Teaching a spine-MRI model without hand-drawn outlines
Spine-GFlow segments vertebrae, discs and the spinal canal in lumbar MRI without any manual annotation, matching a fully supervised model (Dice 0.914 vs 0.916).
The question
Most MRI segmentation models learn from outlines drawn by hand. That is tedious, and more so when several structures must be outlined on every scan.
What we did
Spine-GFlow combines what a neural network learns from the image with anatomical prior knowledge. A rule-based method first produces rough seed areas automatically, a proposal step blends multi-scale image features with anatomy, and a training loss refines pixel classification. No manual annotation is used. We tested it on two independent datasets, one with images from three different scanners.
What we found
Without any manual labels, segmentation of vertebral bodies, discs and the spinal canal was comparable to a fully supervised model (mean Dice score, a measure of overlap with the expert outline where 1 is perfect: 0.914 vs 0.916), and it held up across scanner settings.
Why it matters
Removing the annotation bottleneck makes it far easier to build spine-imaging tools for new hospitals and scanners.