Seg-SkiNet: adaptive deformable fusion convolutional network for skin lesion segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Precise skin lesion segmentation is vital for accurate skin cancer diagnosis.
• Challenges include complex lesion shapes, varying sizes, and color depths.
• Existing methods struggle with intricate lesion features and small targets.

Purpose of the Study:

• To develop a customized deep learning (DL) model for precise skin lesion segmentation.
• To specifically address challenges posed by complex shapes and small target lesions.
• To improve diagnostic accuracy through enhanced segmentation capabilities.

Main Methods:

• Proposed an adaptive deformable fusion convolutional network (Seg-SkiNet).
• Integrated a Dual-Channel Convolution Encoder (Dual-Conv encoder) for edge and internal feature capture.
• Employed a Multi-Scale-Multi-Receptive Field Extraction and Refinement (Multi²ER) module for small lesion segmentation and a Local-Global Information Interaction Fusion Decoder (LGI-FSN decoder) for feature fusion.

Main Results:

• Seg-SkiNet achieved high performance on public datasets (ISIC-2016, ISIC-2017, ISIC-2018).
• Demonstrated Dice coefficients of 93.66%, 89.44%, and 92.29% respectively.
• Validated effectiveness in segmenting challenging complex-shaped and small skin lesions.

Conclusions:

• Seg-SkiNet shows excellent performance in segmenting complex-shaped lesions.
• The model is highly effective for segmenting small target skin lesions.
• This advancement contributes to more accurate skin cancer diagnosis through improved segmentation.