89 0.78 0.76 0.80 0.71 0.8 HER2 0.94 0.88 0.90 0.90 0.84 0.74 0.72 0.78 0.78 0.64 Ki67 0.95 0.88 0.84 0.98 0.74 0.86 0.76 0.79 0.79 0.71 p16 0.96 0.90 0.90 0.94 0.85 0.78 0.70 0.77 0.71 0.67 p53 0.95 0.89 0.91 0.93 0.79 0.74 0.74 0.83 0.80 0.60 treatment methods were also facing the controversy with the progression of DCIS to invasive cancer [40]. For over the treatment of some patients [38]. Therefore, the example, Zhang GJ et  al. [41] found that 79% of DCIS main clinical challenge in DCIS has been to distinguish patients were positive for P53 when studying the occur- between patients who have a better chance of develop- rence and development of breast cancer. Davis et al. [42] ing invasive cancer and require more treatment and demonstrated that high Ki67 expression was an inde- those who are less likely to develop DCIS and need less pendent predictor of postoperative recurrence in patients or no treatment [39]. Immunohistochemical markers can with DCIS. Cornfield DB had found a higher recurrence explain the changes in the biological behaviors of tumors rate with PR > 3.5% using tree structure survival [43]. The on the molecular level. More and more studies have results showed that the changes in the biological behavior pointed out the changes in molecular markers associated Wu et al. BMC Med Imaging (2021) 21:84 Page 10 of 14 Fig. 7 Performance of the radiomics models in the training set. a ER; b PR; c HER2; d Ki67; e p16; f p53 Fig. 8 Calibration curves of the radiomics models in the training set. The oblique dashed line represents the perfect prediction of the ideal model. The solid line represents the performance of the radiomics model, and the dotted line near the diagonal indicates a better prediction. a ER; b PR; c HER2; d Ki67; e p16; f p53 Wu  et al. BMC Med Imaging (2021) 21:84 Page 11 of 14 Fig. 9 Performance of the radiomics models in the test set. a ER; b PR; c HER2; d Ki67; e p16; f p53 of DCIS were closely related to the expression of molecu- and quantitative analysis of the imaging features of the lar biomarkers. whole imaging mass. Therefore, radiomics is a precision About various imaging technologies, they also have medical method for non-invasive diagnosis, evaluation application limitations [44]. Mammography is the main of efficacy, biological behavior [50]. Currently, radiom - method of early breast cancer detection, but it is closely ics mainly relies machine learning algorithms to identify related to the density of the lesion and the possibility of meaningful features of image training data set, and for covering the lesion [45]. However, Chinese women have further interpretation of the information and the opti- dense breasts, so they had certain limitations in finding mization, so as to accurately predict the content of the suspicious lesions in the dense tissues of breasts through research. An independent data set is applied to test the mammography [46, 47]. The traditional mammography universality of the model and provides feedback for fur- diagnosis method will cause trauma to the patient to a ther optimization of the model [51]. To a certain extent, certain extent and reduce the patient’s treatment com- it improves the utilization of image information and ena- pliance. Due to its high sensitivity to soft tissues, US can bles differential diagnosis of diseases on more subtle lev - better show lesions in dense glands and has become the els that cannot be recognized by the naked eye. primary imaging method for Chinese women to screen Breast radiomics studies are mostly applied to the and diagnose breast diseases. For non-mass DCIS, US is prediction of the molecular classification, lymph node difficult to recognize [48]. Therefore, this retrospective metastasis and molecular markers of invasive ductal study only examined mass DCIS, which is a limitation of carcinoma. For example, Demircioglu A et al. [52] con- the