GLeVE: Graph-Guided Lesion Grounding with Proposal Verification in 3D CT

2026-05-21

AI Summary

This study addresses the inaccurate lesion localization arising from the semantic-spatial gap between radiology report texts and 3D CT images by proposing a graph-guided, lesion-level alignment framework. The method models lesion descriptions as atomic semantic units and employs a relation-aware graph neural network to infer organ affiliation, attributes, and inter-lesion relationships, thereby generating discriminative lesion queries. A region-level proposal verification mechanism, guided by anatomical priors, enforces one-to-one correspondence between textual descriptions and lesions. Furthermore, an octree-based autoregressive strategy progressively refines lesion boundaries in a hierarchical manner. Experiments on AbdomenAtlas 3.0 demonstrate that the proposed approach significantly outperforms existing baselines, achieving consistent improvements in both lesion segmentation accuracy and localization precision.

Abstract

Grounding radiology report descriptions to 3D CT volumes is essential for verifiable clinical interpretation, yet remains challenging due to the semantic-spatial gap between free-text narratives and volumetric anatomy. Existing report-assisted and vision-language grounding methods typically rely on phrase-level alignment or dense pixel supervision, resulting in limited lesion-wise correspondence and suboptimal localization accuracy. We propose GLeVE, a graph-guided lesion grounding framework with anatomical prior verification and octree-based autoregressive refinement. GLeVE treats each lesion description as an atomic semantic unit and encodes organ attribution, attributes, and inter-lesion relations through relation-aware graph reasoning to produce discriminative lesion-wise queries. Anatomy-aware proposal generation with region-level verification enforces one-to-one text-lesion alignment, while hierarchical octree refinement progressively improves boundary delineation. Experiments on AbdomenAtlas 3.0 demonstrate consistent gains over classical multimodal foundation models and report-supervised baselines in both segmentation accuracy and lesion-level localization.

Problem

• lesion grounding
• 3D CT
• radiology report
• semantic-spatial gap
• clinical interpretation

Innovation

• graph-guided grounding
• anatomical prior verification
• octree-based refinement
• lesion-wise query
• 3D CT localization