Deng R, Cui C, Remedios LW, Bao S, Womick RM, Chiron S, Li J, Roland JT, Lau KS, Liu Q, Wilson KT, Wang Y, Coburn LA, Landman BA, Huo Y. Cross-scale multi-instance learning for pathological image diagnosis. Med Image Anal. 2024 Feb 27;94:103124. doi: 10.1016/j.media.2024.103124. Epub ahead of print. PMID: 38428271.
In a new study, researchers address the challenge of analyzing high-resolution whole slide images (WSIs) in digital pathology, which often requires understanding information at multiple scales. They introduce a novel cross-scale multi-instance learning (MIL) algorithm that not only processes image patches but also integrates vital inter-scale relationships—crucial elements that mimic the diagnostic process of human pathologists. This cross-scale MIL approach enhances the analysis by aggregating information across different magnifications, leading to improved diagnostic accuracy. The researchers have demonstrated superior performance of their method on both specialized in-house and publicly available datasets. Additionally, they have created a toy dataset with scale-specific features to facilitate understanding and visualization of their cross-scale attention mechanism, making their methods and findings accessible through a public GitHub repository.
