Volumetric chest CT report generation via state-space visual prompting and clinical-fact alignment
Abstract
Objective. To address the memory bottlenecks of three-dimensional (3D) sequence modeling and the factual hallucinations associated with cross-entropy optimization in automated 3D chest computed tomography (CT) report generation. Approach. We propose an end-to-end framework that couples perception refinement with clinical-fact alignment. First, a hybrid perception module integrates spatial visual encoding with a bidirectional state-space model to capture global anatomical topology at linear computational complexity. Second, a lesion-aware visual prompting module extracts key pathological signals through cross-view synergy and adaptive residual gating. Finally, the text decoder is fine-tuned with the group relative policy optimization algorithm under an asymmetric composite reward, reducing its reliance on normal templates. Main results. Experiments on the CTRG-Chest-548K dataset show that the proposed method achieves competitive natural language generation performance and, more importantly, a substantial improvement in clinical accuracy. The RadGraph F1 score increases from 41.89 to 44.63 relative to the strongest baseline, indicating reduced factual hallucination. Significance. Combined with a 71% reduction in peak inference memory, these results indicate that the framework is a practical step toward deployable and clinically reliable automated CT interpretation.