A Clinically-Grounded Two-Stage Framework for Renal CT Report Generation

Abstract: Generating radiology reports from CT scans remains a complex task due to the nuanced nature of medical imaging and the variability in clinical documentation. In this study, we propose a two-stage framework for generating renal radiology reports from 2D CT slices. First, we extract structured abnormality features using a multi-task learning model trained to identify lesion attributes such as location, size, enhancement, and attenuation. These extracted features are subsequently combined with the corresponding CT image and fed into a fine-tuned vision-LLM to generate natural language report sentences aligned with clinical findings. We conduct experiments on a curated dataset of renal CT studies with manually annotated sentence-slice-feature triplets and evaluate performance using both classification metrics and natural language generation metrics. Our results demonstrate that the proposed model outperforms random baselines across all abnormality types, and the generated reports capture key clinical content with reasonable textual accuracy. This exploratory work highlights the feasibility of modular, feature-informed report generation for renal imaging. Future efforts will focus on extending this pipeline to 3D CT volumes and further improving clinical fidelity in multimodal medical AI systems.