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

Learn More →

An interpretable radiomics model to select patients for radiotherapy after surgery for WHO grade 2 meningiomas

An interpretable radiomics model to select patients for radiotherapy after surgery for WHO grade... Objectives: This study investigated whether radiomic features can improve the prediction accuracy for tumor recurrence over clinicopathological features and if these features can be used to identify high‑risk patients requiring adjuvant radiotherapy (ART ) in WHO grade 2 meningiomas. Methods: Preoperative magnetic resonance imaging (MRI) of 155 grade 2 meningioma patients with a median follow‑up of 63.8 months were included and allocated to training (n = 92) and test sets (n = 63). After radiomic feature extraction (n = 200), least absolute shrinkage and selection operator feature selection with logistic regression classifier was performed to develop two models: (1) a clinicopathological model and (2) a combined clinicopathological and radiomic model. The probability of recurrence using the combined model was analyzed to identify candidates for ART. Results: The combined clinicopathological and radiomics model exhibited superior performance for the prediction of recurrence compared with the clinicopathological model in the training set (area under the curve [AUC] 0.78 vs. 0.67, P = 0.042), which was also validated in the test set (AUC 0.77 vs. 0.61, P = 0.192). In patients with a high prob‑ ability of recurrence by the combined model, the 5‑ year progression‑free survival was significantly improved with ART (92% vs. 57%, P = 0.024), and the median time to recurrence was longer (54 vs. 17 months after surgery). Conclusions: Radiomics significantly contributes added value in predicting recurrence when integrated with the clinicopathological features in patients with grade 2 meningiomas. Furthermore, the combined model can be applied to identify high‑risk patients who require ART. Keywords: Magnetic resonance imaging, Meningioma, Radiomics, Radiotherapy, Prognosis Introduction Meningiomas are the most common primary intracranial neoplasms in adults, comprising 36.7% of all intracranial tumors [1]. Since the serial updates of the World Health Chae Jung Park and Seo Hee Choi have contributed equally to this paper Organization (WHO) grading classification, the propor - *Correspondence: yaewonpark@yuhs.ac; YHI0225@yuhs.ac tion of grade 2 meningiomas has gradually increased up to 15–20% [2]. However, despite recent WHO grading Department of Radiology, Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yongin Severance Hospital, scheme, there are limitations in predicting the prognosis Yonsei University College of Medicine, Seoul, Republic of Korea of grade 2 meningiomas [3]. Grade 2 meningiomas are Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion known to have an unpredictable heterogeneous disease Therapy Research Institute, Yonsei University College of Medicine, 50‑1 Yonsei‑ro, Seodaemun‑gu, Seoul 03722, Republic of Korea course; even after gross total resection (GTR), recurrence Full list of author information is available at the end of the article © The Author(s) 2022. 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. Park et al. Radiation Oncology (2022) 17:147 Page 2 of 12 can occur in a substantial number of patients  [4, 5], index (n = 3). Finally, a total of 155 patients were enrolled whereas some patients experience a long indolent clinical (Additional file  1: Fig. S1). The study population was ran - course without adjuvant treatment [6]. domly divided into training and test sets with a ratio of Currently, adjuvant radiotherapy (ART) is recom- 6:4 (n = 92 and n = 63, respectively). mended as a standard therapy for patients who undergo Preoperative MRI was performed using a 3.0-T MRI subtotal resection (STR) in grade 2 meningiomas  [7] scanner (Achieva, Philips Medical Systems, Amsterdam, based on evidence indicating improvement in local con- Netherlands) with an eight-channel sensitivity-encoding trol and survival rates with ART  [8–10]. However, there head coil. Detailed parameters of the MRI sequences are is a lack of consensus on whether ART reduces the risk of provided in Additional file 1: Supplementary Material S1. tumor recurrence after GTR of grade 2 meningiomas [3]. u Th s, it is crucial to establish a model for predicting a Treatments patient’s individual outcome and to identify high-risk All patients underwent either GTR (n = 132) or STR patients who could benefit from ART even after GTR (n = 23) surgery. The extent of resection was defined by of grade 2 meningiomas. Patients who are at a low risk comparing preoperative and postoperative MRI scans; for recurrence can be spared from ART and its potential GTR was defined by a lack of residual enhancing tumor risks. Furthermore, in a set of patients who have under- in the image, and STR was defined by the patient having gone STR, patient subsets with a higher risk of recur- more than 50% of the tumor removed [18]. After surgical rence may undergo more intensified radiotherapy to resection, a multidisciplinary team, consisting of neuro- improve their respective outcomes. surgeons, radiation oncologists, neuropathologists, and Radiomics is an advanced technique that extracts neuroradiologists, decided whether to perform ART, high-dimensional quantitative imaging features, such which consisted of three-dimensional conformal radio- as intensity distributions, spatial relationships, textural therapy and intensity-modulated radiotherapy. A total of heterogeneity, and shape descriptors  [11, 12]. Radiom- 97 patients underwent ART; 80 patients received it after ics aims to discover meaningful “hidden” information GTR (80/132, 60.6%), whereas 17 received it after STR within radiological images that is visually inaccessible. (17/23, 73.9%). Among the patients who underwent ART, Previous studies have demonstrated the use of radiomics intensity-modulated radiotherapy was performed on 86 in differential diagnosis and grade prediction for men - patients (88.7%, median 60.0 Gy), and three-dimensional ingiomas  [13–15]. Some studies have also shown that conformal radiotherapy was performed on 11 patients radiomics can predict tumor recurrence in patients with (11.3%, median 59.4 Gy). meningiomas  [16, 17]. We hypothesized that radiomics can enable risk stratification for tumor recurrence after Response assessment by a neuro‑oncology meningioma surgery in patients with grade 2 meningiomas, which working group may guide patients toward ART. The evaluation of the tumor response and progression Therefore, this study aimed to investigate whether radi - was determined according to the Response Assessment omics features can be used to improve the prediction of in Neuro-Oncology (RANO) criteria  [19] by compar- tumor recurrence over clinicopathological features, and ing serial MRIs of each patient. A radiation oncologist if radiomics can be used to identify high-risk patients and neuroradiologist (with 9 and 8  years of experience, who require ART in WHO grade 2 meningiomas. respectively) performed the evaluation and the results were achieved by consensus. A detailed description of Methods the RANO criteria and the definitions of each assessment Patient population criteria are presented in the Additional file  1: Supple- This study was approved by the institutional review board mentary material S2 and Additional file  1: Table  S1. The of the Yonsei University Health System (9-2021-0047). patients assessed as having a progressive disease in the The institutional review board waived the requirement follow-up periods were considered to have tumor recur- to obtain informed patient consent for this retrospec- rence. Patients assessed as having complete response, tive study. We retrospectively reviewed 199 patients partial response, minor response, and stable disease were with surgically confirmed WHO grade 2 meningiomas considered to be patients without tumor recurrence. The who underwent preoperative conventional magnetic primary endpoint was tumor recurrence assessed by the resonance imaging (MRI) between February 2005 and RANO criteria during the follow-up period. November 2018. The exclusion criteria were as follows: (1) incomplete MRI sequences (n = 30), (2) suboptimal Image preprocessing and radiomic feature extraction image quality (n = 6), (3) patients who received stereo- Preprocessing of the T2 and T1C images was per- tactic radiotherapy (n = 5), and (4) lack of Ki-67 labeling formed to standardize the data analysis across patients. P ark et al. Radiation Oncology (2022) 17:147 Page 3 of 12 Before analysis, the N4 bias correction algorithm was comprising a total of 200 radiomic features. A schematic applied  [20], and images were z-score normalized. of the data processing is shown in Fig. 1. Images were processed using an open-source software package (3D Slicer, version 4.11.0; available at: http:// Model construction and comparison of diagnostic slicer. org/). T1C images were coregistered to T2 images performance via affine transformation with normalized mutual infor - The number of radiomic features was larger than the mation as a cost function [21, 22]. The regions of inter - number of cases; therefore, the least absolute shrinkage est were drawn on every tumor section on the T1C and selection operator (LASSO) was applied to select the images using threshold-based and edge-based algo- significant features, which optimized the feature space rithms. Gross cystic, hemorrhagic, or necrotic areas by removing both irrelevant and redundant features [18]. of the tumors were included in the regions of interest The base radiomics classifiers were constructed using (ROIs). The segmentations were transferred to the T2 logistic regression with tenfold cross validation. In images. A neuroradiologist (8 years of experience) per- addition, each model was trained using random over- formed segmentation which was confirmed by a second sampling examples (ROSE) to overcome any data imbal- neuroradiologist (11  years of experience). Both neuro- ance, and hyperparameters were Bayesian optimized. To radiologists were blind to the corresponding clinical evaluate whether radiomics improves the prediction of information. models, two models were trained as follows: (1) a clinico- Discretization using a fixed bin number of 32 was pathological model trained on clinical features, includ- applied to extract the radiomic features  [23] from the ing age, extent of resection, ART status, and Ki-67 index; ROIs using an open-source Python-based module (PyRa- and (2) a combined model trained on clinicopathological diomics, version 2.0)  [24], which adhered to the Image and radiomics features. For statistical analysis, the Ki-67 Biomarker Standardization Initiative [25]. Fourteen shape index was dichotomized based on a cutoff value of 5% features, 18 first-order features, and 75  s-order features (≤ 5% vs. > 5%). Models were developed from the training (24  Gy-level co-occurrence matrices, 16  Gy-level run set (n = 92) and validated on the test set (n = 63). The area length matrices, 16  Gy-level size zone matrices, 14  Gy- under the receiver operating characteristic curve (AUC), level dependence