Abstract
Although multi-instance learning (MIL) has succeeded in pathological imageclassification, it faces the challenge of high inference costs due toprocessing numerous patches from gigapixel whole slide images (WSIs). Toaddress this, we propose HDMIL, a hierarchical distillation multi-instancelearning framework that achieves fast and accurate classification byeliminating irrelevant patches. HDMIL consists of two key components: thedynamic multi-instance network (DMIN) and the lightweight instancepre-screening network (LIPN). DMIN operates on high-resolution WSIs, while LIPNoperates on the corresponding low-resolution counterparts. During training,DMIN are trained for WSI classification while generating attention-score-basedmasks that indicate irrelevant patches. These masks then guide the training ofLIPN to predict the relevance of each low-resolution patch. During testing,LIPN first determines the useful regions within low-resolution WSIs, whichindirectly enables us to eliminate irrelevant regions in high-resolution WSIs,thereby reducing inference time without causing performance degradation. Inaddition, we further design the first Chebyshev-polynomials-basedKolmogorov-Arnold classifier in computational pathology, which enhances theperformance of HDMIL through learnable activation layers. Extensive experimentson three public datasets demonstrate that HDMIL outperforms previousstate-of-the-art methods, e.g., achieving improvements of 3.13% in AUC whilereducing inference time by 28.6% on the Camelyon16 dataset.