NIMG-03. DEEP LEARNING SURVIVAL ANALYSIS FOR MULTIPLE BRAIN METASTASES

Abstract

Abstract PURPOSE Deep learning survival models show promise for outcome prediction by leveraging the ability to model non-linear relationships between pixel-level imaging predictors and survival data. We hypothesized that a deep learning survival model derived from quantitative imaging predictors would be more effective than traditional models of survival in patients with brain metastases. METHODS We analyzed 831 patients with 3596 total brain metastases treated with primary stereotactic radiosurgery at our institution between 2000-2018. The primary outcome of interest was overall survival following treatment. 851 3D radiomic features were extracted from T1 post contrast MRI images of each brain metastasis and aggregated per patient. Minimum redundancy maximum relevance was used for dimensionality reduction. Relevant features were trained on DeepSurv (Cox proportional hazard neural network architecture) to model overall survival. The disease-specific Graded Prognostic Assessment which uses age, Karnofsky performance status, presence of extracranial metastases, number of brain metastases, and disease-specific molecular characteristics was used as our traditional model. Discriminatory ability between models was assessed with concordance indices (c-index) using 100 bootstrapped samples of 415 patients and evaluated for statistical significance in difference with the 2-sample t-test. RESULTS Median overall survival was 13 months, median age was 63 years, and the most common primary sites were NSCLC (38.5%), melanoma (18.9%), breast (14.9%), SCLC (7.2%), renal (5.3%), and GI (4.7%). The deep learning model using radiomic imaging features (c-index: 0.848, 95% CI [0.811, 0.877]) had superior discriminatory ability compared to the GPA (c-index: 0.380, 95% CI [0.356, 0.403]). Performance was significantly improved with p< 0.001 on the 2-sample t-test. CONCLUSIONS A deep learning model using quantitative radiomic imaging features performed better than a traditional linear model using clinical predictors at modeling survival in patients with multiple brain metastases. This represents a promising method to improve prognostication for patients diagnosed with brain metastases.