A Radiomics-Integrated Deep Learning Model for Identifying Radionecrosis Following Brain Metastasis Stereotactic Radiosurgery (SRS)

Abstract

<h3>Purpose/Objective(s)</h3> To develop a radiomics-integrated deep learning (RIDL) model for identifying radionecrosis in brain metastasis patients with post-SRS radiographic progression. Such a novel model is hypothesized to outperform classic radiomics and deep learning approaches in radionecrosis/recurrence differentiation. <h3>Materials/Methods</h3> The model was developed based on a 51-patient cohort at our institution with post-SRS radiographic progression and known biopsy outcome (37 radionecrosis, 14 recurrence) from Laser Interstitial Thermal Therapy (LITT). Utilizing the 3-month post-SRS high resolution T1+c volume, the RIDL model comprises three key steps: 1) 184 radiomics features (RFs) are extracted from within the SRS planning target volume (PTV) and 60% isodose volume (V₆₀) followed with Z-score normalization; 2) a deep neural network (DNN) mimicking the encoding path of U-net is trained for radionecrosis/recurrence prediction using the 3D volume. Prior to the binary prediction output, latent variables in the DNN are extracted as 512 deep features (DFs); and 3) all extracted features are synthesized as an input of support vector machine (SVM) execution. Key features with higher linear kernel weighting factor values are identified by clustering analysis, and these key features are utilized by SVM to generate the final radionecrosis/recurrence prediction result. During the model training, a 7:3 training/test data sample ratio was adopted, and 50 model versions were acquired with random validation sample assignments. Sensitivity, specificity, accuracy, and ROC of the model were evaluated, and these results were compared with 1) classic radiomics-based prediction (i.e., radiomics features-only for SVM input) and 2) DNN prediction results. Potential associations between the identified key radiomics/deep features and patients' genetic profiles were studied. <h3>Results</h3> As seen in the table, while radiomics-based prediction achieved acceptable accuracy but very low sensitivity, DNN prediction achieved similar accuracy with an improved sensitivity; in contrast, the RIDL achieved the best prediction accuracy and sensitivity results with 32 identified key features (3 RFs+29 DFs), and it also demonstrated superior ROC results. For patients with NSCLC primary disease, 2 RFs extracted in SRS V₆₀ exhibited significant correlation with ALK mutation (R>0.6), and 1 DF exhibited significant correlation with EGFR mutation (R>0.7). <h3>Conclusion</h3> The developed RIDL can accurately differentiate brain metastasis radionecrosis/recurrence using a single post-SRS MR scan. Future work that investigates its integration with genomics as a comprehensive radiogenomic tool is indicated.