SkinFormer: Learning Statistical Texture Representation With Transformer for Skin Lesion Segmentation.

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision

Background:

• Accurate skin lesion segmentation is crucial for effective skin cancer diagnosis.
• Automatic melanoma segmentation is challenging due to difficulties in incorporating texture representations.
• Existing methods often struggle to integrate both local structural and global statistical texture information.

Purpose of the Study:

• To propose a novel Transformer network, SkinFormer, for efficient extraction and fusion of statistical texture representations in skin lesion segmentation.
• To enhance the accuracy of melanoma segmentation in dermoscopic images.

Main Methods:

• Developed a Kurtosis-guided Statistical Counting Operator to quantify statistical texture.
• Introduced Statistical Texture Fusion Transformer to fuse structural and statistical texture information.
• Proposed Statistical Texture Enhance Transformer to improve multi-scale feature statistical texture using global attention mechanisms.

Main Results:

• SkinFormer demonstrated superior performance compared to state-of-the-art (SOAT) methods on three public skin lesion datasets.
• Achieved a high Dice score of 93.2% on the ISIC 2018 dataset.
• The proposed network effectively extracts and fuses statistical texture for improved segmentation.

Conclusions:

• SkinFormer offers an efficient and effective approach for skin lesion segmentation by integrating statistical texture information.
• The method shows significant potential for improving automated skin cancer diagnosis.
• The architecture is adaptable for future applications, including 3D image segmentation.