Abstract 5082: Deep learning model for cancer diagnosis in frozen section sentinel lymph nodes with limited annotations

Abstract Introduction: Frozen section (FS) rapidly examines suspected tumor tissue microscopically during surgery, aiding diagnosis and surgical decisions. However, when applying AI or digital pathology to FS examination, stain and scanner variations pose significant challenges in developing robust deep learning-based diagnostic models. This study proposes a method that reduces stain variation to detect cancer in H&E-stained FS images of breast sentinel lymph nodes. Design: We utilized 19, 881 FS WSIs and 11, 985 FFPE WSIs. The dataset included FS and FFPE WSIs from 33 organs collected from The Cancer Genome Atlas (TCGA), additional FS and FFPE WSIs from a domestic hospital and the Camelyon16 dataset. We reserved 20% of the FS dataset for evaluation; the remaining data were split into training (75%) and validation (25%). We employed instance-based Multiple Instance Learning (MIL) to identify cancerous patches (512×512 pixels at 10× magnification) in each WSI. After achieving sufficient feature discrimination, we fixed the backbone network. We obtained rough annotations from pathologists on 12 lymph node FS WSIs and trained a classifier with various data augmentation. Result: Our method shows that the MIL plus classifier-only approach consistently outperforms MIL with full fine-tuning across diverse datasets (Camelyon16 FFPE and domestic FS). By reducing scanner variation, we significantly improved the model's generalization, which is a primary contribution of this study. Conclusion: These findings suggest that mitigating stain and scanner variations with limited annotations can substantially enhance the performance and generalizability of deep learning models in cancer diagnosis. Citation Format: Joonho Lee, Joonyoung Cho, DoKyung Kim, Yoon-La Choi, Kyungsoo Jung, Tae-Yeong Kwak, Sun Woo Kim, Hyeyoon Chang. Deep learning model for cancer diagnosis in frozen section sentinel lymph nodes with limited annotations [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 5082.