matrices, and 5 neighborhood gray- accuracy, sensitivity, and specificity were obtained. The tone difference matrices) were extracted from the ROIs feature selection and machine learning process were per- in the T1C and T2 images (Additional file  1: Table  S2), formed using Python 3 (Python Software Foundation, Fig. 1 Workflow of image preprocessing, radiomics feature extraction, and machine learning. Park et al. Radiation Oncology (2022) 17:147 Page 4 of 12 Wilmington, Delaware, USA) with the Scikit-Learn value of 0.3 was obtained in the training set. The iden - library module (version 0.21.2). The performances of tical threshold value was applied to the test set, and the models were compared based on the AUC using patients were divided into low- and high-risk groups. DeLong’s method [26]. We retrospectively analyzed whether patients in low- and high-risk groups benefited from ART by comparing Model interpretability with SHapley Additive exPlanations the progression-free survival (PFS). PFS was defined To interpret and analyze the radiomic features of the as the time from initial surgery to tumor recurrence, radiomics model, SHapley Additive exPlanations (SHAP), death, or the last follow-up. Kaplan–Meier curves were which is a game theoretic approach to explain the out- generated and a log-rank test was performed to test the put of a tree-based machine learning model (Additional difference of PFS between patients who did and did not file  1: Supplementary Material S3), was applied  [27]. receive ART for each risk subgroup. SHAP measures the contribution of each feature of a model against the increase or decrease of the probability of a single output (i.e., the probability for tumor recur- Statistical analysis rence in our study) (Fig. 2). The Student’s t-test, Mann–Whitney U-test, and Chi- square test were performed to compare patient char- Analysis of survival outcomes and stratification acteristics between the responders and nonresponders for candidates for ART from training and test sets. The Kaplan–Meier method After constructing the best combined model for pre- was used to estimate survival rates, and the log-rank dicting tumor recurrence, the probability value of each test was performed to compare survival between the patient in the test set was analyzed to stratify candi- two groups. A P-value < 0.05 was considered statisti- dates for ART. Youden’s index was used for the optimal cally significant. All statistical analyses were performed cut-off selection threshold, and a cut-off probability using statistical software R (version 4.0.1; R Foundation for Statistical Computing, Vienna, Austria) and SPSS (version 25.0; SPSS Inc., Chicago, IL). Fig. 2 Comparison of median recurrence interval between patients with and without ART a in the entire cohort and b in high‑risk patients of the test set (probability > 0.3, according to the combined clinicopathological and radiomics model). Kaplan–Meier curves of PFS c comparing patients with and without ART in the entire cohort and d comparing patient subgroups according to the utility of ART in high‑risk patients. ART = adjuvant radiotherapy; PFS = progression‑free survival. Data are presented as the median with a 95% confidence interval for each group P ark et al. Radiation Oncology (2022) 17:147 Page 5 of 12 Treatment outcomes and prognostic factors Results With a median follow-up of 63.8  months (range, 6.6– Patient characteristics 190.7 months), there were 21 patients with tumor recur- Among 155 patients (mean age 56.9 ± 14.3, 102 females rence (12 [13.0%] in the training set, and 9 [14.3%] in the and 53 males), 148 had atypical meningiomas, 6 had test set). Fifteen events occurred within 24  months (7 choroid meningioma, and 1 had clear cell meningi- in the training set, and 8 in the test set, P = 0.292). The oma. In the training set, GTR was more frequently 5-year PFS rate was 79.1% for the entire cohort, and performed in patients without tumor recurrence com- 79.9% and 78.9% for the training and test set, respectively pared to those with tumor recurrence (87.5% vs. 58.3%, (P = 0.529). P = 0.011). No other clinicopathological features were In the training set, tumor recurrences occurred more different between patients with and without tumor frequently in patients who underwent STR than in those recurrence. In the test set, GTR was more frequently who underwent GTR (33.3% vs. 9.1%, P = 0.011). The performed in patients without tumor recurrence com- recurrence rate was higher in patients with a Ki-67 index pared to those with tumor recurrence (90.7% vs. 66.7%, value of 5% or more, although it was not statistically sig- P = 0.045). Patients with tumor recurrence were sig- nificant (19.1% vs. 6.7%, P = 0.076). In the test set, the nificantly older (P = 0.020), showed a male predomi- recurrence rate was higher in patients who underwent nance (P = 0.031), and had a higher Ki-67 labeling index STR (37.5% vs. 10.9%, P = 0.045) and in patients with (≥ 5%) (P = 0.018). ART was performed in 56 (60.9%) Ki-67 ≥ 5% (24.2% vs. 3.3%, P = 0.018). and 41 patients (65.1%) in the training and test sets, respectively. There were no significant differences Radiomics model construction of recurrences between the training and test sets with respect to the and comparison of diagnostic performance clinical and pathologic variables. The clinicopathologi - In the clinicopathological model, the AUC, accuracy, sen- cal characteristics of the training and test sets are sum- sitivity, and specificity were 0.67 (95% confidence interval marized in Table 1. [CI] 0.60–0.74), 66.9%, 48.2%, and 85.3% in the training Table 1 Patient characteristics in the training and test sets a a b Clinical variables Training set (n = 92) P‑ value Test set (n = 63) P‑ value P‑ value Without tumor With tumor Without tumor With tumor recurrence recurrence recurrence recurrence (n = 80) (n = 12) (n = 54) (n = 9) Age (years) 57.3 ± 14.6 54.4 ± 13.6 0.530 55.7 ± 14.1 67.7 ± 13.2 0.020 0.838 Female ratio 52 (65.0%) 9 (75.0%) 0.494 38 (70.4%) 3 (33.3%) 0.031 0.875 Extent of resection 0.011 0.045 0.535 GTR 70 (87.5%) 7 (58.3%) 49 (90.7%) 6 (66.7%) STR 10 (12.5%) 5 (41.7%) 5 (9.3%) 3 (33.3%) Ki‑67 labeling index 6.03 ± 4.81 6.96 ± 2.73 0.515 5.9 ± 4.3 8.2 ± 3.1 0.131 0.874 < 5% 32 (40.0%) 3 (25.0%) 0.076 29 (53.7%) 1 (11.1%) 0.018 ≥ 5% 38 (47.5%) 9 (75.0%) 25 (46.3%) 8 (88.9%) ART 0.659 0.517 0.595 Performed 48 (60.0%) 8 (66.7%) 36 (66.7%) 5 (55.6%) Not performed 32 (40.0%) 4 (33.3%) 18 (33.3%) 4 (44.4%) ART modality 0.851 0.061 0.813 3D‑ CRT 5 (10.4%) 2 (25.0%) 2 (5.6%) 2 (40.0%) IMRT 43 (89.6%) 6 (75.0%) 34 (94.4%) 3 (60.0%) ART dose (Gy) 57.3 ± 4.0 57.2 ± 6.7 0.956 58.3 ± 3.9 58.7 ± 2.6 0.853 0.204 Data are expressed as the mean with standard deviation in parentheses, median with interquartile range in parentheses, or number with percentage in parentheses Calculated from Student’s-t test for continuous variables and Chi-square test for categorical variables to compare the patient characteristics between the responder and non-responders from each training and test set Calculated from Student’s-t test for continuous variables and Chi-square test for categorical variables for the comparison of training and test sets Data obtained from patients who underwent adjuvant radiotherapy following surgery GTR: gross total resection; STR: subtotal resection; ART: adjuvant radiotherapy; RT: radiotherapy; 3D-CRT: three-dimensional conformal radiotherapy; IMRT: intensity modulated radiotherapy Park et al. Radiation Oncology (2022) 17:147 Page 6 of 12 set, respectively, and 0.61 (95% CI 0.44–0.78), 81.3%, 90th percentile from T1C, and Ki-67 index are the 33.3%, and 89.1%, respectively, in the test set. three most important risk factors. In the combined clinicopathological and radiomics model, a total of seven features were selected: two clini- cal features (extent of resection [GTR or STR] and Ki-67 Selection of candidates for ART using the developed index [≥ 5% or < 5%]) and five radiomic features, which combined model th were all first-order features from T1C (10 percentile, In the test set, there were 49 patients in the high-risk th 90 percentile, entropy, mean absolute deviation, and group and 14 patients in the low-risk group accord- minimum) (Additional file  1: Table  S3). The AUC, accu - ing to the combined clinicopathological and radiomics racy, sensitivity, and specificity were 0.78 (95% CI 0.70– model. The characteristics of patients in the high- and 0.85), 75.0%, 76.8%, and 73.1% in the training set, and low-risk groups are summarized in Additional file  1: 0.77 (95% CI 0.60–0.94), 70.3%, 66.7%, and 70.9%, respec- Table  S4. Among all patients, those who received ART tively, in the test set. exhibited a significantly longer PFS (5-year PFS, 87.8% In the training set, the combined model achieved supe- vs. 65.4%, P = 0.035) and delayed recurrence (median rior performance compared with the clinicopathological time to recurrence 45.3 vs. 18.4 months) (Fig.  2a, c). In model (AUC: 0.78 vs. 0.67, P = 0.042) (Fig. 3, Table 2). In the high-risk group, the PFS was significantly longer the test set, the combined model trended toward better in patients who received ART (5-year PFS, 92.3% vs. performance in the test set than in the clinicopathologi- 56.8%, P = 0.024) (Fig .  2b), with significantly delayed cal model (AUC: 0.77 vs. 0.61, P = 0.192) without sta- recurrence (median 53.5 vs. 17.0  months). In the low- tistical significance. The diagnostic performances of the risk group, there was no significant difference in the two models in the training and test set are provided in PFS (P = 0.264) or recurrence interval with respect to Table 2. ART. Only one patient experienced tumor recurrence (39.9 months after surgery). Model interpretability with SHAP When the cohort of patients was divided into four The SHAP values for each selected feature in the com - groups according to ART status and recurrence prob- bined clinicopathological and radiomics model were ability, the PFS was observed to significantly improve calculated. The variance importance plot, summary after ART in patients with a high recurrence probability plot, decision plot, and force plot of the test set are (median 8.7  year, 5-year 92.3%) (Fig.  2d). The signifi - shown in Fig.  4. For each prediction, a positive SHAP cant improvement in PFS was observed regardless of value indicates an increase in the risk of tumor recur- the extent of resection (GTR: 5-year 91.3% vs. 63.5%, rence while a negative SHAP value indicates reduced P = 0.054; STR: 5-year 100.0% vs. 0.0%, P = 0.012). risk. As observed in the plots, the extent of resection, Fig. 3 Receiver operating characteristic curves of the radiomics model in the a training and b test sets P ark et al. Radiation Oncology (2022) 17:147 Page 7 of 12 ‑ ‑ ‑ Table 2 Performances of machine learning models for prediction of tumor recurrence in the training and test set Models Training set Test set AUC (95% CI) Accuracy (%) Sensitivity (%) Specificity (%) P value AUC (95% CI) Accuracy (%) Sensitivity (%) Specificity (%) P value Clinicopatho 0.67 (0.59–0.74) 66.9 48.2 85.3 Reference 0.61 (0.44–0.78) 81.3 33.3 89.1 Reference logical model Clinicopatho 0.78 (0.70–0.85) 75.0 76.8 31.0 0.042 0.77 (0.60–0.94) 70.3 66.7 