Title: RadAgent: An AI Agent Utilizing Tools for Sequential Analysis of Chest Computed Tomography

Abstract: Vision-language models (VLMs) have significantly propelled the advancement of artificial intelligence in interpreting and generating reports for complex medical imaging modalities, including computed tomography (CT). However, current approaches largely position clinicians as passive recipients of final results, failing to provide an interpretable reasoning pathway that allows for inspection, validation, or refinement. To resolve this limitation, we present RadAgent, an AI agent equipped with tool-use capabilities that produces CT reports via a stepwise and transparent methodology. Every generated report is supported by a completely inspectable log of intermediate decisions and tool interactions, enabling healthcare professionals to scrutinize the derivation of reported findings. Experimental results indicate that RadAgent enhances chest CT report generation compared to its 3D VLM equivalent, CT-Chat, across three key metrics. Specifically, clinical accuracy saw improvements of 5.8 points (a 35.4% relative increase) in macro-F1 and 5.1 points (an 18.6% relative increase) in micro-F1. Additionally, robustness against adversarial conditions increased by 24.7 points (41.9% relative). RadAgent also attained a faithfulness score of 37.0%, a capability not present in its 3D VLM counterpart. By framing the interpretation of chest CT as an explicit, iterative reasoning process augmented by tools, RadAgent advances the field toward more transparent and trustworthy AI solutions in radiology.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC