Data quality metrics for unlabelled datasets in medical imaging

Abstract

Deep learning models typically require large, labeled datasets for optimal performance. However, in real-world applications such as medical imaging, labeled data can be scarce. Semi-supervised deep learning addresses this challenge by leveraging both labeled and unlabeled data to enhance model accuracy. Most semi-supervised methods assume similar distributions between labeled and unlabeled datasets, an assumption that may not hold in practice. To ensure data quality and consistency, we introduce Mahalanobis-based and Frobenius-based distance measures in the embedding space of the deep learning model to evaluate the similarity between labeled and unlabeled datasets. Our findings reveal that the Mahalanobis-based distance correlates strongly with the accuracy of the popular semi-supervised method MixMatch, whereas Frobenius distance results show inconsistent behavior. Moreover, the proposed approach is significantly more efficient than existing methods in the field.