DMFF-Net: a multi-scale feature fusion network based on DeepLabV3 for skin lesion segmentation

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Dermatology

Background:

• Precise skin lesion segmentation is vital for early skin cancer detection.
• Challenges include unclear boundaries, low contrast, and varied lesion shapes.
• Existing methods struggle with complex segmentation tasks.

Purpose of the Study:

• To propose DMFF-Net, a multi-scale, multi-attention feature fusion network.
• To improve the accuracy of skin lesion segmentation in medical images.
• To address limitations of current segmentation techniques.

Main Methods:

• Developed DMFF-Net based on DeepLabV3 architecture.
• Integrated Global Grid Coordinate Attention Module (GGCAM) for spatial-channel feature fusion.
• Employed Multi-Scale Depthwise Separable Dilated Convolution (MDSDC) for robust multi-scale feature extraction.
• Utilized Mid-High Level Feature Fusion (MHLFF) to refine features and suppress noise.

Main Results:

• DMFF-Net achieved superior performance on ISIC 2016, 2017, 2018, and PH2 datasets.
• Reported MIoU values of 89.31%, 91.47%, 86.93% on ISIC datasets.
• Achieved high accuracy (up to 97.33%) and F1 scores (up to 96.93%) demonstrating effectiveness.

Conclusions:

• DMFF-Net significantly improves skin lesion segmentation accuracy.
• The network's multi-scale fusion and attention mechanisms preserve spatial details and enhance feature representation.
• DMFF-Net shows potential as a valuable tool for skin lesion diagnosis and medical image segmentation research.