ReconMIL: Synergizing Latent Space Reconstruction with Bi-Stream Mamba for Whole Slide Image Analysis

A research team from multiple institutions has introduced ReconMIL, a breakthrough framework designed to overcome critical limitations in whole slide image (WSI) analysis for pathology. Current methods using multiple instance learning (MIL) struggle with domain gaps between generic foundation model features and specific histological tasks, and they often suffer from over-smoothing where sparse diagnostic signals get lost in background context. ReconMIL addresses these issues through two key innovations.

First, it employs a Latent Space Reconstruction module that adaptively projects generic, task-agnostic features into a compact, task-specific manifold, significantly improving feature separability and boundary delineation for specific diagnostic needs. Second, to prevent information dilution, the framework implements a novel bi-stream architecture: one stream uses Mamba-based sequence modeling to capture long-range dependencies and global contextual priors, while a parallel CNN-based stream preserves subtle local morphological anomalies that are crucial for diagnosis.

These streams are dynamically fused through a scale-adaptive selection mechanism that determines when to prioritize global architectural context versus local saliency. Evaluations across multiple diagnostic and survival prediction benchmarks demonstrate that ReconMIL consistently outperforms current state-of-the-art methods. Visualization results confirm the model's superior ability to localize fine-grained diagnostic regions by effectively balancing global structure with local granularity, while actively suppressing irrelevant background noise that has plagued previous approaches.