A Radiomics Model for Predicting the Response to Bevacizumab in Brain Necrosis after Radiotherapy

Radiotherapy is a key component in the management of head and neck tumors and primary or metastatic cancer of the central nervous system. Various radiation regimens have been developed with the aim of reducing unnecessary radiation to adjacent normal brain tissue. However, these regimens still cause different degrees of damage and lead to brain necrosis in 3% to 24% of patients (1).

Currently, corticosteroid therapy is used for brain necrosis, but with various side effects such as cushingoid symptoms, arthritic sequelae, osteoporosis, peptic ulcer, immunosuppression, and psychiatric disturbances (2), and only 35% of patients may get benefit (3). Other adjuvant therapy, including antiplatelet agents, anticoagulants, and hyperbaric oxygen, were reported to be helpful (1, 2).

Brain necrosis following radiotherapy has been associated with increased permeability of the blood–brain barrier (BBB), inflammation, and elevated VEGF levels. The increased VEGF has been proved to play a critical role in the development of brain edema and necrosis (1, 4–7). Bevacizumab, a humanized mAb against VEGF, has been shown to decrease the volumes of necrosis by 52% to 64%, and the effective response rate is approximately 66% (8–13). Moreover, a randomized controlled trial of bevacizumab therapy for brain necrosis we conducted demonstrated that bevacizumab offers significant symptomatic relief and radiographic response compared with corticosteroids (11). Bevacizumab has been used increasingly for brain necrosis and is promising to be the primary treatment (14). However, some patients failed to derive adequate benefit from bevacizumab or even worsened (15), and some may suffer from severe adverse events (8, 11). Therefore, there is a great need to differentiate patients who are likely to have a good response to bevacizumab from those who are not, so as to provide more personalized targeted therapy and free patients who might have a poor treatment response from risk of those severe side effects. Previously, we developed a random forests model to predict response to bevacizumab in patients with brain necrosis after radiotherapy. This model indicated that patients in different stages of the disease show significant different treatment responses (16). Because characteristics of brain lesions in different stages can be reflected in MRI, we presumed that pretreatment MRI can be incorporated to a model to predict the response to bevacizumab, which may further improve the predictive accuracy.

Radiomics involves high-throughput extraction of large amounts of image-based features from standard medical radiographic images and determination of the relationships between such features and potential pathophysiology (17, 18). Radiomics is now gaining importance in cancer research because of its potential to aid in diagnosis, evaluation of prognosis, and prediction of response to treatment (17, 19) in a noninvasive manner. Studies have shown that pretreatment MRI features may predict the treatment response of certain diseases, such as rectal cancer, cervical cancer, and glioblastoma (20–24). Therefore, this motivates the investigation of the possibility of predicting response to treatment for brain necrosis using MRI features.

Hence, the aim of this study was to develop and validate a radiomics model that integrates a radiomics signature originating from pretreatment MRI and clinical variables for individualized prediction of the response to bevacizumab in patients with brain necrosis. To eliminate the confounding factors of tumor along with necrosis, we only included patients with nasopharyngeal carcinoma (NPC) who developed brain necrosis after radiotherapy for model construction. And the model was further validated in patients with NPC or another cancer type to assess its generalizability.

The Institutional Review Boards approved this retrospective study. This study was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from every patient. In total, 118 consecutive patients who were diagnosed with brain necrosis after radiotherapy for NPC and were treated with bevacizumab between July 2012 and March 2019 in Sun Yat-sen Memorial Hospital (Guangzhou, China) were enrolled in the study. Among them, 77 patients with 101 brain lesions treated between July 2012 and December 2017 were assigned to the training set, whereas 41 patients with 51 brain lesions treated between January 2018 and March 2019 were assigned to the internal validation set. In addition, 31 patients with brain necrosis after radiotherapy (42 lesions) for brain metastases from lung cancer, treated between January 2017 and January 2019 in Tianjin Medical University Cancer Institute and Hospital (Tianjin, China), were used for external validation. The inclusion and exclusion criteria are described in Supplementary Material A1. The patient recruitment pathway and the study flowchart are shown in Supplementary Fig. S1 and Fig. 1A, respectively. The diagnosis of radiation necrosis is mainly based on opinions from both neurologists and radiologists as described in previous studies (8, 9). All patients received bevacizumab 5 mg/kg intravenously once every 2 weeks for up to 4 courses. The details concerning the diagnosis of radiation necrosis, treatment with bevacizumab, and assessments before treatment with bevacizumab are outlined in Supplementary Material A2–A4, respectively.

