Abstract 4916: Enhancing multi-organ frozen section cancer discrimination model by sharing cancer discrimination and organ classification task

Abstract

Abstract Introduction: Frozen section(FS) testing is an important method that removes part or all of a patient's tissue of a suspected tumor during surgery and microscopic examination to determine a diagnosis and surgical method that requires a rapid and accurate diagnosis. In this study, we developed a deep learning model to determine the presence of cancer and the type of organ in H&E FS whole slide images(WSIs) of 5 organs such as Breast(BR), Lung(LU), Stomach(ST), Breast Sentinel Lymph Node(LN), and Prostate(PR). Design: In this study, we utilized a dataset comprising 19,881 H&E FS WSIs and 11,985 H&E-stained formalin-fixed paraffin-embedded (FFPE) WSIs to enhance generalization. The dataset included FS and FFPE WSIs from 33 different organs sourced from The Cancer Genome Atlas (TCGA) and additional FS and FFPE WSIs from a domestic hospital and the Camelyon 16 dataset, respectively. For the FS dataset, 20% was set aside for performance evaluation, while the remaining was randomly split into a 3:1 ratio for train and validation. In our methodology, we utilized instance-based multi-instance learning (MIL) to identify patches (1024*1024 pixels, 20x mag) with a high likelihood of cancer presence in each WSI. Then, we implemented multi-task learning (MTL) to predict the presence of cancer and the organ simultaneously. Result: The table shows the WSI level performance through the 5 organs. The top 5 rows show results of MIL that only discriminate cancer. In comparison, the bottom 5 rows show results of MIL with MTL that share features from the cancer discrimination and organ classification. MIL with MTL make higher roc/auc than MIL only. Conclusion: This study explores the automated analysis of H&E-stained FS WSIs using a deep-learning to discriminate cancer and classify organs. We could figure out that training cancer discrimination and organ classification simultaneously can make general features, and it leads to increased diagnostic performance. Performance Table Between MIL vs MIL + MTL Method Organ F1-score Sensitivity Specificity ROC/AUC Organ Accuracy MIL LN 0.9358 0.9669 0.6170 0.7919 NaN ST 0.9458 0.9583 0.8302 0.8943 LU 0.9577 0.9457 0.9275 0.9366 BR 0.9902 0.9895 0.9757 0.9826 PR 0.9346 0.9328 0.8786 0.9057 MIL + MTL LN 0.9448 0.9448 0.7872 0.8660 0.8816 ST 0.9771 0.9697 0.9623 0.9660 0.7595 LU 0.9700 0.9721 0.9203 0.9462 0.9509 BR 0.9932 0.9910 0.9877 0.9894 0.8979 PR 0.9480 0.9515 0.8913 0.9214 0.9803 Citation Format: Joonho Lee, Joonyoung Cho, Junho Lee, Yoon-La Choi, Kyungsoo Jung, Tae-Yeong Kwak, Sun Woo Kim, HyeYoon Chang. Enhancing multi-organ frozen section cancer discrimination model by sharing cancer discrimination and organ classification task [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4916.