State space model fuses language and vision cues for medical image segmentation

A novel state space model integrating clinical text prompts with imaging outperformed leading multimodal segmentation networks, achieving state-of-the-art Dice scores across three public benchmarks (QaTa-COVID19, MosMedData+, MoNuSeg) with lower computational cost (GFLOPs).

• Design: Retrospective benchmarking on three public datasets (radiology and histopathology) against strong multimodal baselines including LViT and RecLMIS; sample sizes per dataset not specified in the abstract.
• Key innovation: A Multimodal Interactive Guide Decoder (MIGD) for selective visual-text fusion with linear complexity, plus a Multi-Expert Uncertainty Refinement (MEUR) module for calibrated pixel-wise uncertainty.
• Limitation: Retrospective, public-benchmark study without reported prospective or external clinical validation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.