External Validation Of A Radiomics-Based Machine Learning Model For Distinguishing Radiation Necrosis From Progression Of Brain Metastases Treated With Stereotactic Radiosurgery

Abstract

Radiation necrosis (RN) is common and potentially debilitating after stereotactic radiosurgery (SRS) for brain metastases (BM). The goal of this study is to validate a previously reported radiomics signature for distinguishing RN from true progression (TP) using an independent image set. Patients with BM were treated with Gammaknife SRS at Wake Forest University between 2004 and 2012 (WF dataset). Those who developed radiographic evidence of progression in the treated lesion and subsequently had pathologic confirmation were included. The lesions were identified on T1 post-contrast (T1c) and T2 fluid attenuated inversion recovery (T2 FLAIR) MRI. 51 T1c and 51 T2 FLAIR radiomics features were extracted using a software developed in house. An IsoSVM machine learning classifier model trained using our institutional data (JH dataset) was tested using WF dataset. Fifty-three lesions from 43 patients were included in this study, of which all had T1c and 33 had T2 FLAIR images available. Seven lesions were pure RN pathologically, while the TP group included 37 pure tumor and 9 mixed tumor and necrosis. Twenty-two patients (51%) underwent whole brain radiotherapy with a median dose of 35Gy. Median SRS marginal dose was 18Gy (range 10-22Gy). There was no statistically significant difference among baseline characteristics between RN and TP. Of the 6 T1c radiomic features from the JH dataset, 4 demonstrated the same trend in the WF dataset (neighborhood grey tone difference matrix (NGTDM) texture strength, NGTDM coarseness, grey level run length matrix (GLRLM) grey level nonuniformity, and GLRLM run percentage). NGTDM texture strength (mean RN 161.13 vs TP 70.86, p = 0.06), and NGTDM coarseness (mean RN 0.03 vs TP 0.02, p = 0.07) approached significance. Of the 4 FLAIR features, 3 showed similar trend (kurtosis, minimum, NGTDM texture strength). The original IsoSVM model trained with all 6 T1c and 4 FLAIR features showed sensitivity of 65%, specificity of 87%, and AUC of 0.81. Due to the limited number of cases with FLAIR available, we re-trained the model with JH dataset using only the 4 T1c features showing the same trend between the two datasets, resulting in sensitivity and specificity of 80% and 73%, with AUC of 0.8. This model when tested on the WF dataset showed sensitivity and specificity of 71% and 52%, with AUC of 0.62. When the model was re-trained with the 4 T1c and 3 FLAIR features trending similarly between the two datasets, the sensitivity and specificity were 50% and 100%, and the AUC improved to 0.84. Radiomics profiling of post-SRS T1c and T2 FLAIR MRI showed promising trends between RN and TP in 4 T1c features across two independent datasets. Radiomics-based IsoSVM model can distinguish RN vs TP in individual patients. Statistical significance and model performance were inferior in the validation (WF) dataset due to smaller sample size and class imbalance with fewer RN cases. Further studies with larger datasets are needed to optimize and validate the model.