Multi-label Convolutional Neural Network for Multiple Diseases Detection in Chest Radiographs

Identifying multiple diseases from a single chest radiograph is a complex and time-intensive task that requires considerable expertise. Traditional diagnostic methods rely heavily on skilled human interpretation, which can be tedious and prone to variability. Although deep learning techniques have facilitated the development of automated systems, these systems are typically limited to detecting individual pathologies. Chest radiographs, essential for diagnosing diseases such as pneumonia, lung cancer, and various abnormalities like lesions and fractures, contain substantial information that can be difficult to interpret accurately even for expert radiologists. This chapter introduces a multi-label disease diagnosis framework capable of detecting multiple pathologies concurrently. Utilizing a multi-label classification framework, this work explores two initialization strategies: transfer learning with pre-trained weights and training from scratch with random initialization. The dataset used is the NIH Chest X-ray14 dataset, comprising 112,120 chest X-rays. The results demonstrate that the Xception framework, particularly with transfer learning, outperforms the other models, achieving the highest AUC score of 0.826 and an overall test accuracy of 0.878. This study highlights the practical significance of multi-label AI algorithms in enhancing the diagnostic accuracy and efficiency of radiologists, contributing to the development of reliable and trustworthy automated diagnostic systems in the medical field.