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). Identifying which patients benefit from WBRT following chemotherapy, and which patients can be adequately treated with chemotherapy alone remains a persistent clinical challenge. Although WBRT is associated with improved outcomes, it also carries a risk of neuro-cognitive side effects. This study aims to refine patient phenotyping for PCNSL by leveraging deep learning (DL) extracted imaging biomarkers to enable personalized therapy. Methods: Our study included 71 patients treated at our institution between 2009-2021. The primary outcome of interest was overall survival (OS) assessed at one-year, two-year, and median cohort survival cutoffs. The DL model leveraged an 8-layer 2D convolutional neural network which analyzed individual slices of post-contrast T1-weighted pre-treatment MRI scans. Survival predictions were made using 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 statistically evaluated using the log-rank test. External validation of our model was performed with a cohort of 40 patients from an external institution. Results: The cohort's average age was 65.6 years with an average OS of 2.80 years. The one-year, two-year, and median OS models achieved AUCs of 0.73 (95% C.I., 0.60-0.85), 0.70 (95% C.I., 0.58-0.82), and 0.73 (95% C.I., 0.58-0.82) respectively. KM survival curves showcased discrimination between low and high-risk groups in all models. External validation with our one-year model achieved AUC of 0.64 (95% C.I., 0.63-0.65) and significant risk discrimination. A sub-analysis showcased stable 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. Keywords: PCNSL; deep learning; convolutional neural network; magnetic resonance imaging; prognosis; personalized medicine