70.9 0.192 logical + radi omics model AUC: area under the curve; CI: confidence interval; NRI: net reclassification index Park et al. Radiation Oncology (2022) 17:147 Page 8 of 12 (See figure on next page.) Fig. 4 Model interpretability of a combined clinicopathological and radiomics model for the prediction of tumor recurrence with SHAP in the training set. a Variance importance plot that lists the most significant variables in descending order. b Summary plot of feature impact on the decision of the model and interaction between the features in the model. A positive SHAP value indicates an increase in the probability of tumor recurrence. c Decision plot showing how the model predicts tumor recurrence. Starting at the bottom of the plot, the prediction line shows how the SHAP values accumulate from the base value to arrive at the model’s final score at the top of the plot and how each feature contributes to the overall prediction of tumor recurrence. d Force plot of a representative case of a patient with tumor recurrence. Red arrows represent feature effects that drive the prediction value higher, and blue arrows are those effects that drive the prediction value lower. Each arrow’s size represents the magnitude of the corresponding feature’s effect. Note that the extent of resection, 90th percentile from T1C, and Ki‑67 labeling index largely push the model prediction score higher than the base value intensified adjuvant therapies, and patients at low-risk Discussion for tumor recurrence can be spared from experienc- In this study, radiomic features derived from preopera- ing the possible neurotoxicities of ART  [38]. Several tive MRI were applied to predict tumor recurrence in previous studies investigated clinical prognostic fac- grade 2 meningiomas. The model incorporating radi - tors for grade 2 meningiomas, which included age, sex, omic features with clinicopathologic features exhibited extent of resection, tumor invasiveness (i.e., brain or significantly a higher prediction accuracy for tumor bone invasion), and higher MIB-1 labeling index  [39, recurrence. Several risk factors for tumor recurrence 40]. Similar to our study, the clinical significance of were identified using SHAP, and the contribution of the Ki-67 labeling index on local control and survival each feature to the probability of tumor recurrence in high-grade meningiomas has been widely reported. was determined. Further, radiomics enabled a subset A higher Ki-67 labeling index seems to be related to of patients at a high risk for tumor recurrence to be invasiveness, which is closely correlated with an incom- identified, and we observed that the high-risk patients plete resection rate. Although there is some discrep- clearly benefitted from ART. Therefore, radiomics can ancy in the exact cutoff point, several studies have serve as a potential imaging biomarker, as well as a use- suggested the need for ART in patients with a higher ful tool for selecting adequate candidates for ART in Ki-67  [41, 42]. Pretreatment tumor size or volume was patients with meningiomas. also reported as a significant predictor of prognosis [35, Although the majority of meningiomas are benign 43–45]. However, in our study that analyzed various and slow growing, WHO grade 2 meningiomas are radiomic features including tumor size, it was not a sig- considered to have a high recurrence rate of 55% and nificant prognostic factor compared to other features. low survival rates [28–30]. ART after the STR of grade Several studies attempted to identify the prognostic 2 meningiomas is widely practiced; however, the value of radiomics in patients with meningiomas [16, 17, impact of ART on grade 2 meningiomas after a GTR 46]. However, these previous studies combined grade 1–3 remains contentious. Several researchers demonstrated meningiomas and do not provide specific information on improved local control and/or survival  [8–10, 30–33] the utility of the radiomics model for grade 2 meningi- with ART after GTR, but a subset exhibited contrary omas. Moreover, neither the combined clinicopathologi- results after ART  [9, 34–36]. According to two multi- cal model nor radiomics model showed any improved institutional phase II studies (NRG Oncology RTOG performance in predicting prognosis compared with the 0539  [32] and EORTC 22,042–26,042  [37]), which clinicopathological model alone, limiting the real-world enrolled grade 2 meningioma patients who underwent application of the radiomics model  [17, 46]. Decision- GTR and ART, there is a potential survival benefit with making for the treatment of grade 2 meningioma patients a 3-year PFS of 89%–94%. However, those studies did depends on multilevel prognostic information and clin- not directly compare ART with upfront observation, icopathological information, such as the extent of the and risk factors other than surgical resection were tumor resection or Ki-67 index; therefore, we focused not addressed. Therefore, a uniform treatment para - on the role of radiomics in predicting survival given the digm for resected grade 2 meningiomas has yet to be multilevel prognostic information. Our results show that established. radiomics significantly increases the model performance Because there is a lack of standardized treatment when added to a clinicopathological model, thus promot- for resected grade 2 meningiomas, performing nonin- ing the integration of radiomics in clinical practice. vasive risk stratification prior to adjuvant therapies is Radiomics models have limited explainability highly desirable. Even after surgical resection, patients because they rely on complex machine learning algo- at high-risk for tumor recurrence can be considered for rithms, resulting in low clinical utility  [47]. SHAP can P ark et al. Radiation Oncology (2022) 17:147 Page 9 of 12 Fig. 4 (See legend on previous page.) Park et al. Radiation Oncology (2022) 17:147 Page 10 of 12 uncover complex underlying patterns  [27, 48] and has Conclusions recently been utilized to interpret radiomics mod- Multiparametric MRI radiomics has an added prognostic els  [49, 50]. This analysis identified two clinical fea- value for the prediction of tumor recurrence when inte- tures—the extent of resection and the Ki-67 labeling grated with clinicopathologic profiles in patients with index—and five radiomic first-order features from grade 2 meningiomas. With radiomics, we could identify T1C that contributed the most towards the predic- a subset of patients at a high risk for tumor recurrence tion of tumor recurrence. The extent of resection and who might benefit from intensified treatment. Therefore, the Ki-67 index are well-known prognostic factors for radiomics may be used as a potential imaging biomarker tumor recurrence  [35, 51], which was also verified by in patients with grade 2 meningiomas. th th our study. The 90 and 10 percentile values from T1C, which are the two most important radiomic fea- Abbreviations tures, denote the intensity of contrast enhancement. ART : Adjuvant radiotherapy; GTR : Gross total resection; LASSO: Least absolute Heterogeneous contrast enhancement or the presence shrinkage and selection operator; PFS: Progression‑free survival; RANO: Response assessment in neuro‑ oncology; ROI: Region of interest; ROSE: of contrast enhancement of meningiomas  [17] has Random over‑sampling examples; SHAP: Shapley Additive Explanations; STR: shown significant association with high-grade menin- Subtotal resection; T1C: Post‑ contrast T1‑ weighted image; T2: T2‑ weighted giomas or tumor invasiveness  [52, 53]; therefore, the image; WHO: World Health Organization. meningioma grade may be indicated by the percentile- related first-order features, which reflect the distribu- Supplementary Information tion of the contrast-enhancement degree. The online version contains supplementary material available at https:// doi. org/ 10. 1186/ s13014‑ 022‑ 02090‑7. Our study has several limitations. First, our study was based on a single-center, retrospectively col- Additional file 1. Supplementary figures and tables. lected dataset. Further studies with a larger num- ber of patients and external validations are required. Acknowledgements Second, the histopathologic risk factors, such as the None MIB-1 labeling index or mitoses index, were not evalu- Author contributions ated in our study because relevant information for the SHC and CJP analyzed, interpreted the patient data and wrote the manuscript. majority of patients were unavailable during the long SHC, CJP, JE, and HKB are contributors for analysis. YWP and HIY are major study-inclusion period. To prove the predictive val- contributors for conceptualization and interpretation. All authors read and approved the final manuscript. ues of radiomics, it is highly desirable to incorporate detailed histopathologic features into the models in Funding future studies. Third, the probability scores derived This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of from the developed model for tumor recurrence were Education (Republic of Korea; grant number: NRF‑2020R1I1A1A0107164811 to only calculated in the test set, not in the training set, Y. W. Park) and the National Research Foundation of Korea (NRF) Grant funded because the training set was oversampled during the by the Korea government (MSIT ) (NRF‑2021R1A2C1010900 to S.H. Choi). This study was also supported by a faculty research grant of Yonsei University Col‑ development of the model. Therefore, the benefit of lege of Medicine for (6‑2020‑0115). ART was only evaluated in the test set, which inevita- bly decreased the sample size. However, based on our Availability of data and materials Research data are stored in an institutional repository and will be shared upon study results, we at least found a benefit of ART in a request to the corresponding author. subgroup of the test set, which was identified through the combined model. Future studies with larger sample Declarations sizes are required to identify the patients with grade 2 meningiomas who might benefit from ART. Ethics approval and consent to participate This study was conducted according to the guidelines of the Declaration of Our study is the first to show the possibility of radi- Helsinki, and approved by the institutional review board of the Yonsei Univer‑ omics playing an important role in selecting candidate sity Health System (9‑2021‑0047). patients with meningiomas for ART. Over a similar Consent for publication follow-up period (median 54 months in ART group vs. Not applicable. 61 months in non-ART group, P = 0.584), we observed that ART prolongs PFS and significantly delays recur- Competing interests None of the authors have any conflicts of interest to disclose. rence in high-risk patients. Our combined predictive model could be applied effectively and conveniently in Author details both GTR and STR patients. Department of Radiology, Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yongin Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea. Depar tment of Radiation Oncology, Yongin Severance Hospital, Yonsei University College P ark et al. Radiation Oncology (2022) 17:147 Page 11 of 12 of Medicine, Seoul, Republic of Korea. Department of Computer Science, Yon‑ 19. Huang RY, Bi WL, Weller M, Kaley T, Blakeley J, Dunn I, et al. Proposed sei University, Seoul, Republic of Korea. Department of Radiation Oncology, response assessment and endpoints for meningioma clinical trials: Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University report from the Response Assessment in Neuro‑ Oncology Working College of Medicine, 50‑1 Yonsei‑ro, Seodaemun‑gu, Seoul 03722, Republic Group. Neuro Oncol. 