Masked autoencoders with generalizable self-distillation for skin lesion segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Dermatology

Background:

• Accurate skin lesion segmentation is crucial for skin cancer analysis.
• Self-supervised learning, particularly Masked Autoencoders (MAE), shows promise in medical imaging by learning from unlabeled data.
• Existing methods may struggle to capture both global structure and local details in dermoscopy images.

Purpose of the Study:

• To introduce TEDMAE, a teacher-student architecture with self-distillation for enhanced skin lesion image segmentation.
• To improve the learning of global and local features in dermoscopy images using self-supervised learning.
• To enhance the generalizability of pre-trained models for skin lesion segmentation on unseen data.

Main Methods:

• Developed a Teacher-Student architecture (TEDMAE) incorporating self-distillation for holistic feature learning.
• Implemented Exterior Conversion Augmentation (EC) using convolutional kernels and linear interpolation for texture and intensity transformation.
• Utilized Dynamic Feature Generation (DF) with a nonlinear attention mechanism for improved feature expressiveness and merging.

Main Results:

• TEDMAE outperformed existing methods on ISIC2019, ISIC2017, and PH2 datasets.
• Achieved optimal segmentation performance on ISIC2017 (Dice: 82.1%) and PH2 (Dice: 91.2%).
• Demonstrated superior Jaccard (up to 84.5%) and HD95% (as low as 8.9%) scores, indicating strong generalization.

Conclusions:

• TEDMAE effectively learns comprehensive global and local features for skin lesion segmentation.
• The proposed EC and DF strategies significantly enhance model generalizability.
• TEDMAE represents a significant advancement in self-supervised learning for medical image analysis, particularly in dermatology.