Deep learning has made significant strides in medical imaging, leveraging the use of large datasets to improve diagnostics and prognostics. However, large datasets often come with inherent errors through subject selection and acquisition. In this paper, we investigate the use of Diffusion Autoencoder DAE embeddings for uncovering and understanding data characteristics and biases, including biases for protected variables like sex and data abnormalities indicative of unwanted protocol variations.
Evaluations on a large-scale dataset consisting of sagittal T2-weighted MR images of three spine regions show that DAE embeddings effectively separate protected variables such as sex and age. Furthermore, we used t-SNE visualization to identify unwanted variations in imaging protocols, revealing differences in head positioning.
Our embedding can identify samples where a sex predictor will have issues learning the correct sex. Our findings highlight the potential of using advanced embedding techniques like DAEs to detect data quality issues and biases in medical imaging datasets. Identifying such hidden relations can enhance the reliability and fairness of deep learning models in healthcare applications, ultimately improving patient care and outcomes. Historically, deep learning in medicine has relied on large datasets, which have been difficult to obtain.
However, in recent years, significant progress has been made. Notable examples include extensive datasets such as chest radiographs [ 10 , 11 ] , as well as images from comprehensive epidemiological studies like the UK-Biobank [ 1 ] and the German National Cohort "NAKO Gesundheitsstudie" [ 2 ]. Nevertheless, the sheer size of these datasets makes it difficult to detect data shifts and biases visually. Biases can occur through various sources, such as deviations from the examination protocol, the subjects themselves, changes in framing, or data processing errors [ 5 ].
Detecting biases in large data sets is a laborious and time-consuming task and would require a large survey to strongly evaluate the data [ 9 , 13 ]. For disease prediction, it is paramount to know what biases exist to compensate for them or at least observe them if they have an influence on a classifier.
