Core Viewpoint - The research team from Harbin Institute of Technology has proposed a novel Hierarchical Distillation Multi-Instance Learning (HDMIL) framework aimed at rapidly identifying irrelevant patches in whole slide images (WSI) for efficient and accurate classification [1][3][17]. Group 1: Research Background - Pathological images are considered the "gold standard" for cancer diagnosis, with whole slide images (WSI) being a mainstream method due to their high resolution and large data volume [1][2]. - Multi-Instance Learning (MIL) is a primary method for analyzing WSI, but it faces challenges due to the vast amount of information contained in WSI, leading to high costs in data preprocessing and redundancy issues [2][3]. Group 2: Methodology - The HDMIL framework includes two key components: a Dynamic Multi-Instance Network (DMIN) for classifying high-resolution WSI and a Lightweight Instance Prescreening Network (LIPN) tailored for low-resolution WSI [5][9]. - The DMIN utilizes a self-distillation training strategy to identify irrelevant areas in WSI, while the LIPN is designed to quickly identify irrelevant regions in low-resolution WSI, indirectly indicating irrelevant patches in high-resolution WSI [9][12]. Group 3: Experimental Results - HDMIL demonstrated a 28.6% reduction in inference time compared to previous advanced methods across three public datasets [3][15]. - The framework was validated on three datasets: Camelyon16 for breast cancer lymph node metastasis detection, TCGA-NSCLC for lung cancer subtype classification, and TCGA-BRCA for breast cancer subtype classification [4][15]. - HDMIL achieved an AUC of 90.88% and an accuracy of 88.61% on the Camelyon16 dataset, outperforming previous best methods by 3.13% and 3.18%, respectively [14][15]. Group 4: Innovation and Impact - The introduction of the Kolmogorov-Arnold classifier within the HDMIL framework significantly enhances classification performance [11][17]. - The research contributes to the rapid development of digital pathology, particularly in cancer diagnosis, showcasing the potential of AI in medical applications [18][21].