An Evidence-Constrained LLM Framework for Automated Abdominal CTA Aneurysm Analysis and Reporting.
Authors
Abstract
Abdominal computed tomography angiography (CTA) aneurysm analysis requires simultaneous localization, measurement, and anatomical attribution of multiple candidate lesions across a continuous three-dimensional vascular network, followed by translation of these findings into clinically reviewable CTA report drafts. Existing automated approaches largely focus on segmentation, measurement, or candidate detection, whereas direct large language model (LLM) interpretation of CTA images remains vulnerable to factual errors, hallucinations, and safety concerns. We propose an evidence-constrained large language model-assisted computer-aided diagnosis (LLM-CAD) framework that decouples evidence generation from language generation. On the image side, a case-level structured evidence pack is constructed via three-seed vascular tree construction, hierarchical candidate detection, and probabilistic host-vessel attribution, with measurements, host-vessel attribution confidence, and uncertainty cues encoded as separable, traceable fields. On the language side, three categories of hard constraints, namely measurement preservation, evidence priority, and uncertainty explicitness, confine LLM outputs within the evidence pack during both report generation and interactive question answering (QA). Experiments on 105 abdominal CTA cases demonstrate that the proposed framework converts complex three-dimensional vascular evidence into clinically reviewable CTA report drafts. Report-quality assessment, interactive QA, cross-backend LLM comparison, and ablation experiments jointly confirm that the structured evidence pack and the hard constraints play complementary, non-substitutable roles. This work offers an evidence-language decoupling paradigm for applying LLMs to complex three-dimensional medical imaging.