A Two-Stage Contrastive Learning Framework Grounded in Label-Specific Features for Low-Frequency Labels in Chest X-Ray Multi-Label Classification.

Thoracic diseases represent a significant threat to human health. Chest X-ray imaging, owing to its cost-effectiveness and rapid imaging capabilities, has been widely adopted as a primary diagnostic tool in clinical practice. However, existing models are often susceptible to imbalances in disease label distributions. This study proposes a dual-phase convolutional neural network for the classification of thoracic diseases. In the first phase, matrix operations are employed to extract discriminative features corresponding to each disease label, effectively shifting the classification task from the image domain to a label-specific feature domain. The second phase incorporates feature contrastive loss and feature updating mechanisms to further enhance the model's generalization capability. The proposed framework was evaluated on three public datasets (CheXpert, REFLACX, and EGD) to verify its consistent performance across diverse data sources. Experimental results demonstrate that our model achieved an AUC of 0.8296, AUPRC of 0.2969, Precision of 0.3943, and F1-score of 0.3301 on our dataset, outperforming existing chest X-ray classification models. These findings indicate that our proposed framework effectively learns label-specific characteristics and captures intrinsic image features associated with each disease label, offering an advanced technical tool for the diagnosis of thoracic diseases.