2019;21(1):26–36. of Korea. Department of Neurosurgery, Yonsei University College of Medicine, 20. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, 50‑1 Yonsei‑ro, Seodaemun‑gu, Seoul 03722, Republic of Korea. Depar tment et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. of Pathology, Yonsei University College of Medicine, Seoul, Republic of Korea. 2010;29(6):1310–20. 21. Cha J, Kim ST, Kim HJ, Kim BJ, Kim YK, Lee JY, et al. Differentiation of Received: 14 March 2022 Accepted: 23 June 2022 tumor progression from pseudoprogression in patients with posttreat‑ ment glioblastoma using multiparametric histogram analysis. AJNR Am J Neuroradiol. 2014;35(7):1309–17. 22. Maes F, Collignon A, Vandermeulen D, Marchal G, Suetens P. Multimo‑ dality image registration by maximization of mutual information. IEEE Trans Med Imaging. 1997;16(2):187–98. References 23. Carré A, Klausner G, Edjlali M, Lerousseau M, Briend‑Diop J, Sun R, et al. 1. Ostrom QT, Gittleman H, Liao P, Rouse C, Chen Y, Dowling J, et al. CBTRUS Standardization of brain MR images across machines and protocols: statistical report: primary brain and central nervous system tumors diag‑ bridging the gap for MRI‑based radiomics. Sci Rep. 2020;10(1):12340. nosed in the United States in 2007–2011. Neuro Oncol. 2014;16(Suppl 24. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan 4):iv1‑63. V, et al. Computational radiomics system to decode the radiographic 2. Ostrom QT, Gittleman H, Truitt G, Boscia A, Kruchko C, Barnholtz‑Sloan JS. phenotype. Cancer Res. 2017;77(21):e104–7. CBTRUS statistical report: primary brain and other central nervous system 25. Zwanenburg A, Leger S, Vallières M, Löck S. Image biomarker standardi‑ tumors diagnosed in the United States in 2011–2015. Neuro Oncol. sation initiative. arXiv preprint arXiv: 1612. 07003. 2016. 2018;20(suppl_4):iv1–86. 26. DeLong ER, DeLong DM, Clarke‑Pearson DL. Comparing the areas 3. Goldbrunner R, Stavrinou P, Jenkinson MD, Sahm F, Mawrin C, Weber DC, under two or more correlated receiver operating characteristic curves: et al. EANO guideline on the diagnosis and management of meningi‑ a nonparametric approach. Biometrics. 1988;44(3):837–45. omas. Neuro Oncol. 2021;23(11):1821–34. 27. Lundberg SM, Lee S‑I. A unified approach to interpreting model pre ‑ 4. Stark AM, Buhl R, Mehdorn HM. Early gross recurrence of atypical menin‑ dictions. In. Proceedings of the 31st international conference on neural gioma. J Neurooncol. 2005;75(2):223–4. information processing systems; 2017. pp. 4768–77. 5. Jenkinson MD, Waqar M, Farah JO, Farrell M, Barbagallo GM, McManus R, 28. Modha A, Gutin PH. Diagnosis and treatment of atypical and anaplastic et al. Early adjuvant radiotherapy in the treatment of atypical meningi‑ meningiomas: a review. Neurosurgery. 2005;57(3):538–50. oma. J Clin Neurosci. 2016;28:87–92. 29. Rogers L, Barani I, Chamberlain M, Kaley TJ, McDermott M, Raizer J, 6. Lee KD, DePowell JJ, Air EL, Dwivedi AK, Kendler A, McPherson CM. Atypi‑ et al. Meningiomas: knowledge base, treatment outcomes, and uncer‑ cal meningiomas: is postoperative radiotherapy indicated? Neurosurg tainties. A RANO review. J Neurosurg. 2015;122(1):4–23. Focus. 2013;35(6):E15. 30. Aizer AA, Bi WL, Kandola MS, Lee EQ, Nayak L, Rinne ML, et al. Extent of 7. Goldbrunner R, Stavrinou P, Jenkinson MD, Sahm F, Mawrin C, Weber DC, resection and overall survival for patients with atypical and malignant et al. EANO guideline on the diagnosis and management of meningi‑ meningioma. Cancer. 2015;121(24):4376–81. omas. Neuro Oncol. 2021. https:// doi. org/ 10. 1093/ neuonc/ noab1 50. 31. Li H, Zhang YS, Zhang GB, Zhang GJ, Wang B, Li D, et al. Treatment 8. Park HJ, Kang HC, Kim IH, Park SH, Kim DG, Park CK, et al. The role protocol, long‑term follow‑up, and predictors of mortality in 302 cases of adjuvant radiotherapy in atypical meningioma. J Neurooncol. of atypical meningioma. World Neurosurg. 2019;122:e1275–84. 2013;115(2):241–7. 32. Rogers L, Zhang P, Vogelbaum MA, Perry A, Ashby LS, Modi JM, et al. 9. Bagshaw HP, Burt LM, Jensen RL, Suneja G, Palmer CA, Couldwell WT, Intermediate‑risk meningioma: initial outcomes from NRG Oncology et al. Adjuvant radiotherapy for atypical meningiomas. J Neurosurg. RTOG 0539. J Neurosurg. 2018;129(1):35–47. 2017;126(6):1822–8. 33. Rydzewski NR, Lesniak MS, Chandler JP, Kalapurakal JA, Pollom E, 10. Choi Y, Lim DH, Jo K, Nam DH, Seol HJ, Lee JI. Efficacy of postop ‑ Tate MC, et al. Gross total resection and adjuvant radiotherapy most erative radiotherapy for high grade meningiomas. J Neurooncol. significant predictors of improved survival in patients with atypical 2014;119(2):405–12. meningioma. Cancer. 2018;124(4):734–42. 11. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, 34. Zhu H, Bi WL, Aizer A, Hua L, Tian M, Den J, et al. Efficacy of adjuvant they are data. Radiology. 2016;278(2):563–77. radiotherapy for atypical and anaplastic meningioma. Cancer Med. 12. Park YW, Han K, Ahn SS, Choi YS, Chang JH, Kim SH, et al. Whole‑tumor 2019;8(1):13–20. histogram and texture analyses of DTI for evaluation of IDH1‑mutation 35. Zhi M, Girvigian MR, Miller MJ, Chen JC, Schumacher AJ, Rahimian and 1p/19q‑ codeletion status in world health organization grade II J, et al. Long‑term outcomes of newly diagnosed resected atypical gliomas. AJNR Am J Neuroradiol. 2018;39(4):693–8. meningiomas and the role of adjuvant radiotherapy. World Neurosurg. 13. Won SY, Park YW, Ahn SS, Moon JH, Kim EH, Kang SG, et al. Quality 2019;122:e1153–61. assessment of meningioma radiomics studies: Bridging the gap between 36. Graffeo CS, Leeper HE, Perry A, Uhm JH, Lachance DJ, Brown PD, exploratory research and clinical applications. Eur J Radiol. 2021;138: et al. Revisiting adjuvant radiotherapy after gross total resection of world health organization grade II meningioma. World Neurosurg. 14. Park YW, Oh J, You SC, Han K, Ahn SS, Choi YS, et al. Radiomics and 2017;103:655–63. machine learning may accurately predict the grade and histological sub‑ 37. Weber DC, Ares C, Villa S, Peerdeman SM, Renard L, Baumert BG, et al. type in meningiomas using conventional and diffusion tensor imaging. Adjuvant postoperative high‑ dose radiotherapy for atypical and malig‑ Eur Radiol. 2019;29(8):4068–76. nant meningioma: A phase‑II parallel non‑randomized and observa‑ 15. Gu H, Zhang X, di Russo P, Zhao X, Xu T. The current state of radiomics for tion study (EORTC 22042–26042). Radiother Oncol. 2018;128(2):260–5. meningiomas: promises and challenges. Front Oncol. 2020;10: 567736. 38. Wu A, Jin MC, Meola A, Wong H‑n, Chang SD. Efficacy and toxicity of 16. Zhang Y, Chen JH, Chen TY, Lim SW, Wu TC, Kuo Y T, et al. Radiomics particle radiotherapy in WHO grade II and grade III meningiomas: a approach for prediction of recurrence in skull base meningiomas. Neuro‑ systematic review. Neurosurgical Focus FOC. 2019;46(6):E12. radiology. 2019;61(12):1355–64. 39. Wang F, Xu D, Liu Y, Lin Y, Wei Q, Gao Q, et al. Risk factors associated 17. Morin O, Chen WC, Nassiri F, Susko M, Magill ST, Vasudevan HN, et al. with postoperative recurrence in atypical intracranial meningioma: Integrated models incorporating radiologic and radiomic features predict analysis of 263 cases at a single neurosurgical centre. Acta Neurochir. meningioma grade, local failure, and overall survival. Neurooncol Adv. 2019;161(12):2563–70. 2019;1(1):vdz011. 40. Ros‑Sanjuan A, Iglesias‑Moroño S, Carrasco ‑Brenes A, Bautista‑ Ojeda D, 18. Friedman J, Hastie T, Tibshirani R. Regularization paths for generalized Arraez‑Sanchez MA. Atypical meningiomas: histologic and clinical fac‑ linear models via coordinate descent. J Stat Softw. 2010;33(1):1–22. tors associated with recurrence. World Neurosurg. 2019;125:e248–56. Park et al. Radiation Oncology (2022) 17:147 Page 12 of 12 41. Choi Y, Lim DH, Yu JI, Jo K, Nam DH, Seol HJ, et al. Prognostic value of Ki‑67 labeling index and postoperative radiotherapy in WHO grade II meningioma. Am J Clin Oncol. 2018;41(1):18–23. 42. Bruna J, Brell M, Ferrer I, Gimenez‑Bonafe P, Tortosa A. Ki‑67 proliferative index predicts clinical outcome in patients with atypical or anaplastic meningioma. Neuropathology. 2007;27(2):114–20. 43. Detti B, Scoccianti S, Di Cataldo V, Monteleone E, Cipressi S, Bordi L, et al. Atypical and malignant meningioma: outcome and prognostic factors in 68 irradiated patients. J Neurooncol. 2013;115(3):421–7. 44. Milker‑Zabel S, Zabel A, Schulz‑Ertner D, Schlegel W, Wannenmacher M, Debus J. Fractionated stereotactic radiotherapy in patients with benign or atypical intracranial meningioma: long‑term experience and prognos‑ tic factors. Int J Radiat Oncol Biol Phys. 2005;61(3):809–16. 45. Zhang GJ, Zhang YS, Zhang GB, Yan XJ, Li CB, Zhang LW, et al. Prognostic factors, survival, and treatment for intracranial world health organiza‑ tion grade ii chordoid meningiomas and clear‑ cell meningiomas. World Neurosurg. 2018;117:e57–66. 46. Gennatas ED, Wu A, Braunstein SE, Morin O, Chen WC, Magill ST, et al. Preoperative and postoperative prediction of long‑term meningioma outcomes. PLoS ONE. 2018;13(9): e0204161. 47. Du R, Lee VH, Yuan H, Lam K‑ O, Pang HH, Chen Y, et al. Radiomics model to predict early progression of nonmetastatic nasopharyngeal carcinoma after intensity modulation radiation therapy: a multicenter study. Radiol Artif Intell. 2019;1(4):e180075. 48. Arrow K, Barankin E, Blackwell D. Admissible points of convex sets. Con‑ trib Theory Games. 1953;2:87–91. 49. Bang M, Eom J, An C, Kim S, Park YW, Ahn SS, et al. An interpretable mul‑ tiparametric radiomics model for the diagnosis of schizophrenia using magnetic resonance imaging of the corpus callosum. Transl Psychiatry. 2021;11(1):462. 50. Giraud P, Giraud P, Nicolas E, Boisselier P, Alfonsi M, Rives M, et al. Inter‑ pretable machine learning model for locoregional relapse prediction in oropharyngeal cancers. Cancers. 2021;13(1):57. 51. Mirian C, Skyrman S, Bartek J Jr, Jensen LR, Kihlström L, Förander P, et al. The Ki‑67 proliferation index as a marker of time to recurrence in intracra‑ nial meningioma. Neurosurgery. 2020;87(6):1289–98. 52. Adeli A, Hess K, Mawrin C, Streckert EMS, Stummer W, Paulus W, et al. Pre‑ diction of brain invasion in patients with meningiomas using preopera‑ tive magnetic resonance imaging. Oncotarget. 2018;9(89):35974–82. 53. Liu Y, Chotai S, Chen M, Jin S, Qi S‑t, Pan J. Preoperative radiologic clas‑ sification of convexity meningioma to predict the survival and aggressive meningioma behavior. PLoS ONE. 2015;10(3):e0118908‑ e. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ lished maps and institutional affiliations. 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 Radiation Oncology Springer Journals

An interpretable radiomics model to select patients for radiotherapy after surgery for WHO grade 2 meningiomas

Loading next page...
