Using the Mahalanobis Distance as a Dataset Comparison Metric for Semi-Supervised Detection of COVID-19 in Chest X-rays

Deep learning methods rely on large labeled datasets. However, in real-world applications such as medical imaging, only scarcely labeled datasets might be available. Semi-supervised deep learning leverages both labeled and unlabeled datasets to improve the model's accuracy. However, semi-supervised methodologies frequently assume similar distributions of the labeled and unlabeled dataset. In real-world settings, this assumption might be violated. This gives rise to the need of enforcing data quality attributes such as the consistency between labeled and unlabeled data sources. In this work, we propose the usage of a Mahalanobis-based distance in the embedding space of the deep learning model to pre-assess labeled to unlabeled dataset affinity in a semi-supervised setting. We test how this distance can highly correlate with the accuracy of a popular semi-supervised method known as MixMatch. Moreover, the proposed method is considerably more efficient than previous work in the field.