study. MRI has considerable advantages in detecting structed radiomics models for predicting Ki67 expres- breast lesions, but its specificity is limited by several fac - sion in invasive breast cancer based on eight features tors that affect image quality, such as magnetic field and extracted from MRI images, with an AUC of 0.81. gradient strength, coil performance, contrast agent effi - Zhou et al. [53] explored the significance of MRI-radi- cacy and menstrual cycle [49]. omics models for predicting the expression of HER2 Radiomics mainly studies the quantitative features that in patients with invasive breast cancer before surgery; are related to biology in medical images. Radiomics fea- the validation set AUC reached 0.81. There are few tures are considered the invisible tissue infrastructure reports on DICS with radiomics. However, there are components of the object to be imaged, which can serve clinical challenges in the diagnosis and treatment of as a valuable method for studying cancer by imaging, patients with DCIS. Tumor progression and treatment such as MRI. Radiomics can provide in vivo visualization decisions are affected by multiple tumor molecular Wu et al. BMC Med Imaging (2021) 21:84 Page 12 of 14 biomarkers, which require to comprehensively ana- Conclusion lyze and evaluate the molecular biomarkers of DCIS. The application of machine learning-based radiomics To expand the application of radiomics in DCIS, the analysis provided a non-invasive method for predict- investigators carried out this study to assess the fea- ing the expression of multiple molecular biomarkers in sibility of molecular biomarkers of DCIS. The investi- DCIS, with good prediction performance. This study gators believe that information obtained from multiple also demonstrated the potential of radiomics in patho- molecular biomarkers can help explain the underlying logic assessment and individualized precision therapy. pathological process of DCIS. In this study, the first highlight, the first comprehen- Abbreviations sive analysis of molecular markers of DCIS was con- DCIS: Ductal carcinoma in situ; ER: Estrogen receptor; PR: Progesterone recep- ducted based on radiomics. Second highlight, there tor; HER2: Human epidermal growth factor receptor-2; AUC : Area under the curve; CI: Confidence interval; IHC: Immunohistochemistry; US: Ultrasound; were thousands of radiomics features, including eight MRI: Magnetic resonance imaging; BI-RADS: Breast imaging reporting and data classifications: original feature can reflect the number system; ROI: Region of interest; CoLIAGe: Co-occurrence of local anisotropic of voxels in the images, intensity distribution, pixel gradient orientations; LBP: Local binary pattern; PLBP: Phased congruency- based local binary pattern; WILBP: Wavelet-based improved local binary pair frequency, image average gray value, size and pattern; GLCM: Gray level co-occurrence matrix; RLM: Run length matrices; shape of ROI (https:// pyrad iomics. readt hedocs. io/ LASSO: Least absolute shrinkage and selection operator; RF: Random forests; en/ latest/ featu res. html); Ipris features capture nodu- SVM-RFE: Support vector machine-recursive feature elimination; DT: Decision tree; KNN: K-nearest neighbors; LR: Logistics regression; NB: Naive Bayes; SVM: lar heterogeneity and differential growth patterns; Support vector machine; ROC: Receiver operating curve; ACC : Accuracy; PREC: CoLIAGe features can distinguish disease phenotypes Precision; Sn: Sensitivity; Sp: Specificity. that have similar morphologic appearances [54]; wave- lets features represent most of the edge information Supplementary Information in images; Shearlet features are better for processing The online version contains supplementary material available at https:// doi. high-dimensional signals; Gabors features extract the org/ 10. 