Superpixel-Oriented Label Distribution Learning for Skin Lesion Segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Dermatology
• Computational Pathology

Background:

• Accurate skin lesion segmentation is crucial for early skin cancer detection.
• Deep learning models require high-quality, human-annotated data, which is often difficult to obtain for dermatological images due to label uncertainty and inter-observer variability.
• Existing methods using one-hot labels can lead to overconfident and overfitted models, impacting segmentation performance.

Purpose of the Study:

• To develop a robust method for improving skin lesion segmentation by addressing the challenge of uncertain and erroneous labels.
• To introduce a superpixel-oriented label distribution learning approach that generates soft labels incorporating structural prior information.
• To enhance the performance and reliability of deep learning models for dermoscopy image analysis.

Main Methods:

• A superpixel-oriented label distribution learning method was proposed, utilizing the Simple Linear Iterative Clustering (SLIC) algorithm.
• One-hot labels were constrained and converted into soft probability distributions using a defined distance function.
• Knowledge distillation principles were applied to transfer structural prior information from soft labels to the lesion segmentation network.

Main Results:

• The proposed method achieved a Dice coefficient of 84%, sensitivity of 79.6%, and precision of 80.4% on the ISIC 2018 dataset.
• Significant improvements of 19.3% in Dice coefficient, 8.6% in sensitivity, and 2.5% in precision were observed compared to the U-Net baseline.
• The method demonstrated improved performance across various general neural network architectures for skin lesion segmentation.

Conclusions:

• The superpixel-oriented label distribution learning method effectively improves skin lesion segmentation accuracy by generating informative soft labels.
• This approach mitigates issues related to label uncertainty and enhances model robustness against overfitting.
• The proposed method is versatile and can be readily integrated into existing deep learning architectures for enhanced performance in skin cancer analysis.