Gradual Learning for One-Shot Segmentation of Slice Stacks

Johann Christopher Engster, Nele Blum, Thorsten Buzug, Maik Stille

Keywords: Segmentation, Semi-Supervised Learning, One-Shot Learning

Abstract: Diagnostic imaging modalities like magnetic resonance imaging (MRI) or computed tomography (CT) are crucial for medical and industrial inspection. However, labeled datasets are not always available for segmentation of rare cancer types or other defects. Therefore, a new training strategy named gradual learning is proposed for one-shot segmentation, thus requiring only one labeled example slice. A segmentation network trained on this input generates suitable pseudo labels in a local neighborhood, with the quality degrading with distance. These adjacent pseudo labels can be incorporated into the training process repeatedly, to process the unlabeled slices step-by-step. Experiments were conducted on MRI head scans for skull-stripping. A total of 30 models were trained using gradual learning, receiving one scan with one annotated slice each. On a separate test set ($n=30$ scans), the mean intersection over union (mIoU), averaged over all models, increased from 0.885 to 0.935 using gradual learning compared to training without it. When trained with the ground truth (GT) of the same slices instead the models achieved a 0.955 mIoU.

Submission Number: 23