1186/ s12880- 021- 00610-7. edge and gradient information of image and reflect the spatial frequency feature; PLBP, and WILBP: PLBP Additional file 1. 5234 radiomics features matrix file. features are an oriented local texture descriptor that Additional file 2. Modeling matrix file for molecular biomarkers. combines the phase congruency approach with the LBP. The third highlight of this study was to construct Acknowledgements dozens of prediction models by combining multiple Not applicable. classifiers with multiple feature selection to select the Authors’ contributions optimal prediction results. RF and SVM-RFE had sig- Conceptualization: HQ and YL; Methodology: PL, DW and XL; Formal analysis nificant performance in feature selection of multiple and investigation: LW, YZ, PL, HQ, YL, DW, XL, YH, HY; Writing—original draft preparation: LW and YZ; Writing—review and editing: LW and YZ; Resources: molecular markers, KNN and SVM classification per- HY and YH; Supervision: HY and YH. All authors read and approved the final formed well too. Finally, through the verification of the manuscript. test set, the prediction models all showed moderate Funding performance. Not applicable. There were also some shortcomings in our study. First, this retrospective study had the problem of small Availability of data and materials The datasets supporting the conclusions of this article were included within sample size. It was necessary to increase the sample the article and its additional files. size or multi-center cooperation to construct univer- sal models. Second, when the radiologists manually Declarations delineated the ROIs, there were a certain degree of subjectivity to the contours of the lesions, which may Ethics approval and consent to participate This retrospective breast DCIS study was approved by the ethics committee lead to poor robustness of the models. In addition, the of the First Affiliated Hospital of Guangxi Medical University. Informed consent delineation process was done by only one radiologist. was waived. This study on the implementation of all procedures are in line Third, the investigators only investigated the features with the National Research Council of moral standards. extracted from the largest section, which could not Consent for publication represent the whole tumor. Due to the limitations of Not applicable. US, it was not possible to conduct three-dimensional Competing interests studies similar to other imaging studies. The authors declare that they have no competing interests. Wu  et al. BMC Med Imaging (2021) 21:84 Page 13 of 14 Author details 17. Li W, Zhou Q, Xia S, Wu Y, Fei X, Wang Y, Tao L, Fan J, Zhou W. Application Department of Medical Ultrasound, The First Affiliated Hospital of Guangxi of contrast-enhanced ultrasound in the diagnosis of ductal carcinoma Medical University, Nanning, Guangxi Zhuang Autonomous Region 530021, in situ: analysis of 127 cases. J Ultrasound Med. 2020;39(1):39–50. People’s Republic of China. GE Healthcare, Shanghai, People’s Republic 18. Moon HJ, Kim EK, Kim MJ, Yoon JH, Park VY. Comparison of clinical and of China. pathologic characteristics of ductal carcinoma in situ detected on mam- mography versus ultrasound only in asymptomatic patients. Ultrasound Received: 4 December 2020 Accepted: 21 April 2021 Med Biol. 2019;45(1):68–77. 19. Evans A, Vinnicombe S. Overdiagnosis in breast imaging. Breast. 2017;31:270–3. 20. Conti A, Duggento A, Indovina I, Guerrisi M, Toschi N. Radiomics in breast cancer classification and prediction. Semin Cancer Biol. 2020;1:323–4. 21. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, References they are data. Radiology. 2016;278(2):563–77. 1. Wellings S, Jensen HJJI. On the origin and progression of ductal carci- 22. Pinker K, Chin J, Melsaether AN, Morris EA, Moy L. Precision medicine and noma in the human breast. Origin Breast Carcinoma. 