Researchers built an AI system that explains its reasoning before it draws the boundaries of a lesion in a medical scan.
The team introduced two things: ComLesion-14K, a benchmark of complex lesion cases paired with step-by-step reasoning chains, and CORE-Seg, a framework that links a language-reasoning model to a pixel-level segmentation tool through a module they call a Semantic-Guided Prompt Adapter. Training runs in two stages, supervised fine-tuning followed by a reinforcement-learning stage called GRPO, with a reward system designed to keep useful feedback flowing even on hard cases. On the new benchmark, CORE-Seg posted a mean Dice score (a standard measure of overlap between predicted and actual lesion boundaries) of 37.06%, which the researchers say is 14.89% higher than the second-best baseline. It also cut its failure rate to 18.42%.
Segmentation tools are good at drawing boundaries but bad at explaining why. General-purpose multimodal AI models are the opposite: broad common sense, little specialized visual judgment. Combining both in one pipeline, rather than bolting a chatbot onto a segmentation model, is the more interesting bet here, and the new benchmark may end up mattering as much as the model, since reasoning-labeled lesion data has been scarce.
A 14.89% edge on a brand-new benchmark that the same team built is worth noting, not popping champagne over; the real test is whether this reasoning holds up on lesions nobody wrote a training chain for.