Exploring multi-instance learning in whole slide imaging: Current and future perspectives

Abstract

Whole slide images (WSI), due to their gigabyte-scale size and ultra-high resolution, play a significant role in diagnostic pathology. However, the enormous data size makes it difficult to directly input these images into image processing units (GPU) for computation, limiting the development of automated screening and diagnostic algorithms. As an effective computational framework, multi-instance learning (MIL) has provided strong support in addressing this challenge. This review systematically summarizes the research progress and applications of MIL in WSI analysis, based on over 90 articles retrieved from Web of Science, IEEE Xplore and PubMed. It briefly outlines the unique advantages and specific improvements in handling whole slide images, with a focus on analyzing the core characteristics and performance of mainstream techniques in tasks such as cancer detection and subtype classification. The results indicate that methods like data preprocessing, multi-scale feature fusion, representative instance selection, and Transformer-based models significantly enhance the ability of MIL in WSI processing. Furthermore, this paper also summarizes the characteristics of different technologies and proposes future research directions to promote the widespread application of MIL in pathological diagnosis.