1973;50(5):1111–8. radiogenomics in breast cancer: new approaches toward diagnosis and 2. Liu Y, West R, Weber JD, Colditz GA. Race and risk of subsequent treatment. Radiology. 2018;287(3):732–47. aggressive breast cancer following ductal carcinoma in situ. Cancer. 23. Luo WQ, Huang QX, Huang XW, Hu HT, Zeng FQ, Wang W. Predicting 2019;125(18):3225–33. breast cancer in breast imaging reporting and data system (BI-RADS) 3. Villanueva H, Grimm S, Dhamne S, Rajapakshe K, Visbal A, Davis CM, ultrasound category 4 or 5 lesions: a nomogram combining radiomics Ehli EA, Hartig SM, Coarfa C, Edwards DP. The emerging roles of steroid and BI-RADS. Sci Rep. 2019;9(1):11921. hormone receptors in ductal carcinoma in situ (DCIS) of the breast. J 24. Lin F, Wang Z, Zhang K, Yang P, Ma H, Shi Y, Liu M, Wang Q, Cui J, Mao Mammary Gland Biol Neoplasia. 2018;23(4):237–48. N, et al. Contrast-enhanced spectral mammography-based radiomics 4. Kuerer HM, Albarracin CT, Yang WT, Cardiff RD, Brewster AM, Symmans nomogram for identifying benign and malignant breast lesions of sub-1 WF, Hylton NM, Middleton LP, Krishnamurthy S, Perkins GH, et al. Ductal cm. Front Oncol. 2020;10:573630. carcinoma in situ: state of the science and roadmap to advance the field. 25. Cardoso F, Bartlett JMS, Slaets L, van Deurzen CHM, van Leeuwen- J Clin Oncol. 2009;27(2):279–88. Stok E, Porter P, Linderholm B, Hedenfalk I, Schröder C, Martens J, et al. 5. Schnitt SJ. Diagnosis of ductal carcinoma in situ in an era of de-escalation Characterization of male breast cancer: results of the EORTC 10085/ of therapy. Modern. 2020;34:1–7. TBCRC/BIG/NABCG International Male Breast Cancer Program. Ann Oncol. 6. Martínez-Pérez C, Turnbull AK, Ekatah GE, Arthur LM, Sims AH, Thomas 2018;29(2):405–17. JS, Dixon JM. Current treatment trends and the need for better predic- 26. Hu HT, Wang Z, Huang XW, Chen SL, Zheng X, Ruan SM, Xie XY, Lu MD, Yu tive tools in the management of ductal carcinoma in situ of the breast. J, Tian J, et al. Ultrasound-based radiomics score: a potential biomarker for Cancer Treat Rev. 2017;55:163–72. the prediction of microvascular invasion in hepatocellular carcinoma. Eur 7. Shah C, Wobb J, Manyam B, Kundu N, Arthur D, Wazer D, Fernandez E, Radiol. 2019;29(6):2890–901. Vicini F. Management of ductal carcinoma in situ of the breast: a review. 27. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan JAMA Oncol. 2016;2(8):1083–8. V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts H. Computational 8. Giuliano M, Schettini F, Rognoni C, Milani M, Jerusalem G, Bachelot T, De radiomics system to decode the radiographic phenotype. Cancer Res. Laurentiis M, Thomas G, De Placido P, Arpino G, et al. Endocrine treat- 2017;77(21):e104–7. ment versus chemotherapy in postmenopausal women with hormone 28. Peng Y, Lin P, Wu L, Wan D, Zhao Y, Liang L, Ma X, Qin H, Liu Y, Li X, et al. receptor-positive, HER2-negative, metastatic breast cancer: a systematic Ultrasound-based radiomics analysis for preoperatively predicting dif- review and network meta-analysis. Lancet Oncol. 2019;20(10):1360–9. ferent histopathological subtypes of primary liver cancer. Front Oncol. 9. Hida AI, Omanovic D, Pedersen L, Oshiro Y, Ogura T, Nomura T, Kure- 2020;10:1646. bayashi J, Kanomata N, Moriya T. Automated assessment of Ki-67 in breast 29. Lucia F, Visvikis D, Vallières M, Desseroit MC, Miranda O, Robin P, Bonaffini cancer: the utility of digital image analysis using virtual triple staining and PA, Alfieri J, Masson I, Mervoyer A, et al. External validation of a combined whole slide imaging. Histopathology. 2020;77(3):471–80. PET and MRI radiomics model for prediction of recurrence in cervical 10. Guarneri V, Dieci MV, Bisagni G, Frassoldati A, Bianchi GV, De Salvo GL, cancer patients treated with chemoradiotherapy. Eur J Nucl Med Mol Orvieto E, Urso L, Pascual T, Paré L, et al. De-escalated therapy for HR+/ Imaging. 