RADT-12. DERIVING IMAGING BIOMARKERS FOR PRIMARY CENTRAL NERVOUS SYSTEM LYMPHOMA USING DEEP LEARNING

Abstract PURPOSE Primary central nervous system lymphoma (PCNSL) is typically treated with chemotherapy, steroids, and/or whole brain radiotherapy (WBRT). Determining which patients benefit from WBRT after chemotherapy and which are adequately treated with chemotherapy alone remains a clinical challenge. Although WBRT is associated with improved outcomes, it also carries a risk of neurocognitive side effects. This study aims to refine PCNSL patient phenotyping by leveraging deep learning (DL) extracted imaging biomarkers to enable personalized therapy. METHODS Our study included 71 patients treated between 2009-2021. The primary outcome of interest was overall survival (OS) assessed at one- and two-year cutoffs. The DL model leveraged an 8-layer 2D convolutional neural network analyzing individual slices of post-contrast T1-weighted pre-treatment MRI scans. Survival predictions utilized a weighted voting system related to tumor size. Model performance was assessed with accuracy, sensitivity, specificity, and F1 scores. Time-dependent AUCs were calculated and C-statistics were computed to summarize the results. Kaplan-Meier (KM) survival analysis assessed differences between low and high-risk groups and were evaluated using the log-rank test. RESULTS The cohort’s average age was 65.6 years with average OS of 2.80 years. For one-year OS, the model achieved an AUC of 0.73 (0.60-0.85), accuracy of 0.73 (0.61-0.82), sensitivity of 0.72 (0.54–0.85), and specificity of 0.73 (0.58–0.84). For two-year OS, the model achieved an AUC of 0.70 (0.58-0.82), accuracy of 0.70 (0.58-0.79), sensitivity of 0.68 (0.55–0.81), and specificity of 0.71 (0.55–0.84). Model performance is presented with respective 95% confidence intervals. KM curves showed the one and two-year models effectively discriminated between low- and high-risk group, reporting c-statistics of 0.73 (p<.001) and 0.71 (p<.001) respectively. A sub-analysis confirmed consistent model performance across different tumor volumes and focality. CONCLUSION DL classifiers of PCNSL MRIs can stratify patient phenotypes beyond traditional risk paradigms. Given dissensus surrounding PCNSL treatment, DL can augment risk stratification and treatment personalization, especially with regards to WBRT decision making.