Semi-Supervised Burn Depth Segmentation Network with Contrast Learning and Uncertainty Correction.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Accurate burn depth diagnosis is critical for effective treatment and patient outcomes.
• Image-based deep learning offers automated burn depth segmentation but is hindered by limited labeled data.
• Existing semi-supervised methods struggle with burn segmentation due to issues like single-level perturbations and inaccurate pseudo-labels.

Purpose of the Study:

• To develop an advanced semi-supervised network, SBCU-Net, for improved burn depth segmentation.
• To address the limitations of traditional deep learning and current semi-supervised approaches in handling scarce burn data.
• To enhance segmentation accuracy, particularly in complex burn edge regions.

Main Methods:

• Proposed SBCU-Net, a semi-supervised network incorporating contrastive learning and uncertainty correction.
• Introduced multi-level perturbations and two additional decoder branches for enhanced consistency.
• Utilized contrastive learning to refine segmentation outputs and improve feature representation.
• Implemented an uncertainty correction mechanism to mitigate the impact of inaccurate pseudo-labels.

Main Results:

• SBCU-Net demonstrated superior performance in burn depth segmentation compared to state-of-the-art semi-supervised methods.
• The proposed method effectively leveraged unlabeled data to boost segmentation accuracy.
• Improvements were particularly noted in segmenting complex regions like burn edges.

Conclusions:

• SBCU-Net offers a robust solution for semi-supervised burn depth segmentation, overcoming data scarcity challenges.
• The integration of contrastive learning and uncertainty correction significantly enhances segmentation performance.
• This approach holds promise for improving the automation and standardization of burn diagnosis.