The demographic and clinical data were collected before treatment with bevacizumab (Table 1). Diagram depicting the interval between radiotherapy and diagnosis of brain necrosis (IRB), and the interval between diagnosis of brain necrosis and treatment with bevacizumab (IBT) is presented in Fig. 1B. The volumes of the brain lesions were detected by fluid-attenuated inversion recovery (FLAIR) imaging within 3 days before the initiation of bevacizumab treatment and 2 weeks after completion of treatment. Note that the “lesion” represents both necrosis and surrounding edema lesion. The effective treatment response was defined as: (i) an improvement in neurologic manifestations [a reduction ≥1 in the grade of the Late Effects of Normal Tissue/Subjective, Objective, Management, Analytic (LENT/SOMA) scale was defined as an improvement; refs. 8, 11, 25], or (ii) stable neurologic manifestations and a reduction in lesion volume on FLAIR images by ≥25% (8). Of note, a portion of patients have bilateral lesions. Two lesions from the same patient might show different treatment outcomes; therefore, each individual lesion was regarded as a subject to be measured in our study.

The radiomics workflow is shown in Fig. 1C. All patients underwent MRI before and after treatment of bevacizumab. On FLAIR images, the edema surrounding the necrotic core shows increased signal intensity (26), which allows the brain lesion margins to be delineated more accurately. Therefore, the pretreatment coronal FLAIR Digital Imaging and Communications in Medicine (DICOM) images were used for analysis. Image segmentation and radiomics features extraction were performed by using the 3D Slicer software (version 4.9.0; ref. 27). In total, 1,301 radiomics features were extracted from each lesion on FLAIR images. Then, values of these features were normalized in a linear way in the range of 0 to 1 for further analysis. One of the radiomics features, that is, that named “volume”, which denotes the volume of a brain lesion, was also used for assessment of the outcomes of bevacizumab treatment. Details about the radiomics procedure and the radiomics features are presented in Supplementary Material A5 and Supplementary Table S1.

We applied the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm in the training set, which is appropriate for regression of high-dimensional data, to select treatment response–related features with nonzero coefficients (28). A radiomics signature was constructed to estimate the probability of treatment response for each brain lesion using the radiomics score, which was calculated as a linear combination of the selected features weighted by their respective coefficients (29). The discrimination of the radiomics signature was evaluated by the area under the ROC curve (AUC) in the training set and then validated in the internal and external validation sets. Mann–Whitney U tests were performed to assess the potential association of the radiomics score with the treatment response. Stratified analyses were also performed within different subgroups of all patients.

According to the univariable logistic regression analyses, those variables with P < 0.1 were subjected to subsequent multivariable analysis using likelihood ratio test with backward step-down selection in the training set. A variance inflation factor (VIF) was used to estimate the collinearity diagnostics of the multivariable logistic regression. A radiomics model was then developed on the basis of the results of the multivariable logistic analysis.

In the training set, a calibration curve was plotted to measure the calibration of the model, accompanied by the Hosmer–Lemeshow test to assess the goodness-of-fit of the model (30). The AUC was also calculated to quantify the discrimination performance of the model.

The performance of the radiomics model was then tested in two validation sets. Using the formula constructed in the training set, a radiomics score was calculated for each brain lesion in the validation sets. The calibration and Hosmer–Lemeshow tests were then performed, and the AUCs were calculated.

To evaluate whether addition of the radiomics signature could improve the performance of clinical predictors, ROC analyses were used in all patients to compare the discriminatory efficacy of the radiomics model with that of each selected clinical predictor and the combined clinical predictors. Decision curve analysis (DCA) was performed to determine the clinical utility of the radiomics model by quantifying the net benefits at different threshold probabilities (31).

All statistical tests were performed using R statistical software (version 3.4.2; R Foundation for Statistical Computing). The “glmnet” package was used to perform the LASSO logistic regression model analysis. The ROC curves were plotted using the “pROC” package. The VIF values were measured using the “car” package. Calibration plots were performed using the “rms” package and the Hosmer–Lemeshow test was performed using the “vcdExtra” package. DCA was plotted using the “dca.R.” Detailed information about the LASSO algorithm and DCA are presented in Supplementary Material A6 and A7. A two-sided P < 0.05 was considered statistically significant.

