Abstract
Deep learning techniques have provided new research methods for computer-aided diagnosis, allowing researchers to use deep learning methods to process medical imaging data. Chest X-ray examinations are widely used as a primary screening method for chest diseases. Therefore, it is of great importance to study diagnosis of 14 common pathologies in chest X-ray images using deep learning methods. In this paper, we propose a deep learning model named AM_DenseNet for chest X-ray image classification. The model adopts a dense connection network and adds an attention module after each dense block to optimize the model’s ability to extract features, and finally a Focal Loss function is applied to solve the data imbalance problem. The experiments used chest X-ray images as model input and were trained to output the probabilities of 14 chest pathologies. The Area under the ROC curve (AUC) was used to measure the classification results, and the final average AUC was 0.8537. The experimental results show that the AM_DenseNet model could complete the pathology classification of the chest X-ray images effectively.
Highlights
Thorax diseases have severely threatened the health of human beings
We evaluate the classification performance of the AM_DenseNet network model for 14 chest diseases using the Receiver Operating Characteristic (ROC) curve and the Area under the ROC curve (AUC) score
The results show that our experiments outperform the other three experiments in 7 categories of pathologies and have the highest average AUC values, and the AUC value of Hernia is slightly lower than ChestNet
Summary
Thorax diseases have severely threatened the health of human beings. Among them, the pneumonia alone affects approximately 450 million people (i.e., 7% of the world population) and results in about 4 million deaths per year[1]. By adding a CBAM attention module to the DenseNet-121 network model, the training of the deep neural network is focused on disease-relevant regions, adaptively assigning more weight to the learned features in the relevant regions, thereby increasing the model's ability to extract major features that are truly useful. In the chest X-ray datasets, the number of normal samples is much larger than the number of diseased samples, and there are many types of chest diseases, so it is difficult to achieve the ideal training results using BCE loss On this basis, an improved algorithm for BCE loss, named Focal loss[11], was proposed, which introduces a weighting factor α, the value of which is between 0 and 1. It is verified that the AM_DenseNet model can effectively handle the multi-classification problem of chest pathology
Published Version (Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have