Abstract 2437: Deep-learning triage of 3D pathology datasets for accurate and efficient pathologist assessments

Abstract Introduction: Human tissue exhibits heterogeneous morphologies and microenvironments that are readily appreciated in 3D. In the current standard of clinical care, pathologists assess only a few thin 2D sections, typically sampled arbitrarily from one side of a tissue volume/biopsy, in which diagnostically important tissue structures may be missed or misinterpreted. Recent advances in non-destructive 3D pathology enabled by open-top light-sheet (OTLS) microscopy allow for comprehensive imaging of large clinical specimens (e.g. whole biopsies), thus offering the potential to improve the accuracy of diagnostic determinations. While fully automated computational analysis methods are being developed by us and others, a potential early route towards clinical adoption of 3D pathology is to rely upon pathologists for final diagnostic determinations based on viewing familiar 2D H&E-analog cross-sections from the 3D datasets. Because it is not practical for pathologists to manually review massive 3D pathology datasets (equivalent to hundreds of sections per specimen), we present a deep-learning-based triage method to automatically identify the highest-risk 2D cross-sections within 3D pathology datasets for time-efficient pathologist review. Methods: For each 2D cross-section in a 3D pathology dataset, 2D tissue patches are converted into representative low-dimensional feature vectors using the pre-trained CONCH model. We perform attention-based aggregation of the patches within each 2D cross-section and then integrate neighboring 2D sections across depths to predict a context-aware 2.5D risk score. The ground truth labels for model training are based on consensus pathologist diagnoses. Results: We implemented the AI-triaged 3D pathology workflow for two clinical scenarios. For prostate cancer risk stratification based on Gleason grading, our 2.5D method achieved an area under the curve (AUC) of 0.93 for triaging low vs. higher-grade prostate cancer cross-sections on around 110 specimens, outperforming AI analysis on individual 2D cross-sections without depth context (AUC 0.87). In another clinical scenario, where endoscopic biopsies from Barrett’s esophagus patients were screened for neoplasia (dysplasia or adenocarcinoma), the 2.5D method similarly surpassed its 2D counterpart (AUC 0.92 vs. 0.89) for identifying benign vs. neoplastic cross-sections on around 95 specimens. Conclusion: These preliminary results indicate that the integration of additional spatial context along the depth dimension enhances the model's discriminative capabilities to triage high-risk cross-sections from 3D pathology datasets. Finally, preliminary clinical validation studies comparing diagnoses based on AI-triaged 3D pathology vs. conventional 2D histopathology are underway to validate that our methods can improve diagnostic accuracy while minimizing pathologist workloads. Citation Format: Gan Gao, Andrew H. Song, Lindsey Erion Barner, Fiona Wang, Renao Yan, Huai-Ching Hsieh, David Brenes, Rui Wang, Sarah S. Chow, Kevin W. Bishop, Lawrence D. True, Deepti M. Reddi, William M. Grady, Faisal Mahmood, Jonathan T. Liu. Deep-learning triage of 3D pathology datasets for accurate and efficient pathologist assessments [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2437.