The baseline clinical characteristics of the patients are summarized in Table 1. A total of 30.2% (45/149) of all enrolled patients have bilateral lesions. Overall, 68.6% (133/194) of the brain lesions displayed effective treatment response. The treatment outcomes for the two lesions in the same brain were inconsistent in 9 (20.0%) of these 45 patients. Representative images showing treatment outcomes of 3 patients with bilateral lesions are presented in Supplementary Fig. S2.

Satisfactory reproducibility of radiomics feature extraction was achieved in our study (Supplementary Material A5). Using a LASSO logistic regression model, 18 treatment outcome-related features with nonzero coefficients were selected in the training set (Supplementary Fig. S3). The resulting coefficients of the selected features used to calculate the radiomics score are presented in Supplementary Fig. S4.

There was a significant difference in the radiomics scores [median (interquartile range), Mann–Whitney U test] between the lesions that responded favorably to treatment and those that did not in the training set [1.184 (0.765–1.702) vs. 0.370 (−0.039 to 0.633), respectively, P < 0.001]; the same was true in the internal validation set [1.460 (1.009–1.765) vs. 0.734 (0.405–0.981), respectively, P < 0.001] and the external validation set [1.565 (1.129–2.186) vs. 0.485 (−0.167 to 1.339), respectively, P < 0.001; Fig. 2A]. A significant association between the radiomics score and treatment response was also observed in the stratified analyses (Supplementary Table S2).

The radiomics signature showed favorable discrimination, with an AUC of 0.873 [95% confidence interval (CI), 0.804–0.943] in the training set, which was validated in the internal and external validation sets with AUCs of 0.869 (95% CI, 0.755–0.982) and 0.814 (95% CI, 0.676–0.951; Fig. 2B), respectively. An optimal cut-off value of radiomics score was chosen as 0.642 according to the maximum Youden index in the training set. The waterfall plot shows the distribution of radiomics scores and treatment response for all lesions, with the dividing line drawn at the cut-off value (Fig. 2C).

According to the multivariable logistic regression analysis, IRB, IBT, and the radiomics signature were identified as independent predictors of response to treatment (Table 2). The VIF values ranged from 1.132 to 1.512, indicating that there was no collinearity in the collinearity diagnosis. A prediction model that incorporated IRB, IBT, and the radiomics signature was developed and a response score could be calculated for each lesion to reflect the probability of therapeutic response based on the logistic regression formula. The calculating formula was as follows: response score = 3.436 × radiomics score − 0.024 × IRB − 0.043 × IBT + 0.451. The predicted therapeutic response probability was calculated using 1/[1 + exp (−response score)].

In the training set, the calibration curve of the radiomics model estimating the probability of an effective treatment response demonstrated good agreement (Fig. 3A), and an AUC of 0.916 (95% CI, 0.857–0.976) indicated that the model had good discrimination (Fig. 3B). The Hosmer–Lemeshow test yielded a nonsignificant statistic (P = 0.484), suggesting no departure from the perfect fit.

Favorable calibration of the radiomics model was confirmed in the internal and external validation sets (Fig. 3A). The radiomics model also achieved good discrimination performance, with AUCs of 0.912 (95% CI, 0.808–1.000) and 0.827 (95% CI, 0.691–0.962; Fig. 3B) in the internal and external validation sets, respectively. The Hosmer–Lemeshow tests yielded nonsignificant statistics for the model (P = 0.184 and P = 0.244, respectively). The satisfactory performance of the model in the external validation set demonstrated that the model derived from patients with NPC also had favorable prediction efficacy in patients with brain metastases from lung cancer, suggesting the generalizability of the model.

The ROC curves comparing the discriminatory efficiency of each selected clinical predictor (i.e., IRB or IBT), the combined clinical predictors, and the radiomics model in all patients are displayed in Supplementary Fig. S5. The results showed that the model had better prediction efficacy than either IRB, IBT, or the combined clinical predictors. In all three datasets, the DCA suggested that using the radiomics model to detect a treatment response adds more net benefit than either the treat-all or treat-none scheme for a wide range of threshold probability (Supplementary Fig. S6).

A website has been made available to provide a free online calculator based on the presented radiomics model (http://www.radiatcomplication.com/#/bevacizumab_response), which facilitates the prediction of treatment response of bevacizumab in clinical practice (Fig. 4).

