MedPlex: A New Vision-Language Model for Clinically Grounded Medical Segmentation

A new framework called MedPlex (Medical Plexus of Vision and Language) has been developed by researchers, offering an end-to-end Vision-Language Model (VLM) that incorporates text guidance as an integral aspect of medical image segmentation learning. Unlike traditional text-guided segmentation approaches that treat language as a post-processing element, MedPlex utilizes Bi-Fusion (Bidirectional Fusion) to enable the simultaneous evolution of visual and textual representations throughout the encoding hierarchy. This framework also features class-level and region-level concept alignment, organizing shared representations at different granularities. Class-level alignment connects each anatomical target to a consolidated clinical concept profile. The study, accessible on arXiv (2608.13690), emphasizes that medical image segmentation is frequently viewed solely as a vision problem, despite its reliance on textual anatomical knowledge. MedPlex seeks to embed text guidance within segmentation learning in a clinically relevant manner. This research was presented as a cross-type submission on arXiv.