Segmenting medical images with limited data.

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Computer vision is vital for medical image segmentation but struggles with small datasets and unlabeled data.
• Existing methods often require large labeled datasets, limiting their applicability in medical contexts.

Purpose of the Study:

• To introduce a novel semi-supervised, consistency-based approach called the Data-Efficient Medical Segmenter (DEMS).
• To enhance the generalization ability and data efficiency of medical image segmentation models, particularly under data-scarce conditions.

Main Methods:

• DEMS employs an encoder-decoder architecture with Online Automatic Augmentation (OAA) and Residual Robustness Enhancement (RRE) blocks.
• OAA diversifies datasets through image transformations, while RRE enriches feature diversity and introduces perturbations for varied decoder inputs.
• A sensitive loss function is utilized to improve cross-decoder consistency and stabilize training.

Main Results:

• DEMS demonstrated significant effectiveness across multiple datasets, including those with extreme data shortages.
• Achieved superior dice scores compared to U-Net (16.85% improvement) and state-of-the-art methods (10.37% improvement) in low-data scenarios.
• The approach shows strong performance and data efficiency, outperforming existing methods under data limitations.

Conclusions:

• DEMS offers a robust and data-efficient solution for medical image segmentation, particularly valuable in low-data regimes.
• The developed OAA and RRE blocks, along with the sensitive loss, contribute to improved segmentation accuracy and model generalization.
• This method holds potential for advancing medical segmentation applications where labeled data is scarce.