Semi-Supervised Representation Learning for Segmentation on Medical Volumes and Sequences.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep learning excels in 2D image segmentation but struggles with high-dimensional medical data due to annotation costs.
• Existing self/semi-supervised methods lack local feature discrimination and exploitation of volumetric structures.

Purpose of the Study:

• To develop a semi-supervised representation learning method for medical volume/sequence segmentation.
• To enhance encoder and decoder features for improved segmentation performance with limited labeled data.

Main Methods:

• Proposed an asymmetric network with an attention-guided predictor for encoder feature enhancement using slice continuity.
• Introduced semantic contrastive learning for decoder feature regularization based on semantic consistency.
• Jointly trained the model using both labeled and unlabeled medical volumes/sequences.

Main Results:

• Achieved significant performance gains over existing methods on benchmark datasets (ACDC, Prostate, CAMUS) with limited labels.
• Demonstrated improvement on the M&M dataset without domain adaptation techniques.
• Intensive evaluations confirmed the effectiveness of representation mining and overall performance superiority.

Conclusions:

• The proposed semi-supervised method effectively enhances medical image segmentation by improving feature representation.
• The novel encoder and decoder modules contribute to superior performance, especially in low-data regimes.
• The approach shows promise for segmenting data from unknown domains without specialized adaptation.