Core Concepts
Proposing a three-stage framework, CFR, leveraging SAM for semi-supervised 3D medical image segmentation, achieving significant improvements in annotation efficiency and performance.
Abstract
The Content introduces the Concatenate, Fine-tuning, Re-training (CFR) framework for semi-supervised 3D medical image segmentation using SAM. The framework addresses the challenges of contextual information between adjacent slices and mismatch between natural and 3D medical images. By providing robust initialization pseudo-labels and maintaining parameter size efficiency, CFR achieves significant improvements in both moderate and scarce annotation scenarios across multiple datasets.
Stats
Our CFR framework improves the Dice score of Mean Teacher from 29.68% to 74.40% with only one labeled data of LA dataset.
CFR framework achieves extremely close performance to full supervision with moderate annotation on LA dataset.
CFR helps MT improve by 15.06% on BraTS dataset with scarce annotations.
Quotes
"Our CFR framework is plug-and-play, easily compatible with various popular semi-supervised methods."
"Extensive experiments validate that our CFR achieves significant improvements in both moderate annotation and scarce annotation across four datasets."