HDS-Net: Achieving fine-grained skin lesion segmentation using hybrid encoding and dynamic sparse attention.

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Skin cancer is a prevalent malignancy requiring early detection for improved prognoses.
• Accurate segmentation of skin lesions in medical images is vital for diagnosis and treatment planning.
• Deep learning models offer adaptive feature learning for medical image segmentation but face challenges with skin lesions.

Purpose of the Study:

• To introduce a novel network model, HDS-Net (Hybrid Dynamic Sparse Network), for enhanced skin lesion segmentation.
• To address key challenges in skin lesion segmentation, including boundary ambiguity and variations in lesion size, shape, and type.
• To improve the accuracy and reliability of automated skin lesion analysis.

Main Methods:

• Developed a novel HDS-Net incorporating a hybrid encoder for integrated local and global feature extraction.
• Introduced a dynamic sparse attention mechanism to mitigate the impact of irrelevant information by controlling sparsity.
• Evaluated the model's performance on multiple public skin lesion datasets.

Main Results:

• The HDS-Net achieved significant improvements in segmentation accuracy across multiple datasets.
• Demonstrated superior performance with Dice coefficients of 0.914, 0.857, and 0.898 on public datasets.
• The hybrid encoder and dynamic sparse attention effectively handled variations and ambiguities in skin lesion images.

Conclusions:

• HDS-Net offers a robust solution for accurate skin lesion segmentation, outperforming existing methods.
• The proposed model enhances the potential for early and precise diagnosis of skin cancer through improved medical image analysis.
• This advancement in deep learning for medical imaging contributes to better patient outcomes in skin cancer management.