Boosting Semi-Supervised Medical Image Segmentation Through Inter-Instance Information Complementarity.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Expert-annotated data is crucial but scarce for medical image segmentation.
• Limited labeled data hinders the clinical use of accurate segmentation models.
• Inter-instance anatomical complementarity offers a potential solution for data scarcity.

Purpose of the Study:

• To develop a novel semi-supervised approach for medical image segmentation.
• To exploit inter-instance anatomical complementarity for improved model generalization and training efficiency.
• To address the bottleneck of limited expert-annotated data in medical imaging.

Main Methods:

• A semi-supervised model integrating a copy-paste augmentation module (CPAM) and a trainable region calibration mechanism (TRCM) within a mean teacher (MT) framework.
• CPAM enhances data diversity by exchanging informative regions between samples.
• TRCM uses labeled regions to guide unlabeled data calibration, generating high-quality pseudo-labels.

Main Results:

• The proposed model demonstrates robust efficacy across diverse medical image datasets (LA, ACDC, BraTS2019, Pancreas-NIH) and modalities (MRI, CT).
• It consistently outperforms state-of-the-art methods in limited annotated data settings.
• Achieved superior performance across multiple evaluation metrics.

Conclusions:

• The novel semi-supervised approach effectively leverages inter-instance anatomical complementarity.
• The synergistic CPAM and TRCM components significantly enhance medical image segmentation performance.
• The method offers a promising solution for training accurate segmentation models with limited labeled data.