Pseudo-Labeling With Contrastive Perturbation Using CNN & ViT for Chest X-ray Classification

Thanawit Gerdprasert, Shingo Mabu

Abstract: Deep learning has revolutionized artificial intelligence, making its performance acceptable for building Computer-Aided Diagnosis systems in medical images. However, the performance of deep learning heavily relies on the availability of a substantial amount of data during the training step. While the advancement of big data has alleviated this problem across many domains, some domains relying on experts, such as the medical domain, still face the challenge of acquiring substantial amounts of training data which requires essential human resources from the expert to label them manually. To overcome this problem, semi-supervised learning has emerged as one of the promising solutions. Semi-supervised learning operates on the principle of utilizing unlabeled data to enhance performance, eliminating the need for data to be fully labeled before utilization. In this paper, we propose a framework to improve Chest X-ray image classification using pseudo-labeling with consistency regularization in the semi-supervised learning environment. We adopted two deep learning models, a convolutional neural network and a vision transformer, to perform pseudo-labeled tasks on augmented data generated by the perturbation of unlabeled samples. We evaluated our method using the Chest X-ray COVID-19 image classification task, and the accuracy of the model increased by 3.50% in the scenarios where 30% of labeled samples and 60% of unlabeled samples were used for training the model.