MASDF-Net: A Multi-Attention Codec Network with Selective and Dynamic Fusion for Skin Lesion Segmentation

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision

Background:

• Automated segmentation of dermoscopic images aids dermatologists in diagnosis.
• Current deep learning methods struggle with irregular shapes, blurry edges, and artifacts in skin lesion segmentation.

Purpose of the Study:

• To propose a novel multi-attention codec network with selective and dynamic fusion (MASDF-Net) for enhanced skin lesion segmentation.
• To improve the accuracy of delineating lesion boundaries in challenging dermoscopic images.

Main Methods:

• Utilized a pyramid vision transformer as the encoder to capture long-range feature dependencies.
• Introduced three novel modules: Multi-Attention Fusion (MAF) for global context, Selective Information Gathering (SIG) to refine low-level features, and Multi-Scale Cascade Fusion (MSCF) for dynamic decoder feature fusion.
• Employed extensive experiments on ISIC 2016, ISIC 2017, ISIC 2018, and PH2 datasets.

Main Results:

• The MASDF-Net demonstrated superior performance compared to existing state-of-the-art methods.
• The proposed modules effectively addressed challenges such as irregular lesion shapes, blurry edges, and occlusions.
• Achieved significant improvements in accurately segmenting skin lesion boundaries.

Conclusions:

• MASDF-Net offers a robust and effective solution for skin lesion segmentation in dermoscopic images.
• The network's innovative fusion modules enhance the capture of contextual information and refine segmentation accuracy.
• This advancement holds promise for improving diagnostic assistance in clinical dermatology.