EliSeg: New AI Framework for Report-Grounded Abnormality Segmentation

A research article titled 'EliSeg: Verified Target Construction for Report-Grounded Abnormality Segmentation' has been released on arXiv (ID: 2608.07299). This study tackles a significant issue in medical imaging, where radiology reports outline clinical observations without detailing actionable segmentation targets. Such reports may include findings that are present, negated, prior, uncertain, or irrelevant, and multiple valid abnormalities may exist simultaneously. Current segmentation techniques often overlook this ambiguity by relying on a predefined target identity or spatial prompt, functioning as a concealed target oracle. The authors introduce EliSeg, a framework that combines target construction with mask generation through an actor-verify-revise approach. A grammar-constrained Actor suggests target slots and masks, while an independent text-only Verifier reconstructs the potential finding inventory. Revision selectively re-executes the shared Actor when discrepancies arise in their target sets. This paper serves as a cross-type announcement, indicating its submission to a journal following a conference presentation, and highlights the intersection of artificial intelligence and medical imaging, offering potential for automated radiology analysis.