Skin lesion segmentation using a semi-supervised U-NetSC model with an adaptive loss function.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin lesion segmentation is vital for accurate cancer detection.
• Deep learning models excel at segmentation but require substantial pixel-level annotated data.
• Semi-supervised learning offers a solution to reduce annotation dependency.

Purpose of the Study:

• To propose a semi-supervised strategy for skin lesion segmentation.
• To develop a unified architecture (U-NetSC) that integrates segmentation and classification tasks.
• To improve the efficiency of training segmentation models by reducing the need for pixel-level annotations.

Main Methods:

• Developed the U-Net Segmentation and Classification (U-NetSC) model, a unified architecture with collaborative segmentation and classification modules.
• Utilized feature maps from the segmentation model's last layer within the classification module.
• Employed an adaptive loss function enabling training with class-level or both class-level and pixel-level ground truth.

Main Results:

• The U-NetSC model demonstrated improved performance compared to a supervised attention-gated U-Net model.
• Achieved approximately 2% improvement in Jaccard Index and Dice coefficient.
• Reported ~3% increase in precision and ~1% increase in accuracy.

Conclusions:

• The proposed semi-supervised approach effectively segments skin lesions, reducing the reliance on large annotated datasets.
• The U-NetSC model enhances collaboration between segmentation and classification for improved performance.
• This method holds clinical relevance for automated skin lesion analysis in cancer detection.