Redes Neurais Convolucionais Explicáveis para detecção de doenças pulmonares utilizando imagens de raios-x do tórax

Abstract

This work presents the development of a Deep Learning model for the automated diagnosis of multiple pathologies in chest X-rays, using explainability techniques such as Grad-CAM++ and LIME. The goal is not only to identify diseases in medical images but also to provide a visual explanation that helps validate the model's predictions, fostering greater confidence in its use in clinical environments. The study employed the NIH Chest X-ray Dataset, one of the largest publicly available chest X-ray databases, containing over 100,000 images labeled with multiple pathologies. The dataset was reduced and balanced to ensure a more equitable distribution of disease classes and better representativity during model training. The DenseNet-121 architecture with transfer learning was utilized, chosen for its efficiency in image classification tasks and its extensive use in the medical literature for diagnosing pathologies. Additionally, techniques like dropout were applied to reduce overfitting, and the model was evaluated using standard performance metrics such as AUC-ROC and Hamming Loss. The validation process included an essential explainability step, where we used Grad-CAM++ to generate heatmaps highlighting the regions of the image that contributed the most to the diagnosis, and LIME to identify the most relevant features in the model's decision-making process. This approach was validated by healthcare professionals who evaluated the model's predictions and the visual explanations generated, providing both qualitative and quantitative feedback. The analysis of the results demonstrated that the model showed good performance in diseases such as Cardiomegaly and Edema, with AUC-ROC values above 0.85. More challenging conditions, like Infiltration, presented more modest values, reflecting the inherent complexity of diagnosing these diseases. The results suggest that the use of Deep Learning models, combined with explainability techniques, can be a valuable tool in clinical practice, especially by providing greater transparency in predictions, thus facilitating the adoption of these technologies by healthcare professionals. However, there are limitations, such as the low number of specialists involved in the validation and the challenges in accurately identifying multiple concomitant diseases. Future work could include broader validation with additional specialists, as well as exploring new network architectures and more robust explainability techniques.