Enhancing prognostic accuracy in lung cancer brain metastases through histopathologic feature integration and AI-based image analysis.

e13608 Background: The management of brain metastases from lung cancer has emerged as a critical aspect for enhancing both longevity and quality of life as systemic disease control has improved. Previous studies aiming to forecast patient survival following diagnosis and treatment have frequently overlooked the significance of histopathologic features, thereby potentially missing valuable insights into disease prognosis. Methods: We assembled a dataset comprising whole slide images (WSIs) of surgically resected or biopsied intracranial lung adenocarcinoma metastases, along with corresponding clinical features at diagnosis. Employing publicly available pretrained feature encoders, we generated patches and extracted features from each slide. Weakly-supervised attention-based models were trained on these extracted features to classify WSIs based on their clinical and histologic characteristics. Subsequently, we evaluated the performance of learned features in estimating overall survival using Cox proportional hazard models and neural-network-based models against ground-truth labels at the time of new brain metastasis diagnosis. Results: Neural network-based prognostic models that incorporated pathologic features like tumor-infiltrating lymphocytes (TILs), necrosis, and fibrosis from intracranial metastases biopsies demonstrated superior performance compared to models without these features (logrank test: p = 0.01 vs p = 0.17). Models using learned features from attention-based networks also outperformed those relying on clinical pathologist assessments (logrank test: p < 0.005 vs p = 0.01). Cox proportional hazards models incorporating clinical data with either ground-truth pathologic data or learned histopathologic features showed similar performance (Concordance: 0.64 vs 0.65). Conclusions: The integration of histopathologic features derived from surgically resected lung adenocarcinoma metastases with clinical features enhanced overall survival prediction compared to using clinical features alone. Artificial intelligence-based models trained on whole slide images (WSIs) demonstrated comparable performance to ground-truth data assessed by clinical pathologists. Notably, combining these AI-driven models with clinical features further improved the predictive accuracy of overall survival estimates.