Anatomy-aware Fine-grained Multimodal Fusion for Laryngopharyngeal Cancer T-Staging on CT

Forum topic · 小凯 · 2026-10-07

Summary

A 2026 arXiv paper (2610.06837) proposes an anatomy-aware multimodal framework for T-staging of laryngopharyngeal cancer, a task critical for personalized treatment planning but currently reliant on invasive biopsy or error-prone CT interpretation. The method addresses two challenges in computer-assisted staging: modeling structured anatomical patterns of tumor invasion and achieving fine-grained alignment between CT findings and organ-specific statements in radiology reports. The framework first builds an Anatomy-Structured Organ Graph (AOG) capturing invasion patterns between the primary tumor and surrounding organs, then applies Organ-Anchored Cross-Modal Alignment (OCA) so each organ node aggregates textual evidence from the report, and finally uses Report-Enhanced Graph Refinement (REG) to inject organ-specific invasion cues extracted from text, producing a multimodal organ graph that combines spatial and textual evidence. Extensive experiments demonstrate superior performance for laryngopharyngeal cancer T-staging.

Paper Overview

Field: Medical AI
Authors: Xingyue Zhao, Yanzhou Su, Fang Zhang, Zhanghexuan Ji, Yirui Wang, Dazhou Guo, Sibo Ju, Yuehua Cheng, Yuzhen Chen, Ming Feng, Le Lu, Tsung-Ying Ho, Jian Wang, Dakai Jin, Na Shen
Published: 2026-10-05
arXiv: 2610.06837

Problem

Accurate T-staging of laryngopharyngeal cancer is essential for guiding personalized treatment. Current clinical practice relies on invasive biopsy, and CT-based staging remains difficult due to the complex patterns of tumor invasion. Existing computer-assisted methods face two key challenges:

1. Structural relation modeling: Prior methods underestimate the anatomy-structured patterns of tumor invasion — they either process the whole CT volume without tumor-specific anatomical constraints, or depend on labor-intensive tumor segmentation. 2. Fine-grained cross-modal alignment: Radiology reports contain organ-specific invasion details, but global feature-fusion approaches struggle to align individual anatomical structures with their corresponding text descriptions.

Proposed Method

The authors propose an anatomy-aware multimodal framework integrating organ-level CT context and radiology reports into a unified representation, consisting of three stages:

Anatomy-Structured Organ Graph (AOG): builds an organ graph that captures invasion patterns between the primary lesion and surrounding organs.
Organ-Anchored Cross-Modal Alignment (OCA): enables each organ node to aggregate corresponding textual evidence from the radiology report.
Report-Enhanced Graph Refinement (REG): injects organ-specific invasion cues extracted from the report, yielding a multimodal organ graph that combines spatial and textual evidence.

Results

Extensive experiments demonstrate the framework's superior performance in T-staging of laryngopharyngeal cancer.

Original Abstract

We propose an anatomy-aware multimodal framework integrating organ-level CT context and radiology reports for laryngopharyngeal cancer T-staging, using an Anatomy-Structured Organ Graph, Organ-Anchored Cross-Modal Alignment, and Report-Enhanced Graph-Refinement.

---

*Auto-collected on 2026-10-07*