Classificação de pneumonia em radiografias pediátricas com deep learning e inteligência artificial explicável: uma análise comparativa entre as arquiteturas ResNet50 e YOLOv11s-cls.

Abstract

Pneumonia is one of the leading causes of child morbidity and mortality worldwide, and early diagnosis through chest X-rays is essential to reduce clinical complications. In this context, the use of deep learning techniques has shown promise as a tool to support medical diagnosis. This work proposes a comparative analysis between the YOLOv11s-cls and ResNet50 models applied to binary pneumonia classification in pediatric chest X-rays, using the Pediatric Pneumonia Chest X-ray Dataset, composed of 5,856 images divided into Normal and Pneumonia classes. Both models were trained with weights pre-trained on the ImageNet dataset using a two-phase progressive transfer learning strategy, combined with data augmentation techniques and class weight balancing to mitigate class imbalance. Performance evaluation was conducted using metrics such as accuracy, precision, recall, F1-score, and AUC-ROC. Additionally, the Explainable Artificial Intelligence technique Gradient-weighted Class Activation Mapping was applied to investigate model explainability, identifying the X-ray regions that most influenced predictions. The results demonstrated that both models achieved high performance, with AUC-ROC values above 0.99. ResNet50 obtained superior performance across all evaluated metrics.