In the current study, we developed an MRI-based and noninvasive radiomics model incorporating the radiomics signature and clinical variables for individualized prediction of the response to bevacizumab in patients with brain necrosis after radiotherapy. This may provide an illustrative example of precision medicine and affect treatment strategies.

Radiomics is an emerging field in quantitative imaging, which may have important applications in personalized medicine. Previously, several prediction models have been developed to differentiate between brain necrosis after radiotherapy and local recurrent brain metastases by using a radiomics approach (32–34). In this study, we first attempted to select the key predictive features among the large arrays of radiomics features extracted from pretreatment MRI images of brain necrosis and then developed a radiomics signature for the prediction of the response to bevacizumab. Of note, our radiomics signature showed favorable discrimination (AUC 0.873) in the training set, which was further validated in the internal and external validation sets (AUC 0.869 and 0.814, respectively).

To improve the prediction efficacy, predictors beyond radiomics should also be incorporated with radiomics signature to achieve a holistic model (17). Therefore, in our study, the radiomics signature was combined with patient characteristics to further increase the power of the decision support model. As a result, the radiomics signature, IRB, and IBT, with corresponding odds ratios of 1.410, 0.977, and 0.958, were selected by multivariable logistic regression analysis. The ORs for IRB and IBT suggest that the earlier brain necrosis occurs after radiotherapy or the earlier bevacizumab is administrated after diagnosis of brain necrosis, the greater the probability of a favorable response to bevacizumab. We consider the above results may be associated with the pathophysiologic progress of brain necrosis. Although the mechanism of brain necrosis after radiotherapy remains uncertain, microvascular injury is generally believed to play a critical role in the development of brain necrosis. In the early stage, increase of VEGF and dysfunction of the BBB allows plasma proteins to leak into the cerebral parenchyma, leading to cerebral edema. After that, endothelial cell swelling as well as apoptosis and subsequent collapse of capillaries are followed by necrosis (5, 6, 35). Brain necrosis at the early stage after radiotherapy may be associated with the increase of VEGF and capillary permeability, and bevacizumab could inhibit VEGF to alleviate edema. While during the late stage, besides collapse or occlusion of capillaries, mechanisms such as activation of glia cells and neuroinflammation might also be involved in the pathophysiologic process, which are not prone to be reversed simply by anti-VEGF therapy alone. Therefore, regular follow-up examinations to detect disease at an early stage and administer timely treatment are important for favorable outcomes of subsequent therapy.

After the selection of candidate predictors using multivariable logistic regression analysis, an MRI-based radiomics model was built incorporating the radiomics signature, IRB, and IBT. This model displayed good calibration and discrimination in the training set (AUC 0.916) and also performed well in the internal and external validation sets (AUC 0.912 and 0.827, respectively). Furthermore, the DCA showed a higher overall net benefit of the radiomic model, thus highlighting its value as a better tool for assisting in clinical decision-making. Using the radiomics model, if a patient is predicted to have favorable response to bevacizumab, then administration of bevacizumab should be recommended. This is particularly important for those with bilateral lesions and have been predicted to have consistently good treatment outcomes. In addition, corticosteroids may be a favorable alternative in patients with bilateral lesions predicted to display consistently poor treatment outcomes. In patients with bilateral lesions predicted to show inconsistent treatment outcomes, care should be taken to consider the clinical manifestations, and locations and volumes of lesions to evaluate the advantages and disadvantages of each regimen, determining treatment with bevacizumab or other therapies.

This is the first attempt to develop a radiomics model for the prediction of the response to bevacizumab in patients with brain necrosis after radiotherapy. On the whole, this study has several strengths. First, by incorporating the radiomics signature and clinical features based on a larger cohort, our model has superior predictive accuracy as compared with the previous random forests model for evaluating the response to bevacizumab in patients with brain necrosis (16). The high-dimensional features used in our present study provide more detailed information about brain lesions and are more sensitive and comprehensive when evaluating the response to treatment. Moreover, three-dimensional volumes of interest of lesions were extracted from image slices for acquisition of radiomics features rather than two-dimensional regions of interest of the lesions, and demonstrated the heterogeneity of the entire lesion more accurately (36). Second, the research subjects were brain lesions rather than individual patients. Because 30.2% of patients had bilateral lesions and in 20.0% of these patients the treatment outcomes for two lesions within the same brain were different in this study, our model is more precise and may provide more accurate direction for therapy in patients with bilateral lesions. Third, only patients with NPC were included to develop the model in this study, which can minimize the confounding factors of tumor along with necrosis and contribute to a more accurate prediction model with potential to be applied to other tumors. Furthermore, the model also performs well in patients with brain necrosis after radiotherapy for brain metastases from lung cancer. Therefore, our model may serve as a reliable and non-invasive tool for the prediction of the response to bevacizumab in patients with brain necrosis.

