Anatomy and Image Contrast Metadata Verification using Self-supervised Pretraining

Automated metadata verification is essential for data quality checking. This study evaluates the extent to which self-supervised pretraining can improve performance on the anatomy and image contrast metadata verification tasks. On a small but diverse dataset, pretraining coupled with supervised finetuning, outperforms training from an ImageNet initialization, suggesting improved near out-of-distribution performance. On a larger brain-only dataset, training a linear classifier on the self-supervised pretrained network embeddings outperforms the corresponding ImageNet pretraining or random initialization on the image contrast prediction task. Cross-checking predictions against the DICOM metadata was an effective method for detecting artifacts and other quality issues.