A Multi-phase Method for Computer-Aided Diagnosis in Chest X-Rays Using Convolutional Neural Networks Transfer Learning and Multi-label Classification

Recent advancements in deep learning have enabled the development of convolutional neural network (CNN) architectures, which have proven to be valuable tools in computer-aided diagnosis (CAD) systems. These systems assist radiologists in identifying regions of interest associated with pathologies in chest X-ray images, a diagnostic tool recognized as essential by the World Health Organization (WHO). The WHO highlights that chest X-rays are an accessible and cost-effective method, crucial for evaluating respiratory and thoracic diseases, particularly in resource-limited settings and during global health emergencies. In this study, the Vindr-CXR dataset was used, known for providing labeled chest X-ray images suitable for multi-label classification tasks. The process began with data preparation, where images and labels were grouped in a binary format and split into training and validation sets. Subsequently, pre-trained neural network architectures, such as VGG16, InceptionV3, ResNet50, and EfficientNetB0, were utilized with weights initialized from ImageNet. The initial layers of these architectures were frozen, and dense layers with sigmoid activation were added for multi-label classification. During training, the binary crossentropy loss function and the Adam optimizer were employed. The models were trained for a fixed number of epochs, with validation conducted at the end of each epoch to evaluate metrics such as accuracy and loss. Finally, predictions were generated on the validation set, and key metrics such as the ROC curve, precision, recall, and F1-Score were calculated. The models achieved a promising performance, with an accuracy of 0.72 in detecting thoracic pathologies. These findings highlight the potential of deep learning to enhance diagnostic precision and support clinical decision-making, reaffirming the critical role of chest X-rays.