Addressing data imbalance in Vietnamese chest X-ray diagnosis using deep neural networks

Abstract

Pulmonary diseases such as pneumonia, tuberculosis, and particularly lung cancer represent serious public health concerns, necessitating early and accurate detection methods, in which chest X-ray classification plays a pivotal role. However, an inherent challenge in medical datasets is the issue of class imbalance, where rare but critical pathologies often have significantly fewer samples compared to normal cases or more common conditions. This study systematically proposes and evaluates a deep learning–based approach for automatic chest X-ray classification, with a focus on addressing data imbalance to improve the detection of minority classes. The approach involves data normalization, the application of appropriate data augmentation techniques, and loss function reweighting through class weighting. We conducted experiments and performance comparisons using state-of-the-art convolutional neural network (CNN) architectures, including DenseNet-121, ResNet-50, EfficientNet-B0, and MobileNet-V3 Small, on two chest X-ray datasets: a publicly available dataset from Kaggle and the Vietnam VRPACs dataset. Experimental results demonstrate that DenseNet-121, when combined with imbalance-handling techniques, achieved the highest balanced accuracy (BACC) of 0.85, indicating a substantial improvement in minority-class classification performance compared with methods without imbalance handling. This study provides a potential solution and a scientific foundation for the development and deployment of automated diagnostic support systems in healthcare facilities, particularly in Vietnam.