Work

LG-MIL: Language-Guided Multiple Instance Learning for Coordinate-Aligned Attention Analysis in Whole-Slide Images

Year: 2026
Type: preprint
Abstract: Language-Guided Multiple Instance Learning (LG-MIL) is introduced as a weakly supervised framework for whole-slide image (WSI) analysis. The proposed approach integrates visual and textual information to support interpretable slide-level prediction. Patch-level features extracted from WSIs are combined with class-level textual prompts encoded in a shared multimodal space. To enable semantically informed aggregation under weak supervision, text embeddings guide patch weighting through a cross-modal attention mechanism. A coordinate retrofit strategy is further employed to recover per-patch (x,y) locations at inference, enabling the generation of attention heatmaps, top-K tile visualizations, and saliency-based summaries for spatial interpretability. To regulate the sharpness and semantic alignment of attention distributions, the framework incorporates two interpretable controls: the attention temperature τ(tau) and the contrastive loss weight α. Experiments on patient-exclusive validation and test splits demonstrate competitive slide-level classification performance, with measurable variation across random initializations and hyperparameter settings. Saliency analyses further investigate the relationship between model confidence and attention dispersion, while ablation studies show that attention temperature influences both attention concentration and predictive performance. Lower temperature values produce more concentrated attention distributions, whereas higher values result in more diffuse attention patterns. Overall, LG-MIL provides a reproducible pipeline for language-guided, coordinate-aligned attention analysis in WSIs, enabling transparent investigation of model behavior under weak supervision. The framework highlights the potential of text guidance, spatial alignment, and saliency quantification for advancing interpretability in large-scale WSI modeling without requiring dense or clinical annotations.
Language: English
Open Access status: green
OpenAlex ID: W7236700949
Cites: 25
Cited by:
Related to: 10