Résumé:
Radiological reports play an essential role in the diagnostic process, particularly for thoracic pathologies visible on chest X-rays. The manual preparation of these reports by radiologists is a demanding, time-consuming task that is subject to subjective variations. The emergence of deep learning offers promising prospects for automating this task and improving clinical productivity. In this work, we propose an automatic radiology report generation system based on a hybrid deep learning architecture. The system integrates a pre-trained convolutional neural network (EfficientNetB0) for visual feature extraction, coupled with a Transformer-based decoder for diagnostic text generation. The model is trained on the Indiana University Chest X-ray database, after structured pre-processing of the images and text reports. In order to improve the linguistic consistency and terminological accuracy of the reports generated, a post-processing phase is introduced, based on the BioGPT model, which specialises in the biomedical field. This step improves the fluidity, readability and clinical accuracy of the reports produced. The experimental results obtained demonstrate the effectiveness of the system. The BLEU-4 score increased from 0.4191 (LSTM model) to 0.8286 (Transformer model with BioGPT), while the BERTScore reached 0.9628, reflecting strong semantic similarity with the reference reports. These performances confirm the potential of the proposed approach to assist radiologists and improve the quality of AI-assisted diagnoses.
