핵심 개념
This paper introduces a novel inference-time pseudo-labeling technique to improve the performance of few-shot 3D medical image segmentation models, particularly in addressing the challenge of limited annotated data.
통계
The study reports Dice scores based on 5-fold cross-validation.
Experiments were conducted using three annotated slices (K = 3) for each query volume.
A window size of 7 was found to be optimal for incorporating query prototypes.
Two iterations of the pseudo-labeling process yielded the best results.