Despite the above strengths, some limitations were unavoidable in this study. First, possibly due to the heterogeneity of the patient population and the difference in MRI scanners with different imaging parameters, the performance of the radiomics model was relatively less than satisfactory in patients with brain metastases from lung cancer. Hence, further study should be conducted to investigate the generalizability of the model to other tumor types and to other MRI scanners as well as various imaging parameters. Even prospective clinical trials are warranted to provide high-level evidence before implementation in clinical practice. Second, the homogeneous patient population in the training set might lead to relatively low variance of the candidate predictor named radiation dose to the brain per fraction. For this reason, even in the case where a true correlation exists, it might fail to reach statistical significance. This might also potentially influence the performance of the model. So far, no studies have demonstrated whether dose fraction has significant impact on the treatment response to bevacizumab in brain necrosis. Therefore, further studies are still needed to investigate the potential association of radiation dose fraction with treatment response in larger datasets. Third, the candidate clinical variables associated with a therapeutic effect of bevacizumab monotherapy on brain necrosis were determined from clinical experience and previous studies (4, 37). Subsequent studies are needed to address these limitations.

In summary, we developed and validated a radiomics model that incorporates the pretreatment MRI-based radiomics signature and clinical variables for the prediction of the response to bevacizumab in patients with brain necrosis after radiotherapy. The proposed radiomics model performs well and may provide an effective and easy-to-use tool for clinical decision-making.

M.L.K. Chua reports personal fees from Janssen and Astellas, personal fees and nonfinancial support from Varian, personal fees from MSD Oncology and Bayer, personal fees and nonfinancial support from AstraZeneca (research collaborative agreement), personal fees from Illumina, nonfinancial support from PVMed (research collaboration on AI-based radiotherapy planning), and nonfinancial support from Decipher Biosciences (research collaborative agreement on molecular profiling of prostate cancers) outside the submitted work. No potential conflicts of interest were disclosed by the other authors.

J. Cai: Conceptualization, formal analysis, methodology, writing-original draft, writing-review and editing. J. Zheng: Conceptualization, formal analysis, methodology, writing-original draft, writing-review and editing. J. Shen: Resources, formal analysis, methodology, writing-review and editing. Z. Yuan: Resources, project administration, writing-review and editing. M. Xie: Formal analysis, writing-review and editing, acquisition of data. M. Gao: Writing-review and editing, acquisition of data. H. Tan: Formal analysis, methodology, writing-review and editing. Z. Liang: Writing-review and editing, acquisition of data. X. Rong: Writing-review and editing, acquisition of data. Y. Li: Project administration, writing-review and editing, acquisition of data. H. Li: Writing-review and editing, acquisition of data. J. Jiang: Writing-review and editing, acquisition of data. H. Zhao: Formal analysis, writing-review and editing. A.A. Argyriou: Formal analysis, writing-review and editing. M.L.K. Chua: Formal analysis, writing-review and editing. Y. Tang: Conceptualization, resources, supervision, writing-original draft, project administration, writing-review and editing.

This work was supported by the National Key R&D Program of China (2017YFC1307500, 2017YFC1307504), the National Natural Science Foundation of China (81925031, 81820108026), Science and Technology Program of Guangzhou (202007030001), and the Tip-top Scientific and Technical Innovative Youth Talents of Guangdong special support program (2016TQ03R559) to Y. Tang; the Science and Technology Planning Project of Guangzhou (201704030033), and the National Natural Science Foundation of China (81872549) to Y. Li; the National Medical Research Council Clinician-Scientist Award (NMRC/CSA/0027/2018), the National Research Foundation Clinical Research Program (NRF/CRP17-2017-05), and the Duke-NUS Oncology Clinical Program-Proton Therapy Research Fund to M.L.K. Chua.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.