ClinCrossT5: Design and Deployment of a Retrieval-Augmented System for Automated Chest X-ray Report Generation

Abstract

Automated radiology report generation can improve reporting efficiency, but implementing research models is challenging. In this paper, we introduce ClinCrossT5, an end-to-end deployment framework for automated chest X-ray report generation, in the form of an interactive web application. Our system combines Vision Transformer (ViT-B/16) for image feature extraction, retrieval of visually similar prior cases using cosine similarity, and Clinical-T5 for language generation to produce structured Findings and Impression sections from chest radiographs and clinical indications.
The application was built using Gradio and deployed on Hugging Face Spaces. Users can upload chest X-ray images, enter clinical indications, generate structured radiology reports and export the results as Microsoft Word (.docx) documents. This deployment integrates image preprocessing, visual feature extraction, similarity retrieval, multimodal inference, and automated document generation into a browser-accessible interface.
Deployment experiments verified the successful end-to-end execution of the complete inference pipeline that included image preprocessing, feature extraction, case retrieval, report generation, visualization and document export. Latency evaluation on the production deployment using an NVIDIA T4 GPU showed a mean warm-inference latency of 6.07 s for repeated requests to a single image (n = 9) and 8.98 s across 20 chest X-ray images with three repeated runs per image (n = 60). We show representative examples generated by the deployed application on unseen chest X-ray images. The paper further presents the deployment architecture, implementation workflow and operational evaluation in terms of latency, providing a practical platform for research and decision-support evaluation of retrieval-augmented multi-modal radiology report generation systems.