Abstract 2308: Semi-automated Ki-67 index assessment using top-k hotspot recommendation in Ki-67 IHC stained WSIs

Abstract

Abstract Ki-67 index is commonly used as a breast cancer proliferation marker. However, it is a laborious and time-consuming task for a pathologist to directly measure the Ki-67 index. Recently, automation techniques using machine learning techniques such as deep learning have been proposed. In this study, we propose a top-k hotspot recommendation method to help Ki-67 index assessment based on nuclei detection detected by a deep learning model. The dataset included a total of 32 diaminobenzidine-hematoxylin (DAB-H) Ki-67 IHC whole slide images (WSIs) of breast cancer obtained by core biopsy. Each of these WSIs was evaluated by a pathologist using the interval of the Ki-67 index estimate and the scores of 1+, 2+, 3+, and 4+. Additionally, the search area was manually designated to exclude internal control. We developed a Ki-67 index analysis model based on deep learning and image analysis. Nuclei segmentation used the StarDist model. The detected results were filtered based on nucleus size and nucleus eccentricity. The filtering threshold was determined experimentally. To recommend top-k hotspots, a coarse search is performed to move the circle region horizontally and vertically in specific units, and the top k regions based on Ki-67 index are found among areas containing more than 500 cells. Next, a fine search is performed around the k areas in detail. Finally, the top-k hotspots and each detected positive and negative cells, count of positive cells and negative cells, and predicted Ki-67 index are provided. It was confirmed that the Ki-67 score estimated for the detected top-1 hotspot and the pathologist's Ki-67 score had a correlation of Pearson correlation coefficient=0.8153, R2=0.6648. As k increased, the top-k accuracy was observed to increase from 68.75% when k = 1 and 3 to 75.00% when k = 5, based on the pathologist's score. When a pathologist reviewed 10 cases where the Ki-67 score differed from the prediction, 6 cases were judged to be worth reexamining considering the proposed hotspot and Ki-67 index, and 2 cases included stromal cells in cell detection. In two cases, the WSI was judged to be blurry. Cell analysis and top-k hotspot recommendation using deep learning and image analysis were performed on Ki-67 IHC stained WSIs and compared with the pathologist's Ki-67 score. Through semi-automatic top-k hotspot recommendation, the reliability of diagnosis can be increased as an auxiliary test to Ki-67 index assessment. Citation Format: Hyeon Seok Yang, Yunseob Hwang, Yongeun Lee, Kyungsoo Jung, Minjung Sung, Tae-Yeong Kwak, Sun Woo Kim, Hyeyoon Chang. Semi-automated Ki-67 index assessment using top-k hotspot recommendation in Ki-67 IHC stained WSIs [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 2308.