Banca de QUALIFICAÇÃO: HERICLYS S BORGES
Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.DISCENTE: HERICLYS S BORGES
DATA: 22/05/2026
HORA: 10:00
LOCAL: Google Meet
TÍTULO: Integration of Clinical Context and Chest X-ray Images for Automatic Radiology Report Generation
PALAVRAS-CHAVES: Automatic Radiology Report Generation; Chest X-ray Analysis; Multimodal Deep Learning; Transformer-based Medical Imaging
PÁGINAS: 12
GRANDE ÁREA: Engenharias
ÁREA: Engenharia Elétrica
RESUMO:
Automatic radiology report generation from chest X-rays has attracted increasing attention as a way to assist clinicians and reduce the workload associated with medical image interpretation. Despite recent advances in multimodal learning, generating clinically coherent and factually consistent reports remains a challenging task. This paper presents a multimodal transformer-based framework for automatic chest X-ray report generation that integrates radiographic images and patient clinical history. Visual representations are extracted from frontal and lateral chest X-ray images using a ResNet-50 backbone with progressive fine-tuning. Clinical context is encoded using Bio_ClinicalBERT, allowing the model to incorporate domain-specific medical knowledge. These multimodal representations are fused and processed by a Transformer encoder–decoder architecture that generates radiology reports autoregressively. Experiments conducted on the MIMIC-CXR dataset demonstrate that the proposed model can produce structured radiology reports that capture clinically relevant findings. The model achieves a BLEU-4 score of 0.089, ROUGE-L of 0.263, and a METEOR score of 0.282, indicating strong semantic similarity between generated and reference reports. These results suggest that integrating clinical context with visual features is a promising direction for improving automated radiology report generation systems.
MEMBROS DA BANCA:
Presidente - 2025063 - ROMUERE RODRIGUES VELOSO E SILVA
Interno - 1126212 - ANTONIO OSEAS DE CARVALHO FILHO
Interno - 3420215 - LEONARDO PIO VASCONCELOS