The lack of labeled data is one of the main prohibiting issues on the development of deep learning models, as they rely on large labeled datasets in order to achieve high accuracy in complex tasks. Our objective is to evaluate the performance gain of having additional unlabeled data in the training of a deep learning model when working with medical imaging data. We present a semi-supervised learning algorithm that utilizes a teacher-student paradigm in order to leverage unlabeled data in the classification of chest X-ray images. Using our algorithm on the ChestX-ray14 dataset, we manage to achieve a substantial increase in performance when using small labeled datasets. With our method, a model achieves an AUROC of 0.822 with only 2% labeled data and 0.865 with 5% labeled data, while a fully supervised method achieves an AUROC of 0.807 with 5% labeled data and only 0.845 with 10%.
Semi-supervised Learning for Bone Mineral Density Estimation in Hip X-Ray Images