2019;46(4):864–77. HER2+ breast cancer patients with Ki67 response after 2-week letrozole: 30. Atabaki-Pasdar N, Ohlsson M, Viñuela A, Frau F, Pomares-Millan H, results of the PerELISA neoadjuvant study. Ann Oncol. 2019;30(6):921–6. Haid M, Jones AG, Thomas EL, Koivula RW, Kurbasic A, et al. Predicting 11. Kim M, Katayose Y, Rojanala L, Shah S, Sgagias M, Jang L, Jung YJ, Lee SH, and elucidating the etiology of fatty liver disease: a machine learning Hwang SG, Cowan KH. Induction of apoptosis in p16INK4A mutant cell modeling and validation study in the IMI DIRECT cohorts. PLoS Med. lines by adenovirus-mediated overexpression of p16INK4A protein. Cell 2020;17(6):e1003149. Death Differ. 2000;7(8):706–11. 31. Kapwata T, Gebreslasie MT. Random forest variable selection in spatial 12. Shan M, Zhang X, Liu X, Qin Y, Liu T, Liu Y, Wang J, Zhong Z, Zhang Y, Geng malaria transmission modelling in Mpumalanga Province, South Africa. J, et al. P16 and p53 play distinct roles in different subtypes of breast Geospat Health. 2016;11(3):434. cancer. PLoS ONE. 2013;8(10):e76408. 32. Tian XP, Su N, Wang L, Huang WJ, Liu YH, Zhang X, Huang HQ, Lin TY, 13. Juan MW, Yu J, Peng GX, Jun LJ, Feng SP, Fang LP. Correlation between Ma SY, Rao HL, et al. A CpG Methylation Classifier to Predict Relapse DCE-MRI radiomics features and Ki-67 expression in invasive breast in Adults with T-Cell Lymphoblastic Lymphoma. Clin Cancer Res. cancer. Oncol Lett. 2018;16(4):5084–90. 2020;26(14):3760–70. 14. Holmberg L, Wong YN, Tabár L, Ringberg A, Karlsson P, Arnesson LG, Sand- 33. Naeem SM, Mabrouk MS, Marzouk SY, Eldosoky MA. A diagnostic elin K, Anderson H, Garmo H, Emdin S. Mammography casting-type calci- genomic signal processing (GSP)-based system for automatic feature fication and risk of local recurrence in DCIS: analyses from a randomised analysis and detection of COVID-19. Brief Bioinform. 2020;22:1197–206. study. Br J Cancer. 2013;108(4):812–9. 34. Dey N, Rajinikanth V, Fong SJ, Kaiser MS, Mahmud M. Social group 15. Won SY, Park HS, Kim EK, Kim SI, Moon HJ, Yoon JH, Park VY, Park S, Kim MJ, optimization-assisted Kapur’s entropy and morphological segmentation Cho YU, et al. Survival rates of breast cancer patients aged 40 to 49 years for automated detection of COVID-19 infection from computed tomogra- according to detection modality in Korea: screening ultrasound versus phy images. Cognit Comput. 2020;54:1–13. mammography. Korean J Radiol. 2020;22:159–67. 35. Rosner D, Bedwani RN, Vana J, Baker HW, Murphy GP. Noninvasive breast 16. Watanabe T, Yamaguchi T, Tsunoda H, Kaoku S, Tohno E, Yasuda H, Ban K, carcinoma: results of a national survey by the American College of Sur- Hirokaga K, Tanaka K, Umemoto T, et al. Ultrasound image classification of geons. Ann Surg. 1980;192(2):139–47. ductal carcinoma in situ (DCIS) of the breast: analysis of 705 DCIS lesions. 36. Groen EJ, Elshof LE, Visser LL, Rutgers EJT, Winter-Warnars HAO, Lips EH, Ultrasound Med Biol. 2017;43(5):918–25. Wesseling J. Finding the balance between over- and under-treatment of ductal carcinoma in situ (DCIS). Breast. 2017;31:274–83. Wu et al. BMC Med Imaging (2021) 21:84 Page 14 of 14 37. Davey C, White V, Warne C, Kitchen P, Villanueva E, Erbas B. Understanding with breast ultrasonography versus mammography for breast cancer a ductal carcinoma in situ diagnosis: patient views and surgeon descrip- screening in women at average risk. Cochrane Database Syst Rev. tions. Eur J Cancer Care. 2011;20(6):776–84. 2013;2013(4):Cd009632. 38. Hunter NB, Kilgore MR, Davidson NE. The long and winding road for 47. Pu H, Peng J, Xu F, Liu N, Wang F, Huang X, Jia Y. Ultrasound and clinical breast cancer biomarkers to reach clinical utility. Clin Cancer Res. characteristics of false-negative results in mammography screening of 2020;26:5543–5. dense breasts. Clin Breast Cancer. 2020;20(4):317–25. 