With the rapid increase in the volume of medical imaging data, self-supervised learning has shown significant potential in the intelligent diagnosis of chest X-ray images. This paper proposes a self-supervised pre-training method based on the SimCLR framework. It uses the large-scale ChestX-ray14 dataset for unsupervised pre-training and then performs transfer fine-tuning on the multi-label classification task. Specifically, the paper adopt EfficientNet-B0 as the encoder structure and optimize the network's feature extraction ability through contrastive learning. For the downstream multi-label classification task, the paper uses a fully connected layer for output, Sigmoid activation, and combine the Focal Loss function to handle the problem of class imbalance. The experimental results show that compared with the model without pre-training, the pre-trained model has an approximately 2.51% increase in the average AUC value and an approximately 3.04% increase in the F1 value on the validation set. In addition, the paper used the Grad-CAM visualization technique to analyze the regions of interest of the model and conducted in-depth analysis on the categories with a high error rate. The experiments also included dynamic threshold optimization, training curve plotting, and performance comparison across other categories to confirm the method's efficacy.
Kaiwen Luo (Wed,) studied this question.