Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis
A new framework called Trajectory-Integral Feedback GRPO (TIF-GRPO) has been proposed to enhance the accuracy of medical vision-language models in 3D Computed Tomography analysis. This framework addresses the issue of 'Evaluation Hallucinations' by integrating control-theoretic principles into policy optimization. Experiments show that TIF-GRPO significantly improves abnormality detection and clinical faithfulness in radiology reports.
- ▪Medical vision-language models have advanced but face challenges in 3D CT analysis due to mismatched optimization objectives.
- ▪The proposed Clinical Abnormality Benchmarking Substrate (CABS) decomposes radiology reports into verifiable clinical semantic units.
- ▪TIF-GRPO regulates anatomy-aware rewards through an integral feedback loop, improving clinical reasoning and reducing errors.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20277 (cs) [Submitted on 19 May 2026] Title:Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis Authors:Tianwei Lin, Zhongwei Qiu, Jie Cao, Jiang Liu, Wenjie Yan, Bo Zhang, Yu Zhong, Wenqiao Zhang, Yingda Xia, Ling Zhang View a PDF of the paper titled Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis, by Tianwei Lin and 9 other authors View PDF HTML (experimental) Abstract:Medical vision-language models (VLMs) have rapidly advanced as general-purpose multimodal assistants, yet their deployment in 3D Computed Tomography (CT) analysis remains constrained by a persistent mismatch between optimization…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