 
/lp/springer-journals/an-interpretable-radiomics-model-to-select-patients-for-radiotherapy-Pq2gYzxB6h

References (55)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2022
eISSN
1748-717X
DOI
10.1186/s13014-022-02090-7
Publisher site
See Article on Publisher Site

Abstract

Objectives: This study investigated whether radiomic features can improve the prediction accuracy for tumor recurrence over clinicopathological features and if these features can be used to identify high‑risk patients requiring adjuvant radiotherapy (ART ) in WHO grade 2 meningiomas. Methods: Preoperative magnetic resonance imaging (MRI) of 155 grade 2 meningioma patients with a median follow‑up of 63.8 months were included and allocated to training (n = 92) and test sets (n = 63). After radiomic feature extraction (n = 200), least absolute shrinkage and selection operator feature selection with logistic regression classifier was performed to develop two models: (1) a clinicopathological model and (2) a combined clinicopathological and radiomic model. The probability of recurrence using the combined model was analyzed to identify candidates for ART. Results: The combined clinicopathological and radiomics model exhibited superior performance for the prediction of recurrence compared with the clinicopathological model in the training set (area under the curve [AUC] 0.78 vs. 0.67, P = 0.042), which was also validated in the test set (AUC 0.77 vs. 0.61, P = 0.192). In patients with a high prob‑ ability of recurrence by the combined model, the 5‑ year progression‑free survival was significantly improved with ART (92% vs. 57%, P = 0.024), and the median time to recurrence was longer (54 vs. 17 months after surgery). Conclusions: Radiomics significantly contributes added value in predicting recurrence when integrated with the clinicopathological features in patients with grade 2 meningiomas. Furthermore, the combined model can be applied to identify high‑risk patients who require ART. Keywords: Magnetic resonance imaging, Meningioma, Radiomics, Radiotherapy, Prognosis Introduction Meningiomas are the most common primary intracranial neoplasms in adults, comprising 36.7% of all intracranial tumors [1]. Since the serial updates of the World Health Chae Jung Park and Seo Hee Choi have contributed equally to this paper Organization (WHO) grading classification, the propor - *Correspondence: yaewonpark@yuhs.ac; YHI0225@yuhs.ac tion of grade 2 meningiomas has gradually increased up to 15–20% [2]. However, despite recent WHO grading Department of Radiology, Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yongin Severance Hospital, scheme, there are limitations in predicting the prognosis Yonsei University College of Medicine, Seoul, Republic of Korea of grade 2 meningiomas [3]. Grade 2 meningiomas are Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion known to have an unpredictable heterogeneous disease Therapy Research Institute, Yonsei University College of Medicine, 50‑1 Yonsei‑ro, Seodaemun‑gu, Seoul 03722, Republic of Korea course; even after gross total resection (GTR), recurrence Full list of author information is available at the end of the article © The Author(s) 2022. 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. Park et al. Radiation Oncology (2022) 17:147 Page 2 of 12 can occur in a substantial number of patients  [4, 5], index (n = 3). Finally, a total of 155 patients were enrolled whereas some patients experience a long indolent clinical (Additional file  1: Fig. S1). The study population was ran - course without adjuvant treatment [6]. domly divided into training and test sets with a ratio of Currently, adjuvant radiotherapy (ART) is recom- 6:4 (n = 92 and n = 63, respectively). mended as a standard therapy for patients who undergo Preoperative MRI was performed using a 3.0-T MRI subtotal resection (STR) in grade 2 meningiomas  [7] scanner (Achieva, Philips Medical Systems, Amsterdam, based on evidence indicating improvement in local con- Netherlands) with an eight-channel sensitivity-encoding trol and survival rates with ART  [8–10]. However, there head coil. Detailed parameters of the MRI sequences are is a lack of consensus on whether ART reduces the risk of provided in Additional file 1: Supplementary Material S1. tumor recurrence after GTR of grade 2 meningiomas [3]. u Th s, it is crucial to establish a model for predicting a Treatments patient’s individual outcome and to identify high-risk All patients underwent either GTR (n = 132) or STR patients who could benefit from ART even after GTR (n = 23) surgery. The extent of resection was defined by of grade 2 meningiomas. Patients who are at a low risk comparing preoperative and postoperative MRI scans; for recurrence can be spared from ART and its potential GTR was defined by a lack of residual enhancing tumor risks. Furthermore, in a set of patients who have under- in the image, and STR was defined by the patient having gone STR, patient subsets with a higher risk of recur- more than 50% of the tumor removed [18]. After surgical rence may undergo more intensified radiotherapy to resection, a multidisciplinary team, consisting of neuro- improve their respective outcomes. surgeons, radiation oncologists, neuropathologists, and Radiomics is an advanced technique that extracts neuroradiologists, decided whether to perform ART, high-dimensional quantitative imaging features, such which consisted of three-dimensional conformal radio- as intensity distributions, spatial relationships, textural therapy and intensity-modulated radiotherapy. A total of heterogeneity, and shape descriptors  [11, 12]. Radiom- 97 patients underwent ART; 80 patients received it after ics aims to discover meaningful “hidden” information GTR (80/132, 60.6%), whereas 17 received it after STR within radiological images that is visually inaccessible. (17/23, 73.9%). Among the patients who underwent ART, Previous studies have demonstrated the use of radiomics intensity-modulated radiotherapy was performed on 86 in differential diagnosis and grade prediction for men - patients (88.7%, median 60.0 Gy), and three-dimensional ingiomas  [13–15]. Some studies have also shown that conformal radiotherapy was performed on 11 patients radiomics can predict tumor recurrence in patients with (11.3%, median 59.4 Gy). meningiomas  [16, 17]. We hypothesized that radiomics can enable risk stratification for tumor recurrence after Response assessment by a neuro‑oncology meningioma surgery in patients with grade 2 meningiomas, which working group may guide patients toward ART. The evaluation of the tumor response and progression Therefore, this study aimed to investigate whether radi - was determined according to the Response Assessment omics features can be used to improve the prediction of in Neuro-Oncology (RANO) criteria  [19] by compar- tumor recurrence over clinicopathological features, and ing serial MRIs of each patient. A radiation oncologist if radiomics can be used to identify high-risk patients and neuroradiologist (with 9 and 8  years of experience, who require ART in WHO grade 2 meningiomas. respectively) performed the evaluation and the results were achieved by consensus. A detailed description of Methods the RANO criteria and the definitions of each assessment Patient population criteria are presented in the Additional file  1: Supple- This study was approved by the institutional review board mentary material S2 and Additional file  1: Table  S1. The of the Yonsei University Health System (9-2021-0047). patients assessed as having a progressive disease in the The institutional review board waived the requirement follow-up periods were considered to have tumor recur- to obtain informed patient consent for this retrospec- rence. Patients assessed as having complete response, tive study. We retrospectively reviewed 199 patients partial response, minor response, and stable disease were with surgically confirmed WHO grade 2 meningiomas considered to be patients without tumor recurrence. The who underwent preoperative conventional magnetic primary endpoint was tumor recurrence assessed by the resonance imaging (MRI) between February 2005 and RANO criteria during the follow-up period. November 2018. The exclusion criteria were as follows: (1) incomplete MRI sequences (n = 30), (2) suboptimal Image preprocessing and radiomic feature extraction image quality (n = 6), (3) patients who received stereo- Preprocessing of the T2 and T1C images was per- tactic radiotherapy (n = 5), and (4) lack of Ki-67 labeling formed to standardize the data analysis across patients. P ark et al. Radiation Oncology (2022) 17:147 Page 3 of 12 Before analysis, the N4 bias correction algorithm was comprising a total of 200 radiomic features. A schematic applied  [20], and images were z-score normalized. of the data processing is shown in Fig. 1. Images were processed using an open-source software package (3D Slicer, version 4.11.0; available at: http:// Model construction and comparison of diagnostic slicer. org/). T1C images were coregistered to T2 images performance via affine transformation with normalized mutual infor - The number of radiomic features was larger than the mation as a cost function [21, 22]. The regions of inter - number of cases; therefore, the least absolute shrinkage est were drawn on every tumor section on the T1C and selection operator (LASSO) was applied to select the images using threshold-based and edge-based algo- significant features, which optimized the feature space rithms. Gross cystic, hemorrhagic, or necrotic areas by removing both irrelevant and redundant features [18]. of the tumors were included in the regions of interest The base radiomics classifiers were constructed using (ROIs). The segmentations were transferred to the T2 logistic regression with tenfold cross validation. In images. A neuroradiologist (8 years of experience) per- addition, each model was trained using random over- formed segmentation which was confirmed by a second sampling examples (ROSE) to overcome any data imbal- neuroradiologist (11  years of experience). Both neuro- ance, and hyperparameters were Bayesian optimized. To radiologists were blind to the corresponding clinical evaluate whether radiomics improves the prediction of information. models, two models were trained as follows: (1) a clinico- Discretization using a fixed bin number of 32 was pathological model trained on clinical features, includ- applied to extract the radiomic features  [23] from the ing age, extent of resection, ART status, and Ki-67 index; ROIs using an open-source Python-based module (PyRa- and (2) a combined model trained on clinicopathological diomics, version 2.0)  [24], which adhered to the Image and radiomics features. For statistical analysis, the Ki-67 Biomarker Standardization Initiative [25]. Fourteen shape index was dichotomized based on a cutoff value of 5% features, 18 first-order features, and 75  s-order features (≤ 5% vs. > 5%). Models were developed from the training (24  Gy-level co-occurrence matrices, 16  Gy-level run set (n = 92) and validated on the test set (n = 63). The area length matrices, 16  Gy-level size zone matrices, 14  Gy- under the receiver operating characteristic curve (AUC), level dependence matrices, and 