39. van Seijen M, Lips EH, Thompson AM, Nik-Zainal S, Futreal A, Hwang ES, 48. Gunawardena DS, Burrows S, Taylor DB. Non-mass versus mass-like ultra- Verschuur E, Lane J, Jonkers J, Rea DW, et al. Ductal carcinoma in situ: to sound patterns in ductal carcinoma in situ: is there an association with treat or not to treat, that is the question. Br J Cancer. 2019;121(4):285–92. high-risk histology? Clin Radiol. 2020;75(2):140–7. 40. Visser LL, Elshof LE, Van de Vijver K, Groen EJ, Almekinders MM, Sanders 49. Partridge SC, Nissan N, Rahbar H, Kitsch AE, Sigmund EE. Diffusion- J, Bierman C, Peters D, Hofland I, Broeks A, et al. Discordant marker expres- weighted breast MRI: clinical applications and emerging techniques. J sion between invasive breast carcinoma and corresponding synchronous Magn Reson Imaging. 2017;45(2):337–55. and preceding DCIS. Am J Surg Pathol. 2019;43(11):1574–82. 50. Alderson PO, Summers RM. The evolving status of radiomics. J Natl Can- 41. Zhang GJ, Kimijima I, Abe R, Kanno M, Katagata N, Hara K, Watanabe T, cer Inst. 2020;112:869–70. Tsuchiya A. Correlation between the expression of apoptosis-related 51. Lundervold AS, Lundervold A. An overview of deep learning in medical bcl-2 and p53 oncoproteins and the carcinogenesis and progression of imaging focusing on MRI. Z Med Phys. 2019;29(2):102–27. breast carcinomas. Clin Cancer Res. 1997;3(12 Pt 1):2329–35. 52. Demircioglu A, Grueneisen J, Ingenwerth M, Hoffmann O, Pinker- 42. Davis JE, Nemesure B, Mehmood S, Nayi V, Burke S, Brzostek SR, Singh M. Domenig K, Morris E, Haubold J, Forsting M, Nensa F, Umutlu L. A rapid Her2 and Ki67 biomarkers predict recurrence of ductal carinoma in situ. volume of interest-based approach of radiomics analysis of breast MRI Appl Immunohistochem Mol Morphol. 2016;24(1):20–5. for tumor decoding and phenotyping of breast cancer. PLoS ONE. 43. Cornfield DB, Palazzo JP, Schwartz GF, Goonewardene SA, Kovatich AJ, 2020;15(6):e0234871. Chervoneva I, Hyslop T, Schwarting R. The prognostic significance of 53. Zhou J, Tan H, Li W, Liu Z, Wu Y, Bai Y, Fu F, Jia X, Feng A, Liu H, et al. multiple morphologic features and biologic markers in ductal carcinoma Radiomics signatures based on multiparametric MRI for the preoperative in situ of the breast: a study of a large cohort of patients treated with prediction of the her2 status of patients with breast cancer. Acad Radiol. surgery alone. Cancer. 2004;100(11):2317–27. 2020;23:568–98. 44. Lee CH, Dershaw DD, Kopans D, Evans P, Monsees B, Monticciolo D, 54. Prasanna P, Tiwari P, Madabhushi A. Co-occurrence of local anisotropic Brenner RJ, Bassett L, Berg W, Feig S, et al. Breast cancer screening with gradient orientations (CoLlAGe): a new radiomics descriptor. Sci Rep. imaging: recommendations from the Society of Breast Imaging and the 2016;6:37241. ACR on the use of mammography, breast MRI, breast ultrasound, and other technologies for the detection of clinically occult breast cancer. J Publisher’s Note Am Coll Radiol. 2010;7(1):18–27. Springer Nature remains neutral with regard to jurisdictional claims in pub- 45. Vourtsis A, Berg WA. Breast density implications and supplemental lished maps and institutional affiliations. screening. Eur Radiol. 2019;29(4):1762–77. 46. Gartlehner G, Thaler K, Chapman A, Kaminski-Hartenthaler A, Ber- zaczy D, Van Noord MG, Helbich TH. Mammography in combination Re Read ady y to to submit y submit your our re researc search h ? Choose BMC and benefit fr ? Choose BMC and benefit from om: : fast, convenient online submission thorough peer review by experienced researchers in your field rapid publication on acceptance support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations maximum visibility for your research: over 100M website views per year At BMC, research is always in progress. Learn more biomedcentral.com/submissions

Journal

BMC Medical Imaging – Springer Journals

Published: May 17, 2021

There are no references for this article.