MSS-UNet: A Multi-Spatial-Shift MLP-based UNet for skin lesion segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Multilayer perceptron (MLP) networks offer fewer parameters than convolutional neural networks (CNNs) and transformers but increase computational complexity with depth for image processing.
• Automated segmentation of skin lesions is crucial for diagnosis and treatment planning in dermatology.

Purpose of the Study:

• To develop a lightweight yet effective model for automated skin lesion segmentation from dermoscopic images.
• To address the computational challenges of deep MLP models in processing local image features.

Main Methods:

• Proposed MSS-UNet, a hybrid CNN-MLP model incorporating a convolutional module for local feature extraction.
• Introduced a Double-Spatial-Shift MLP (DSS-MLP) module to enhance spatial communication within MLPs.
• Developed a Multi-Spatial-Shift External Attention (MSSEA) module to expand the receptive field and capture lesion boundary continuity.

Main Results:

• Achieved high Intersection over Union (IoU) scores on ISIC 2017 (85.01%±0.65), ISIC 2018 (83.65%±1.05), and PH2 (92.71%±1.03) datasets.
• Demonstrated significantly reduced parameter size (0.33M) and computational complexity (15.98G) compared to existing methods.
• Outperformed most state-of-the-art models in skin lesion segmentation accuracy and efficiency.

Conclusions:

• MSS-UNet provides an efficient and accurate solution for automated skin lesion segmentation.
• The proposed DSS-MLP and MSSEA modules effectively enhance MLP capabilities for medical image analysis.
• This lightweight model holds promise for practical clinical applications in dermatology.