autoSMIM: Automatic Superpixel-Based Masked Image Modeling for Skin Lesion Segmentation.

Area of Science:

• Dermatology
• Computer Vision
• Medical Imaging

Background:

• Skin lesion segmentation is crucial for early diagnosis but challenging due to lesion variability and blurry boundaries.
• Existing datasets often lack sufficient segmentation labels, hindering model development.
• Self-supervised learning offers a promising avenue to leverage abundant unlabeled data.

Purpose of the Study:

• To propose a novel automatic superpixel-based masked image modeling method (autoSMIM) for self-supervised skin lesion segmentation.
• To address the limitations of existing methods in handling skin lesion variability and limited labeled data.
• To improve the accuracy and efficiency of skin lesion segmentation from dermoscopic images.

Main Methods:

• Developed autoSMIM, a self-supervised masked image modeling approach utilizing superpixels.
• Employed Bayesian Optimization for an optimal superpixel masking policy.
• Pre-trained a masked image modeling model on unlabeled dermoscopic images and finetuned it for segmentation.

Main Results:

• Demonstrated the effectiveness of superpixel-based masked image modeling through ablation studies.
• Showcased the adaptability of autoSMIM across different skin lesion segmentation datasets (ISIC 2016, 2017, 2018).
• Achieved superior performance compared to state-of-the-art methods in skin lesion segmentation.

Conclusions:

• autoSMIM offers a robust and effective self-supervised solution for skin lesion segmentation.
• The proposed method overcomes challenges posed by lesion variability and limited labeled data.
• autoSMIM shows significant potential for advancing automated skin disease diagnosis.