Skip to main content
Back to timeline
arXivSource publication:

Brain MRI lesions no longer need real patients: PathoDraw draws pathology into label maps so one model spots stroke and tumors zero-shot

Synopsis

PathoDraw treats pathology as a conditional distribution over healthy anatomy, drawing lesions into the label map before rendering synthetic brain MRI at any contrast, so training needs no real lesion image, lesion mask or patient-derived threshold. On public stroke, glioma and white matter hyperintensity datasets plus archived clinical scans, a single contrast-agnostic model detects pathology zero-shot under a recent unsupervised anomaly detection benchmark protocol, matching or exceeding contrast-specific detectors trained on real scans while keeping healthy anatomy intact.

Source-provided article image: Pathology Randomization for Synthetic Priors and Zero-Shot Detection
arXiv

Interpretation

Brain MRI anomaly detection can now be covered by one model that never sees real pathology, across several diseases and acquisition settings, instead of one detector per contrast or per disease. Earlier synthetic pathology generators required real lesion masks or real pathological scans, or modeled a single disease process; mainstream anomaly detectors are typically contrast specific and need real healthy or patient data. Under a recent unsupervised anomaly detection benchmark protocol, the same model reaches Dice comparable to or above contrast-specific detectors such as FAE and DISYRE on ATLAS-O, BraTS, WMH and PRES, for example 0.537 versus FAE's 0.437 on BraTS and 0.229 versus DISYRE's 0.130 on WMH.

The method draws pathology as a conditional distribution over healthy anatomy into the label map, then renders it through a contrast-randomized renderer, and the network learns only two things: the healthy anatomy beneath the lesion and the signed intensity deviation of the scan from it. The network never predicts a lesion class; a lesion voxel is scored against the host tissue it overwrites, so detection reduces to a single global threshold on the deviation map. Training draws healthy anatomy from 1,033 label maps of healthy subjects whose images never enter training; detection, super-resolution and registration share one 3D UNet backbone with only the task head changed.

Perspective

The demonstrated scope is public stroke, glioma and white matter hyperintensity datasets plus one hospital's archived clinical cohort, scored under a recent unsupervised anomaly detection benchmark protocol, with no real image, lesion mask or patient-derived threshold in training.

Synthetic pathology is randomized over shape, placement and intensity rather than over disease, so detectable conditions must lie on this manifold; real lesions fall within the generator's shape manifold at high recall, but shape enclosure is only necessary, and the residual error lies in appearance.