Résumé:
Automated medical report generation from chest radiographs has emerged as a criti- cal challenge in medical imaging, particularly in addressing information bottlenecks and poor clinical accuracy for rare pathological conditions. This work presents ChestBXG, a novel multi-modal architecture that integrates classification-guided visual encoding with domain-adaptive language generation to bridge the semantic gap between radiographic features and clinical text. Our approach employs Efficient Net-B4 for visual feature extrac- tion, coupled with BioGPT for medical domain-specific text generation, interconnected through sophisticated co-attention mechanisms that prevent information loss during cross- modal alignment. The architecture incorporates a confidence-based classification head that guides report generation, particularly enhancing performance on minority patho- logical cases. Experimental evaluation on a curated subset of the MIMIC-CXR dataset demonstrates substantial improvements across standard metrics, achieving decent results on multiple metrics while focusing on harder samples. The proposed framework addresses fundamental limitations in existing methodologies while maintaining computational effi- ciency, establishing a foundation for clinically viable automated reporting systems that enhance diagnostic accuracy and workflow efficiency in radiological practice.
Keywords: X-ray images, Vision feature extraction, Language generation, Encoder- Decoder
