Self-supervised Modelling of Neuroimaging Data

This teaching session set out a four-day project on detecting synthetic motion artefacts in brain MRI. Philipp Goebl and Arman Eshaghi framed it as a direct comparison between a fine-tuned self-supervised classifier and a supervised baseline.

The question

Does self-supervised pretraining improve motion-artefact classification when compared with a supervised model on the same task?

Deck outline

• Classifying brain MRI as good or bad, with uncertainty
• Generating ground-truth examples of synthetic motion artefacts
• Self-supervised pretraining with MONAI
• Fine-tuning the pretrained model for classification
• Training a supervised comparison model
• Comparing performance and presenting the results

Take-away

The exercise treats self-supervision as a testable modelling choice: keep the task fixed, then compare performance and uncertainty.