Effective latent hierarchical feature fusion in multiple instance learning for Whole Slide Image classification

Abstract

Deep learning applications in computational pathology have revolutionized cancer diagnostics through histopathology tissue analysis of Whole Slide Images (WSIs). However, the gigapixel scale of WSIs presents significant challenges for traditional approaches. While Multiple Instance Learning (MIL) frameworks address these challenges by treating WSIs as bags of patches, existing methods often focus solely on information extraction modules, neglecting effective decoding of latent features. This paper introduces LHFF-MIL, a novel framework that emphasizes latent feature decoding and fusion in MIL. Our key contribution is the Latent Feature Distribution Decoder (LFDD), which efficiently decodes diverse information from high-dimensional semantics across different WSI resolutions, enabling explicit measurement of image informativeness for tumor detection. Evaluated on three real-world datasets of breast and gastric cancer, LHFF-MIL consistently outperforms competing methods, demonstrating statistically significant diagnostic accuracy improvement from 0.27% to 1.44% with at least 95% of confidence level. These improvements advance computational pathology by enhancing classification performance, potentially enabling more reliable computer-aided diagnosis systems in clinical settings. Code will be available upon acceptance.