Vision-to-Report: A Design Blueprint for an Automated Hepatology Reporting System Integrating nnU-Net and Generative Language Models

Domínguez S.A.G.; Iparraguirre M.G.; Silva P.D.R. · Proceedings - 2025 5th International Conference on Information Systems and Software Technologies, ICI2ST 2025 (2025) · DOI: 10.1109/ici2st68465.2025.00028

Resumen

The manual interpretation of hepatic computed tomography (CT) scans entails a high operational workload and diagnostic variability due to its dependence on individual medical expertise. This article presents the architectural design of a multimodal system that automates the generation of preliminary radiological reports through artificial intelligence. The system integrates an advanced automatic segmentation framework (nnU-Net) to identify and delineate hepatic lesions in CT images, together with a large language model specialized in medical context (GPT-4o mini), capable of generating coherent, detailed, and clinically relevant technical reports. The main contribution of this work lies in the effective design of the interface between the computer vision and natural language processing modules, in addition to proposing a validation plan to assess the clinical effectiveness of the system in terms of diagnostic accuracy and reduction of operational times.

Instituciones: Universidad Peruana de Ciencias Aplicadas