MedRegion-CT reads CT scans with region-specific precision, outperforming existing AI

Current CT report generation frameworks rely on global feature representations, often missing region-specific abnormalities. To address this, researchers developed MedRegion-CT, a multimodal large language model with three key innovations. First, a Region-based SlowFast Tokenizer jointly models global and fine-grained information by extracting tokens guided by clinically meaningful anatomical regions. Second, generated pseudo-masks direct the model to attend to diagnostically important areas, enabling systematic scan context understanding. Third, quantitative lesion data—size, diameter, and spatial location—is encoded as structured textual prompts for context-aware, clinically informed report generation.

Validated on multi-institutional structured report generation benchmarks, MedRegion-CT achieves state-of-the-art performance, surpassing existing approaches in both linguistic quality and clinical accuracy. Accepted at ECCV 2026, the work is fully open-source. This innovation promises to reduce missed abnormalities in CT reports, offering radiologists a powerful AI assistant that captures critical region-specific details.