KHiM-Mamba injects pathology knowledge into Mamba for sharper WSI analysis

Whole slide images (WSIs) are gigapixel-scale pathology scans where diagnostically decisive regions are often sparse and surrounded by irrelevant tissue. Standard multiple instance learning (MIL) pipelines treat slides as bags of patches, but purely vision-driven models struggle to focus on what matters. Recent selective state-space models (SSMs) like Mamba offer linear complexity for long sequences, yet they still ignore medical knowledge during slide encoding, allowing irrelevant evidence to accumulate and dilute critical cues.

KHiM-Mamba, developed by Qixiang Zhang and colleagues, tackles this by redesigning Mamba's core selective mechanism: a knowledge-aware hidden-state modulation layer that injects explicit pathology priors at each encoding step. This guides exactly what visual evidence gets accumulated and retrieved, keeping the model's state aligned with diagnostically relevant patterns. Additionally, the architecture includes a local-adaptive vocabulary retrieval module that leverages large language models to assign each image patch fine-grained, tissue-specific semantic descriptions, enabling precise modulation across diverse tasks. Tested on 11 public benchmarks spanning 4 pathology tasks (including cancer subtyping and grading), KHiM-Mamba consistently outperforms existing MIL and SSM-based methods, setting a new state of the art.