Abstract 5056: Automated Gleason grading of digitized frozen section prostate tissue slide images

Abstract

Abstract Frozen sections are used in intraoperative consultations which need rapid microscopic analysis and in certain pathological procedures that require fresh tissue. Due to the various artifacts present on frozen section tissue slides, it is difficult to automatically analyze them. We tried automatic analysis of frozen sections. Two types of segmentation deep learning models detecting Gleason patterns were used. Both models were trained from hematoxylin and eosin (H&E) stained formalin-fixed, paraffin-embedded prostate tissue slide images. Each slide image was reviewed by an experienced pathologist, and all prostate cancer lesions were annotated with corresponding Gleason grades. One model was trained with H&E stained images (RS model), and the other model was trained with Hematoxylin-only (H-only) stained images (H-only model). A color deconvolution method was used to generate H-only stained images from original H&E stained images. In evaluation, frozen section prostate cancer slide images from The Cancer Genome Atlas (TCGA) were used. To diagnose the slide, we splitted whole slide images (WSI) into patches and analyze those patches to segment the Gleason pattern area. For the WSI result, we reconstructed the result as a whole size heatmap, then we counted the Gleason pattern to diagnose the slide as the prostate grade group. In detecting malignant cases, the RS model achieved a sensitivity of 98% and the H only model of 97%, respectively. In detecting clinically significant risk cases (grade group 2 or over), sensitivities of 99% and 96% were achieved, respectively. Most of the errors were all false-positive cases. In particular, the RS model mainly had false positives because of the ice crystal artifact, and the H-only model had many false positives for the folded part and the part suspected of being a prostatic intraepithelial neoplasia (PIN). False negatives were common in the compression artifacts for both models. We confirmed that both models showed high performance in cancer detection. In addition, they showed high accuracy in the classification of high-risk and low-risk groups of cancer. Further performance improvement can be expected with additional training of data related to artifacts or PIN. Citation Format: Joonyoung Cho, Tae-Yeong Kwak, Sun Woo Kim, Hyeyoon Chang. Automated Gleason grading of digitized frozen section prostate tissue slide images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5056.