Med-Scout: Curing MLLMs' Geometric Blindness in Medical Perception via Geometry-Aware RL Post-Training

Despite recent Multimodal Large Language Models (MLLMs)' linguistic prowess in medical diagnosis, we find even state-of-the-art MLLMs suffer from a critical perceptual deficit: geometric blindness. This failure to ground outputs in objective geometric constraints leads to plausible yet factually incorrect hallucinations, rooted in training paradigms that prioritize linguistic fluency over geometric fidelity. This paper introduces Med-Scout, a novel framework that "cures" this blindness via Reinforcement Learning (RL) that leverages the intrinsic geometric logic latent within unlabeled medical images. Instead of relying on costly expert annotations, Med-Scout derives verifiable supervision signals through three strategic proxy tasks inspired by the systematic reading and reasoning patterns of clinicians: Hierarchical Scale Localization, Topological Jigsaw Reconstruction, and Anomaly Consistency Detection. To rigorously quantify this deficit, we present Med-Scout-Bench, a new benchmark specifically designed to evaluate geometric perception. Extensive evaluations show that Med-Scout significantly mitigates geometric blindness, outperforming leading proprietary and open-source MLLMs by over 40% on our benchmark. Furthermore, this enhanced geometric perception generalizes to broader medical understanding, achieving superior results on radiological and comprehensive medical VQA tasks. The project page is: https://github.com/HKUSTGZ-ML4Health-Lab/Med-Scout.

AI models designed for medical images often suffer from a critical flaw known as "geometric blindness". While they excel at using complex medical terminology, they frequently fail to understand actual spatial relationships, leading them to confidently hallucinate incorrect locations for physical abnormalities. To address this, we developed Med-Scout, a novel training framework that teaches these models visual logic. Instead of relying on expensive human experts, Med-Scout uses automated, game-like tasks to train the AI. These tasks require the model to pinpoint coordinates when zooming in, reconstruct anatomical jigsaw puzzles, and play "spot the difference" to find subtle anomalies. By rewarding the AI for mastering these spatial challenges, Med-Scout successfully cures its geometric blindness. This approach significantly boosts the model's accuracy when analyzing and answering questions about real medical scans. Ultimately, it paves the way for more reliable and trustworthy clinical AI tools that strictly respect the physical reality of the human body.