Inteligência artificial aplicada à radiografia de tórax: comparação de modelos de redes neurais convolucionais na detecção de opacidades.

Abstract

Introduction: Artificial intelligence, particularly through convolutional neural networks (CNNs), has shown increasing applicability in the automated analysis of medical images and may assist in identifying abnormalities on chest radiographs. Objective: To develop and compare the performance of different convolutional neural network architectures for classifying chest radiographs into Normal, Opacity, and Other Abnormalities categories, with emphasis on opacity detection. Methods: This retrospective, analytical, observational study analyzed 3,349 chest radiographs from patients aged 18 years or older. Three architectures were evaluated: EfficientNetB0 and MobileNetV2 using transfer learning, and a CNN developed and trained from scratch. Model performance was assessed using accuracy, precision, sensitivity (recall), F1-score, area under the receiver operating characteristic curve (AUC), and confusion matrices. Results: EfficientNetB0 achieved the best overall performance, with an accuracy of 85.9% and a macro AUC of approximately 0.96. Specifically for opacity detection, it achieved 84.6% precision, 82.2% sensitivity, and an F1-score of 83.4%. The CNN trained from scratch achieved an overall accuracy of 72.3%, with 83.5% precision, 57.6% sensitivity, and an F1-score of 68.2% for opacity detection, as well as evidence of overfitting. MobileNetV2 showed lower overall performance, with an accuracy of approximately 67%. Conclusion: EfficientNetB0 combined with transfer learning outperformed the other evaluated architectures, particularly in detecting opacities on chest radiographs. These findings demonstrate the potential of CNNs as auxiliary tools for radiographic analysis; however, external validation, greater image diversity, and improved model interpretability are required before clinical implementation.