SegReg-Rep: Region-Aware Vision-Language Alignment for Fine-Grained Radiology Report Generation from 3D Medical Images

Radiology report generation from 3D CT scans remains a challenging task due to the high dimensionality of volumetric data and the sparse, variable structure of clinical language. Existing methods often rely on global image-text alignment, which struggles to capture fine-grained anatomical semantics and localized abnormalities critical for clinical reasoning. In this work, we present SegReg-Rep, a structured and region-aware vision-language framework designed to bridge 3D medical imaging and radiology reports with spatial and semantic precision. Firstly, each radiology report is decomposed into structured region-description pairs using domain-adapted named entity recognition. A prompt-guided 3D segmentation model is then employed to localize the anatomical regions referenced in the report, producing sub-volumes that are explicitly aligned with corresponding text descriptions. These region-text pairs are leveraged to train a CLIP-style contrastive vision-language model with segmentation-based enhancement region. To enable fine-grained and clinically interpretable report generation, we directly integrate the pretrained region-level vision encoder into a large medical language model (Qwen2.5-3B-Instruct). By conditioning the language model on spatially grounded visual features through prompt-based mechanisms, our decoupled architecture facilitates accurate, modular, and deployment-ready synthesis of radiology reports. Experimental results on the CT-RATE dataset demonstrate that SegReg-Rep achieves superior report generation performance, enhanced anatomical interpretability compared to existing baselines. Our framework offers a scalable, modular solution for clinical report automation, with improved computational efficiency and better deployment practicality compared with standard pipelines. The codes and models are available at https://github.com/BlueRainbow2000/SegReg-Rep.