Guideline-Informed MLLM Reasoning for Pathology-Aware Post-Operative Prostate CTV Segmentation

Yinhao Wu, Hengrui Zhao, Haiqing Li, Wenliang Zhong, Hehuan Ma, Yuzhi Guo, Dan Nguyen, Daniel X. Yang, Steve B. Jiang, Junzhou Huang

Keywords: Clinical Target Volume, Radiotherapy Planning, 3D Image Segmentation.

TL;DR: Multimodal Reasoning for CTV Segmentation

Abstract: Accurate segmentation of the Clinical Target Volume (CTV) is a critical prerequisite for precise radiotherapy planning, pursuing complete irradiation of microscopic disease while minimizing toxicity to surrounding healthy organs. However, achieving automated CTV segmentation remains highly challenging due to the invisible microscopic disease on planning CT and the necessity of incorporating clinical context into delineation decisions. Unlike previous methods that rely solely on visual features or coarse global text reasoning, we propose ReaCT, a unified framework that reformulates CTV segmentation as a multimodal reasoning task by explicitly integrating pathological information with visual context. Specifically, we introduce a Guideline-Informed LLM Agent designed to emulate the information-retrieval workflow of radiation oncologists. By distilling knowledge from authoritative clinical guidelines, this agent filters and structures lengthy pathology reports into a concise set of determinative pathological attributes, effectively bridging the semantic gap between unstructured clinical records and segmentation networks. Furthermore, we develop an Attribute-Specific MLLM Reasoner built upon a 3D residual U-Net that performs fine-grained spatial reasoning. By leveraging a sequence of attribute-specific query tokens, the model disentangles the distinct target implications of individual pathological attributes, enabling fine-grained anatomical alignment via multi-scale fusion using Two-Way Transformers. Experiments on a postoperative prostate cancer dataset demonstrate that ReaCT achieves state-of-the-art segmentation performance and exhibits strong robustness, with pronounced improvements under limited-annotation settings.

Primary Subject Area: Segmentation

Secondary Subject Area: Foundation Models

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 80