1071 - BrAIn_Path: A Novel Deep Learning Histopathology Score for Prognostication and Risk Stratification In Resected Brain Metastases
Presenter(s)
Z. Yazdani1,2, R. R. Patel1, A. Skakodub1,3, N. Kumar4, M. I. Parker1, K. Tsai1, K. M. Boehm1, H. Walch5, R. Homsi5, S. Nanda4, S. Singi4, B. S. Imber1, N. S. Moss6, J. N. Stember7, K. Yu6, D. J. Foran2,8, C. Kinslow1, N. Schultz5, C. M. Vanderbilt4, and L. R. G. Pike1; 1Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 2Rutgers Robert Wood Johnson Medical School, New Brunswick, NJ, 3Department of Radiation Oncology, Icahn School of Medicine at Mount Sinai, New York, NY, 4Department of Pathology and Laboratory Medicine, Memorial Sloan Kettering Cancer Center, New York, NY, 5Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, NY, 6Department of Neurosurgery, Memorial Sloan Kettering Cancer Center, New York, NY, 7Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, NY, 8Rutgers Cancer Institute of New Jersey, New Brunswick, NJ
Purpose/Objective(s): To develop a deep learning histopathology risk score trained on hematoxylin and eosin (H&E) whole-slide images (WSI) from resected brain metastases and assess whether it provides independent prognostic value beyond the established Diagnosis-Specific Graded Prognostic Assessment (DS-GPA) and a baseline clinical features model.
Materials/Methods: The primary endpoint was overall survival (OS), measured from date of resection to death or last follow-up. In a single-institution retrospective cohort, patients with resected brain metastases were consecutively identified. One representative H&E whole-slide image per patient was selected using pre-specified criteria, prioritizing viable tumor, minimal artifact, and overall slide quality. We developed a deep learning Cox proportional hazards model using foundation model-derived WSI features to produce a patient-level risk score (BrAIn_Path). Five-fold cross-validation was performed (80% train; 20% validation per fold), generating out-of-fold BrAIn_Path scores for all patients. Hazard ratios (HRs) were reported per 1-standard deviation increase in BrAIn_Path.
Results: Among 618 patients, median follow-up was 18.8 months and median OS was 21.3 months (425 events). The most common histologies were NSCLC (N=194, 31%), breast (N=117, 19%), melanoma (N=79, 13%), and GI (N=75, 12%), followed by all other histologies (N=153). In the DS-GPA eligible cohort (N=502), BrAIn_Path was independently prognostic in a Cox model with DS-GPA (HR/SD=1.26, p<0.001) and improved overall concordance (C-index 0.62?0.64). Independent prognostic value was consistent across NSCLC (HR/SD=1.38, p<0.001; ?C-index=+0.05), breast (HR/SD=1.33, p=0.009; ?C-index=+0.02), and GI (HR/SD=1.35, p=0.035; ?C-index=+0.01); the melanoma subgroup did not reach statistical significance. In a pre-specified baseline clinical model that included primary histology, age, KPS, number and size of brain metastases, intracranial location, postoperative radiotherapy, and extracranial disease status, BrAIn_Path remained independently prognostic (N=618; HR/SD=1.18, p<0.001; ?C-index=+0.01). A median split of BrAIn_Path stratified patients into high- and low-risk groups (median OS high-risk vs low-risk: 15.1 vs 24.2 months; log-rank p<0.001). Pathologist review of high-risk tumors identified enrichment for necrosis and high-grade morphologic features.
Conclusion: BrAIn_Path provides independent prognostic information and, when incorporated, outperforms gold-standard DS-GPA alone for patients with resected brain metastases. External validation at an independent institution is underway. These findings suggest that WSI-derived phenotypic information from resected brain metastases has the potential to augment current prognostic frameworks, refine clinical trial stratification, and guide patient care.