Teaching a quality-control system to spot MRI artefacts using simulated physics

Any large-scale imaging analysis has an unglamorous prerequisite: knowing which scans are good enough to use. Automating that check is harder than it sounds, because machine learning needs plenty of examples of bad scans to learn from — and bad scans, especially with rarer artefacts, are exactly what nobody has collected and labelled.

This Medical Image Analysis paper, senior-authored by Dr Arman Eshaghi, sidesteps the problem by generating the training data instead of hunting for it. The framework has four parts: a set of artefact generators derived from MR physics that corrupt clean brain scans in controlled, realistic ways; a large pool of abstract and engineered image features designed to capture nine distinct structural-MRI artefacts; a feature-selection step that picks the best features separately for each class of artefact; and a set of support vector machine classifiers trained to identify them.

The result is a semi-supervised quality-control system that never needed a labour-intensive collection of rare real-world failures. For a programme like MS-PINPOINT, built on repurposing heterogeneous hospital MRI archives, automated quality control is not a side concern — it is the gate everything else passes through, and it has to work on scans from hundreds of different scanners.

Read the paper: An efficient semi-supervised quality control system trained using physics-based MRI-artefact generators and adversarial training.