5 neighborhood gray- accuracy, sensitivity, and specificity were obtained. The tone difference matrices) were extracted from the ROIs feature selection and machine learning process were per- in the T1C and T2 images (Additional file  1: Table  S2), formed using Python 3 (Python Software Foundation, Fig. 1 Workflow of image preprocessing, radiomics feature extraction, and machine learning. Park et al. Radiation Oncology (2022) 17:147 Page 4 of 12 Wilmington, Delaware, USA) with the Scikit-Learn value of 0.3 was obtained in the training set. The iden - library module (version 0.21.2). The performances of tical threshold value was applied to the test set, and the models were compared based on the AUC using patients were divided into low- and high-risk groups. DeLong’s method [26]. We retrospectively analyzed whether patients in low- and high-risk groups benefited from ART by comparing Model interpretability with SHapley Additive exPlanations the progression-free survival (PFS). PFS was defined To interpret and analyze the radiomic features of the as the time from initial surgery to tumor recurrence, radiomics model, SHapley Additive exPlanations (SHAP), death, or the last follow-up. Kaplan–Meier curves were which is a game theoretic approach to explain the out- generated and a log-rank test was performed to test the put of a tree-based machine learning model (Additional difference of PFS between patients who did and did not file  1: Supplementary Material S3), was applied  [27]. receive ART for each risk subgroup. SHAP measures the contribution of each feature of a model against the increase or decrease of the probability of a single output (i.e., the probability for tumor recur- Statistical analysis rence in our study) (Fig. 2). The Student’s t-test, Mann–Whitney U-test, and Chi- square test were performed to compare patient char- Analysis of survival outcomes and stratification acteristics between the responders and nonresponders for candidates for ART from training and test sets. The Kaplan–Meier method After constructing the best combined model for pre- was used to estimate survival rates, and the log-rank dicting tumor recurrence, the probability value of each test was performed to compare survival between the patient in the test set was analyzed to stratify candi- two groups. A P-value < 0.05 was considered statisti- dates for ART. Youden’s index was used for the optimal cally significant. All statistical analyses were performed cut-off selection threshold, and a cut-off probability using statistical software R (version 4.0.1; R Foundation for Statistical Computing, Vienna, Austria) and SPSS (version 25.0; SPSS Inc., Chicago, IL). Fig. 2 Comparison of median recurrence interval between patients with and without ART a in the entire cohort and b in high‑risk patients of the test set (probability > 0.3, according to the combined clinicopathological and radiomics model). Kaplan–Meier curves of PFS c comparing patients with and without ART in the entire cohort and d comparing patient subgroups according to the utility of ART in high‑risk patients. ART = adjuvant radiotherapy; PFS = progression‑free survival. Data are presented as the median with a 95% confidence interval for each group P ark et al. Radiation Oncology (2022) 17:147 Page 5 of 12 Treatment outcomes and prognostic factors Results With a median follow-up of 63.8  months (range, 6.6– Patient characteristics 190.7 months), there were 21 patients with tumor recur- Among 155 patients (mean age 56.9 ± 14.3, 102 females rence (12 [13.0%] in the training set, and 9 [14.3%] in the and 53 males), 148 had atypical meningiomas, 6 had test set). Fifteen events occurred within 24  months (7 choroid meningioma, and 1 had clear cell meningi- in the training set, and 8 in the test set, P = 0.292). The oma. In the training set, GTR was more frequently 5-year PFS rate was 79.1% for the entire cohort, and performed in patients without tumor recurrence com- 79.9% and 78.9% for the training and test set, respectively pared to those with tumor recurrence (87.5% vs. 58.3%, (P = 0.529). P = 0.011). No other clinicopathological features were In the training set, tumor recurrences occurred more different between patients with and without tumor frequently in patients who underwent STR than in those recurrence. In the test set, GTR was more frequently who underwent GTR (33.3% vs. 9.1%, P = 0.011). The performed in patients without tumor recurrence com- recurrence rate was higher in patients with a Ki-67 index pared to those with tumor recurrence (90.7% vs. 66.7%, value of 5% or more, although it was not statistically sig- P = 0.045). Patients with tumor recurrence were sig- nificant (19.1% vs. 6.7%, P = 0.076). In the test set, the nificantly older (P = 0.020), showed a male predomi- recurrence rate was higher in patients who underwent nance (P = 0.031), and had a higher Ki-67 labeling index STR (37.5% vs. 10.9%, P = 0.045) and in patients with (≥ 5%) (P = 0.018). ART was performed in 56 (60.9%) Ki-67 ≥ 5% (24.2% vs. 3.3%, P = 0.018). and 41 patients (65.1%) in the training and test sets, respectively. There were no significant differences Radiomics model construction of recurrences between the training and test sets with respect to the and comparison of diagnostic performance clinical and pathologic variables. The clinicopathologi - In the clinicopathological model, the AUC, accuracy, sen- cal characteristics of the training and test sets are sum- sitivity, and specificity were 0.67 (95% confidence interval marized in Table 1. [CI] 0.60–0.74), 66.9%, 48.2%, and 85.3% in the training Table 1 Patient characteristics in the training and test sets a a b Clinical variables Training set (n = 92) P‑ value Test set (n = 63) P‑ value P‑ value Without tumor With tumor Without tumor With tumor recurrence recurrence recurrence recurrence (n = 80) (n = 12) (n = 54) (n = 9) Age (years) 57.3 ± 14.6 54.4 ± 13.6 0.530 55.7 ± 14.1 67.7 ± 13.2 0.020 0.838 Female ratio 52 (65.0%) 9 (75.0%) 0.494 38 (70.4%) 3 (33.3%) 0.031 0.875 Extent of resection 0.011 0.045 0.535 GTR 70 (87.5%) 7 (58.3%) 49 (90.7%) 6 (66.7%) STR 10 (12.5%) 5 (41.7%) 5 (9.3%) 3 (33.3%) Ki‑67 labeling index 6.03 ± 4.81 6.96 ± 2.73 0.515 5.9 ± 4.3 8.2 ± 3.1 0.131 0.874 < 5% 32 (40.0%) 3 (25.0%) 0.076 29 (53.7%) 1 (11.1%) 0.018 ≥ 5% 38 (47.5%) 9 (75.0%) 25 (46.3%) 8 (88.9%) ART 0.659 0.517 0.595 Performed 48 (60.0%) 8 (66.7%) 36 (66.7%) 5 (55.6%) Not performed 32 (40.0%) 4 (33.3%) 18 (33.3%) 4 (44.4%) ART modality 0.851 0.061 0.813 3D‑ CRT 5 (10.4%) 2 (25.0%) 2 (5.6%) 2 (40.0%) IMRT 43 (89.6%) 6 (75.0%) 34 (94.4%) 3 (60.0%) ART dose (Gy) 57.3 ± 4.0 57.2 ± 6.7 0.956 58.3 ± 3.9 58.7 ± 2.6 0.853 0.204 Data are expressed as the mean with standard deviation in parentheses, median with interquartile range in parentheses, or number with percentage in parentheses Calculated from Student’s-t test for continuous variables and Chi-square test for categorical variables to compare the patient characteristics between the responder and non-responders from each training and test set Calculated from Student’s-t test for continuous variables and Chi-square test for categorical variables for the comparison of training and test sets Data obtained from patients who underwent adjuvant radiotherapy following surgery GTR: gross total resection; STR: subtotal resection; ART: adjuvant radiotherapy; RT: radiotherapy; 3D-CRT: three-dimensional conformal radiotherapy; IMRT: intensity modulated radiotherapy Park et al. Radiation Oncology (2022) 17:147 Page 6 of 12 set, respectively, and 0.61 (95% CI 0.44–0.78), 81.3%, 90th percentile from T1C, and Ki-67 index are the 33.3%, and 89.1%, respectively, in the test set. three most important risk factors. In the combined clinicopathological and radiomics model, a total of seven features were selected: two clini- cal features (extent of resection [GTR or STR] and Ki-67 Selection of candidates for ART using the developed index [≥ 5% or < 5%]) and five radiomic features, which combined model th were all first-order features from T1C (10 percentile, In the test set, there were 49 patients in the high-risk th 90 percentile, entropy, mean absolute deviation, and group and 14 patients in the low-risk group accord- minimum) (Additional file  1: Table  S3). The AUC, accu - ing to the combined clinicopathological and radiomics racy, sensitivity, and specificity were 0.78 (95% CI 0.70– model. The characteristics of patients in the high- and 0.85), 75.0%, 76.8%, and 73.1% in the training set, and low-risk groups are summarized in Additional file  1: 0.77 (95% CI 0.60–0.94), 70.3%, 66.7%, and 70.9%, respec- Table  S4. Among all patients, those who received ART tively, in the test set. exhibited a significantly longer PFS (5-year PFS, 87.8% In the training set, the combined model achieved supe- vs. 65.4%, P = 0.035) and delayed recurrence (median rior performance compared with the clinicopathological time to recurrence 45.3 vs. 18.4 months) (Fig.  2a, c). In model (AUC: 0.78 vs. 0.67, P = 0.042) (Fig. 3, Table 2). In the high-risk group, the PFS was significantly longer the test set, the combined model trended toward better in patients who received ART (5-year PFS, 92.3% vs. performance in the test set than in the clinicopathologi- 56.8%, P = 0.024) (Fig .  2b), with significantly delayed cal model (AUC: 0.77 vs. 0.61, P = 0.192) without sta- recurrence (median 53.5 vs. 17.0  months). In the low- tistical significance. The diagnostic performances of the risk group, there was no significant difference in the two models in the training and test set are provided in PFS (P = 0.264) or recurrence interval with respect to Table 2. ART. Only one patient experienced tumor recurrence (39.9 months after surgery). Model interpretability with SHAP When the cohort of patients was divided into four The SHAP values for each selected feature in the com - groups according to ART status and recurrence prob- bined clinicopathological and radiomics model were ability, the PFS was observed to significantly improve calculated. The variance importance plot, summary after ART in patients with a high recurrence probability plot, decision plot, and force plot of the test set are (median 8.7  year, 5-year 92.3%) (Fig.  2d). The signifi - shown in Fig.  4. For each prediction, a positive SHAP cant improvement in PFS was observed regardless of value indicates an increase in the risk of tumor recur- the extent of resection (GTR: 5-year 91.3% vs. 63.5%, rence while a negative SHAP value indicates reduced P = 0.054; STR: 5-year 100.0% vs. 0.0%, P = 0.012). risk. As observed in the plots, the extent of resection, Fig. 3 Receiver operating characteristic curves of the radiomics model in the a training and b test sets P ark et al. Radiation Oncology (2022) 17:147 Page 7 of 12 ‑ ‑ ‑ Table 2 Performances of machine learning models for prediction of tumor recurrence in the training and test set Models Training set Test set AUC (95% CI) Accuracy (%) Sensitivity (%) Specificity (%) P value AUC (95% CI) Accuracy (%) Sensitivity (%) Specificity (%) P value Clinicopatho 0.67 (0.59–0.74) 66.9 48.2 85.3 Reference 0.61 (0.44–0.78) 81.3 33.3 89.1 Reference logical model Clinicopatho 0.78 (0.70–0.85) 75.0 76.8 31.0 0.042 0.77 (0.60–0.94) 70.3 66.7 70.9 0.192 logical + radi omics model AUC: area under the curve; CI: confidence interval; NRI: net reclassification index Park et al. Radiation Oncology (2022) 17:147 Page 8 of 12 (See figure on next page.) Fig. 4 Model interpretability of a combined clinicopathological and radiomics model for the prediction of tumor recurrence with SHAP in the training set. a Variance importance plot that lists the most significant variables in descending order. b Summary plot of feature impact on the decision of the model and interaction between the features in the model. A positive SHAP value indicates an increase in the probability of tumor recurrence. c Decision plot showing how the model predicts tumor recurrence. Starting at the bottom of the plot, the prediction line shows how the SHAP values accumulate from the base value to arrive at the model’s final score at the top of the plot and how each feature contributes to the overall prediction of tumor recurrence. d Force plot of a representative case of a patient with tumor recurrence. Red arrows represent feature effects that drive the prediction value higher, and blue arrows are those effects that drive the prediction value lower. Each arrow’s size represents the magnitude of the corresponding feature’s effect. Note that the extent of resection, 90th percentile from T1C, and Ki‑67 labeling index largely push the model prediction score higher than the base value intensified adjuvant therapies, and patients at low-risk Discussion for tumor recurrence can be spared from experienc- In this study, radiomic features derived from preopera- ing the possible neurotoxicities of ART  [38]. Several tive MRI were applied to predict tumor recurrence in previous studies investigated clinical prognostic fac- grade 2 meningiomas. The model incorporating radi - tors for grade 2 meningiomas, which included age, sex, omic features with clinicopathologic features exhibited extent of resection, tumor invasiveness (i.e., brain or significantly a higher prediction accuracy for tumor bone invasion), and higher MIB-1 labeling index  [39, recurrence. Several risk factors for tumor recurrence 40]. Similar to our study, the clinical significance of were identified using SHAP, and the contribution of the Ki-67 labeling index on local control and survival each feature to the probability of tumor recurrence in high-grade meningiomas has been widely reported. was determined. Further, radiomics enabled a subset A higher Ki-67 labeling index seems to be related to of patients at a high risk for tumor recurrence to be invasiveness, which is closely correlated with an incom- identified, and we observed that the high-risk patients plete resection rate. Although there is some discrep- clearly benefitted from ART. Therefore, radiomics can ancy in the exact cutoff point, several studies have serve as a potential imaging biomarker, as well as a use- suggested the need for ART in patients with a higher ful tool for selecting adequate candidates for ART in Ki-67  [41, 42]. Pretreatment tumor size or volume was patients with meningiomas. also reported as a significant predictor of prognosis [35, Although the majority of meningiomas are benign 43–45]. However, in our study that analyzed various and slow growing, WHO grade 2 meningiomas are radiomic features including tumor size, it was not a sig- considered to have a high recurrence rate of 55% and nificant prognostic factor compared to other features. low survival rates [28–30]. ART after the STR of grade Several studies attempted to identify the prognostic 2 meningiomas is widely practiced; however, the value of radiomics in patients with meningiomas [16, 17, impact of ART on grade 2 meningiomas after a GTR 46]. However, these previous studies combined grade 1–3 remains contentious. Several researchers demonstrated meningiomas and do not provide specific information on improved local control and/or survival  [8–10, 30–33] the utility of the radiomics model for grade 2 meningi- with ART after GTR, but a subset exhibited contrary omas. Moreover, neither the combined clinicopathologi- results after ART  [9, 34–36]. According to two multi- cal model nor radiomics model showed any improved institutional phase II studies (NRG Oncology RTOG performance in predicting prognosis compared with the 0539  [32] and EORTC 22,042–26,042  [37]), which clinicopathological model alone, limiting the real-world enrolled grade 2 meningioma patients who underwent application of the radiomics model  [17, 46]. Decision- GTR and ART, there is a potential survival benefit with making for the treatment of grade 2 meningioma patients a 3-year PFS of 89%–94%. However, those studies did depends on multilevel prognostic information and clin- not directly compare ART with upfront observation, icopathological information, such as the extent of the and risk factors other than surgical resection were tumor resection or Ki-67 index; therefore, we focused not addressed. Therefore, a uniform treatment para - on the role of radiomics in predicting survival given the digm for resected grade 2 meningiomas has yet to be multilevel prognostic information. Our results show that established. radiomics significantly increases the model performance Because there is a lack of standardized treatment when added to a clinicopathological model, thus promot- for resected grade 2 meningiomas, performing nonin- ing the integration of radiomics in clinical practice. vasive risk stratification prior to adjuvant therapies is Radiomics models have limited explainability highly desirable. Even after surgical resection, patients because they rely on complex machine learning algo- at high-risk for tumor recurrence can be considered for rithms, resulting in low clinical utility  [47]. SHAP can P ark et al. Radiation Oncology (2022) 17:147 Page 9 of 12 Fig. 4 (See legend on previous page.) Park et al. Radiation Oncology (2022) 17:147 Page 10 of 12 uncover complex underlying patterns  [27, 48] and has Conclusions recently been utilized to interpret radiomics mod- Multiparametric MRI radiomics has an added prognostic els  [49, 50]. This analysis identified two clinical fea- value for the prediction of tumor recurrence when inte- tures—the extent of resection and the Ki-67 labeling grated with clinicopathologic profiles in patients with index—and five radiomic first-order features from grade 2 meningiomas. With radiomics, we could identify T1C that contributed the most towards the predic- a subset of patients at a high risk for tumor recurrence tion of tumor recurrence. The extent of resection and who might benefit from intensified treatment. Therefore, the Ki-67 index are well-known prognostic factors for radiomics may be used as a potential imaging biomarker tumor recurrence  [35, 51], which was also verified by in patients with grade 2 meningiomas. th th our study. The 90 and 10 percentile values from T1C, which are the two most important radiomic fea- Abbreviations tures, denote the intensity of contrast enhancement. ART : Adjuvant radiotherapy; GTR : Gross total resection; LASSO: Least absolute Heterogeneous contrast enhancement or the presence shrinkage and selection operator; PFS: Progression‑free survival; RANO: Response assessment in neuro‑ oncology; ROI: Region of interest; ROSE: of contrast enhancement of meningiomas  [17] has Random over‑sampling examples; SHAP: Shapley Additive Explanations; STR: shown significant association with high-grade menin- Subtotal resection; T1C: Post‑ contrast T1‑ weighted image; T2: T2‑ weighted giomas or tumor invasiveness  [52, 53]; therefore, the image; WHO: World Health Organization. meningioma grade may be indicated by the percentile- related first-order features, which reflect the distribu- Supplementary Information tion of the contrast-enhancement degree. The online version contains supplementary material available at https:// doi. org/ 10. 1186/ s13014‑ 022‑ 02090‑7. Our study has several limitations. First, our study was based on a single-center, retrospectively col- Additional file 1. Supplementary figures and tables. lected dataset. Further studies with a larger num- ber of patients and external validations are required. Acknowledgements Second, the histopathologic risk factors, such as the None MIB-1 labeling index or mitoses index, were not evalu- Author contributions ated in our study because relevant information for the SHC and CJP analyzed, interpreted the patient data and wrote the manuscript. majority of patients were unavailable during the long SHC, CJP, JE, and HKB are contributors for analysis. YWP and HIY are major study-inclusion period. To prove the predictive val- contributors for conceptualization and interpretation. All authors read and approved the final manuscript. ues of radiomics, it is highly desirable to incorporate detailed histopathologic features into the models in Funding future studies. Third, the probability scores derived This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of from the developed model for tumor recurrence were Education (Republic of Korea; grant number: NRF‑2020R1I1A1A0107164811 to only calculated in the test set, not in the training set, Y. W. Park) and the National Research Foundation of Korea (NRF) Grant funded because the training set was oversampled during the by the Korea government (MSIT ) (NRF‑2021R1A2C1010900 to S.H. Choi). This study was also supported by a faculty research grant of Yonsei University Col‑ development of the model. Therefore, the benefit of lege of Medicine for (6‑2020‑0115). ART was only evaluated in the test set, which inevita- bly decreased the sample size. However, based on our Availability of data and materials Research data are stored in an institutional repository and will be shared upon study results, we at least found a benefit of ART in a request to the corresponding author. subgroup of the test set, which was identified through the combined model. Future studies with larger sample Declarations sizes are required to identify the patients with grade 2 meningiomas who might benefit from ART. Ethics approval and consent to participate This study was conducted according to the guidelines of the Declaration of Our study is the first to show the possibility of radi- Helsinki, and approved by the institutional review board of the Yonsei Univer‑ omics playing an important role in selecting candidate sity Health System (9‑2021‑0047). patients with meningiomas for ART. Over a similar Consent for publication follow-up period (median 54 months in ART group vs. Not applicable. 61 months in non-ART group, P = 0.584), we observed that ART prolongs PFS and significantly delays recur- Competing interests None of the authors have any conflicts of interest to disclose. rence in high-risk patients. Our combined predictive model could be applied effectively and conveniently in Author details both GTR and STR patients. Department of Radiology, Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yongin Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea. Depar tment of Radiation Oncology, Yongin Severance Hospital, Yonsei University College P ark et al. Radiation Oncology (2022) 17:147 Page 11 of 12 of Medicine, Seoul, Republic of Korea. Department of Computer Science, Yon‑ 19. Huang RY, Bi WL, Weller M, Kaley T, Blakeley J, Dunn I, et al. Proposed sei University, Seoul, Republic of Korea. Department of Radiation Oncology, response assessment and endpoints for meningioma clinical trials: Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University report from the Response Assessment in Neuro‑ Oncology Working College of Medicine, 50‑1 Yonsei‑ro, Seodaemun‑gu, Seoul 03722, Republic Group. Neuro Oncol. 