CheXDouble: Dual-Supervised interpretable disease diagnosis model

Chest X-ray imaging, commonly used for diagnosing cardiopulmonary diseases, typically requires radiologists to devote considerable effort to reading and interpreting the images. Moreover, diagnostic outcomes can vary due to differences in radiologists’ experience. Deep learning for chest X-ray disease diagnosis holds great promise for enhancing diagnostic accuracy and reducing the workload of radiologists. However, traditional deep learning models for medical image classification are often difficult to interpret. To address this, we introduce the Global Attention Alignment Module, which utilizes cardiopulmonary mask for supervised training. This provides the model with spatial location priors during training, thereby enhancing the interpretability of the saliency maps and the disease classification performance. Additionally, most chest X-ray datasets suffer from severe imbalances between positive and negative samples for diseases, leading to classification imbalance issues when training models. Thus, we propose the Improved Focal Loss, which dynamically adjusts the weight of negative samples in the loss function based on sample statistics, effectively mitigating the imbalance issue in the dataset. Moreover, the training of deep learning models for medical image classification requires substantial data support. Therefore, we conducted a quantitative analysis to explore the impact of five different data augmentation methods on model classification performance across various input image sizes, identifying the most effective data augmentation strategy. Ultimately, through these proposed methods, we developed the dual-supervised medical imaging disease diagnosis model CheXDouble, which surpasses previous state-of-the-art models with its highly competitive disease classification performance.