2019;21(1):26–36. of Korea. Department of Neurosurgery, Yonsei University College of Medicine, 20. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, 50‑1 Yonsei‑ro, Seodaemun‑gu, Seoul 03722, Republic of Korea. Depar tment et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. of Pathology, Yonsei University College of Medicine, Seoul, Republic of Korea. 2010;29(6):1310–20. 21. Cha J, Kim ST, Kim HJ, Kim BJ, Kim YK, Lee JY, et al. Differentiation of Received: 14 March 2022 Accepted: 23 June 2022 tumor progression from pseudoprogression in patients with posttreat‑ ment glioblastoma using multiparametric histogram analysis. AJNR Am J Neuroradiol. 2014;35(7):1309–17. 22. Maes F, Collignon A, Vandermeulen D, Marchal G, Suetens P. Multimo‑ dality image registration by maximization of mutual information. IEEE Trans Med Imaging. 1997;16(2):187–98. References 23. Carré A, Klausner G, Edjlali M, Lerousseau M, Briend‑Diop J, Sun R, et al. 1. Ostrom QT, Gittleman H, Liao P, Rouse C, Chen Y, Dowling J, et al. CBTRUS Standardization of brain MR images across machines and protocols: statistical report: primary brain and central nervous system tumors diag‑ bridging the gap for MRI‑based radiomics. Sci Rep. 2020;10(1):12340. nosed in the United States in 2007–2011. Neuro Oncol. 2014;16(Suppl 24. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan 4):iv1‑63. V, et al. Computational radiomics system to decode the radiographic 2. Ostrom QT, Gittleman H, Truitt G, Boscia A, Kruchko C, Barnholtz‑Sloan JS. phenotype. Cancer Res. 2017;77(21):e104–7. CBTRUS statistical report: primary brain and other central nervous system 25. Zwanenburg A, Leger S, Vallières M, Löck S. Image biomarker standardi‑ tumors diagnosed in the United States in 2011–2015. Neuro Oncol. sation initiative. arXiv preprint arXiv: 1612. 07003. 2016. 2018;20(suppl_4):iv1–86. 26. DeLong ER, DeLong DM, Clarke‑Pearson DL. Comparing the areas 3. Goldbrunner R, Stavrinou P, Jenkinson MD, Sahm F, Mawrin C, Weber DC, under two or more correlated receiver operating characteristic curves: et al. EANO guideline on the diagnosis and management of meningi‑ a nonparametric approach. Biometrics. 1988;44(3):837–45. omas. Neuro Oncol. 2021;23(11):1821–34. 27. Lundberg SM, Lee S‑I. A unified approach to interpreting model pre ‑ 4. Stark AM, Buhl R, Mehdorn HM. Early gross recurrence of atypical menin‑ dictions. In. Proceedings of the 31st international conference on neural gioma. J Neurooncol. 2005;75(2):223–4. information processing systems; 2017. pp. 4768–77. 5. Jenkinson MD, Waqar M, Farah JO, Farrell M, Barbagallo GM, McManus R, 28. Modha A, Gutin PH. Diagnosis and treatment of atypical and anaplastic et al. Early adjuvant radiotherapy in the treatment of atypical meningi‑ meningiomas: a review. Neurosurgery. 2005;57(3):538–50. oma. J Clin Neurosci. 2016;28:87–92. 29. Rogers L, Barani I, Chamberlain M, Kaley TJ, McDermott M, Raizer J, 6. Lee KD, DePowell JJ, Air EL, Dwivedi AK, Kendler A, McPherson CM. Atypi‑ et al. Meningiomas: knowledge base, treatment outcomes, and uncer‑ cal meningiomas: is postoperative radiotherapy indicated? Neurosurg tainties. A RANO review. J Neurosurg. 2015;122(1):4–23. Focus. 2013;35(6):E15. 30. Aizer AA, Bi WL, Kandola MS, Lee EQ, Nayak L, Rinne ML, et al. Extent of 7. Goldbrunner R, Stavrinou P, Jenkinson MD, Sahm F, Mawrin C, Weber DC, resection and overall survival for patients with atypical and malignant et al. EANO guideline on the diagnosis and management of meningi‑ meningioma. Cancer. 2015;121(24):4376–81. omas. Neuro Oncol. 2021. https:// doi. org/ 10. 1093/ neuonc/ noab1 50. 31. Li H, Zhang YS, Zhang GB, Zhang GJ, Wang B, Li D, et al. Treatment 8. Park HJ, Kang HC, Kim IH, Park SH, Kim DG, Park CK, et al. The role protocol, long‑term follow‑up, and predictors of mortality in 302 cases of adjuvant radiotherapy in atypical meningioma. J Neurooncol. of atypical meningioma. World Neurosurg. 2019;122:e1275–84. 2013;115(2):241–7. 32. Rogers L, Zhang P, Vogelbaum MA, Perry A, Ashby LS, Modi JM, et al. 9. Bagshaw HP, Burt LM, Jensen RL, Suneja G, Palmer CA, Couldwell WT, Intermediate‑risk meningioma: initial outcomes from NRG Oncology et al. Adjuvant radiotherapy for atypical meningiomas. J Neurosurg. RTOG 0539. J Neurosurg. 2018;129(1):35–47. 2017;126(6):1822–8. 33. Rydzewski NR, Lesniak MS, Chandler JP, Kalapurakal JA, Pollom E, 10. Choi Y, Lim DH, Jo K, Nam DH, Seol HJ, Lee JI. Efficacy of postop ‑ Tate MC, et al. Gross total resection and adjuvant radiotherapy most erative radiotherapy for high grade meningiomas. J Neurooncol. significant predictors of improved survival in patients with atypical 2014;119(2):405–12. meningioma. Cancer. 2018;124(4):734–42. 11. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, 34. Zhu H, Bi WL, Aizer A, Hua L, Tian M, Den J, et al. Efficacy of adjuvant they are data. Radiology. 2016;278(2):563–77. radiotherapy for atypical and anaplastic meningioma. Cancer Med. 12. Park YW, Han K, Ahn SS, Choi YS, Chang JH, Kim SH, et al. Whole‑tumor 2019;8(1):13–20. histogram and texture analyses of DTI for evaluation of IDH1‑mutation 35. Zhi M, Girvigian MR, Miller MJ, Chen JC, Schumacher AJ, Rahimian and 1p/19q‑ codeletion status in world health organization grade II J, et al. Long‑term outcomes of newly diagnosed resected atypical gliomas. AJNR Am J Neuroradiol. 2018;39(4):693–8. meningiomas and the role of adjuvant radiotherapy. World Neurosurg. 13. Won SY, Park YW, Ahn SS, Moon JH, Kim EH, Kang SG, et al. Quality 2019;122:e1153–61. assessment of meningioma radiomics studies: Bridging the gap between 36. Graffeo CS, Leeper HE, Perry A, Uhm JH, Lachance DJ, Brown PD, exploratory research and clinical applications. Eur J Radiol. 2021;138: et al. Revisiting adjuvant radiotherapy after gross total resection of world health organization grade II meningioma. World Neurosurg. 14. Park YW, Oh J, You SC, Han K, Ahn SS, Choi YS, et al. Radiomics and 2017;103:655–63. machine learning may accurately predict the grade and histological sub‑ 37. Weber DC, Ares C, Villa S, Peerdeman SM, Renard L, Baumert BG, et al. type in meningiomas using conventional and diffusion tensor imaging. Adjuvant postoperative high‑ dose radiotherapy for atypical and malig‑ Eur Radiol. 2019;29(8):4068–76. nant meningioma: A phase‑II parallel non‑randomized and observa‑ 15. Gu H, Zhang X, di Russo P, Zhao X, Xu T. The current state of radiomics for tion study (EORTC 22042–26042). Radiother Oncol. 2018;128(2):260–5. meningiomas: promises and challenges. Front Oncol. 2020;10: 567736. 38. Wu A, Jin MC, Meola A, Wong H‑n, Chang SD. Efficacy and toxicity of 16. Zhang Y, Chen JH, Chen TY, Lim SW, Wu TC, Kuo Y T, et al. Radiomics particle radiotherapy in WHO grade II and grade III meningiomas: a approach for prediction of recurrence in skull base meningiomas. Neuro‑ systematic review. Neurosurgical Focus FOC. 2019;46(6):E12. radiology. 2019;61(12):1355–64. 39. Wang F, Xu D, Liu Y, Lin Y, Wei Q, Gao Q, et al. Risk factors associated 17. Morin O, Chen WC, Nassiri F, Susko M, Magill ST, Vasudevan HN, et al. with postoperative recurrence in atypical intracranial meningioma: Integrated models incorporating radiologic and radiomic features predict analysis of 263 cases at a single neurosurgical centre. Acta Neurochir. meningioma grade, local failure, and overall survival. Neurooncol Adv. 2019;161(12):2563–70. 2019;1(1):vdz011. 40. Ros‑Sanjuan A, Iglesias‑Moroño S, Carrasco ‑Brenes A, Bautista‑ Ojeda D, 18. Friedman J, Hastie T, Tibshirani R. Regularization paths for generalized Arraez‑Sanchez MA. Atypical meningiomas: histologic and clinical fac‑ linear models via coordinate descent. J Stat Softw. 2010;33(1):1–22. tors associated with recurrence. World Neurosurg. 2019;125:e248–56. Park et al. Radiation Oncology (2022) 17:147 Page 12 of 12 41. Choi Y, Lim DH, Yu JI, Jo K, Nam DH, Seol HJ, et al. Prognostic value of Ki‑67 labeling index and postoperative radiotherapy in WHO grade II meningioma. Am J Clin Oncol. 2018;41(1):18–23. 42. Bruna J, Brell M, Ferrer I, Gimenez‑Bonafe P, Tortosa A. Ki‑67 proliferative index predicts clinical outcome in patients with atypical or anaplastic meningioma. Neuropathology. 2007;27(2):114–20. 43. Detti B, Scoccianti S, Di Cataldo V, Monteleone E, Cipressi S, Bordi L, et al. Atypical and malignant meningioma: outcome and prognostic factors in 68 irradiated patients. J Neurooncol. 2013;115(3):421–7. 44. Milker‑Zabel S, Zabel A, Schulz‑Ertner D, Schlegel W, Wannenmacher M, Debus J. Fractionated stereotactic radiotherapy in patients with benign or atypical intracranial meningioma: long‑term experience and prognos‑ tic factors. Int J Radiat Oncol Biol Phys. 2005;61(3):809–16. 45. Zhang GJ, Zhang YS, Zhang GB, Yan XJ, Li CB, Zhang LW, et al. Prognostic factors, survival, and treatment for intracranial world health organiza‑ tion grade ii chordoid meningiomas and clear‑ cell meningiomas. World Neurosurg. 2018;117:e57–66. 46. Gennatas ED, Wu A, Braunstein SE, Morin O, Chen WC, Magill ST, et al. Preoperative and postoperative prediction of long‑term meningioma outcomes. PLoS ONE. 2018;13(9): e0204161. 47. Du R, Lee VH, Yuan H, Lam K‑ O, Pang HH, Chen Y, et al. Radiomics model to predict early progression of nonmetastatic nasopharyngeal carcinoma after intensity modulation radiation therapy: a multicenter study. Radiol Artif Intell. 2019;1(4):e180075. 48. Arrow K, Barankin E, Blackwell D. Admissible points of convex sets. Con‑ trib Theory Games. 1953;2:87–91. 49. Bang M, Eom J, An C, Kim S, Park YW, Ahn SS, et al. An interpretable mul‑ tiparametric radiomics model for the diagnosis of schizophrenia using magnetic resonance imaging of the corpus callosum. Transl Psychiatry. 2021;11(1):462. 50. Giraud P, Giraud P, Nicolas E, Boisselier P, Alfonsi M, Rives M, et al. Inter‑ pretable machine learning model for locoregional relapse prediction in oropharyngeal cancers. Cancers. 2021;13(1):57. 51. Mirian C, Skyrman S, Bartek J Jr, Jensen LR, Kihlström L, Förander P, et al. The Ki‑67 proliferation index as a marker of time to recurrence in intracra‑ nial meningioma. Neurosurgery. 2020;87(6):1289–98. 52. Adeli A, Hess K, Mawrin C, Streckert EMS, Stummer W, Paulus W, et al. Pre‑ diction of brain invasion in patients with meningiomas using preopera‑ tive magnetic resonance imaging. Oncotarget. 2018;9(89):35974–82. 53. Liu Y, Chotai S, Chen M, Jin S, Qi S‑t, Pan J. Preoperative radiologic clas‑ sification of convexity meningioma to predict the survival and aggressive meningioma behavior. PLoS ONE. 2015;10(3):e0118908‑ e. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in pub‑ lished maps and institutional affiliations. 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

Radiation OncologySpringer Journals

Published: Aug 22, 2022

Keywords: Magnetic resonance imaging; Meningioma; Radiomics; Radiotherapy; Prognosis

There are no references for this article.