ICL-Net: Global and Local Inter-Pixel Correlations Learning Network for Skin Lesion Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Dermatology

Background:

• Skin lesion segmentation is crucial for computer-aided melanoma diagnosis.
• Existing methods face challenges with lesion variability (shape, size, boundaries) leading to classification inconsistencies.

Purpose of the Study:

• To develop a novel method for skin lesion segmentation that improves intra-class similarity and inter-class variance.
• To enhance feature representation by modeling inter-pixel correlations from global and local perspectives.

Main Methods:

• An encoder-decoder architecture incorporating a Pyramid Transformer Inter-Pixel Correlations (PTIC) module for global context and non-local information.
• A Local Neighborhood Metric Learning (LNML) module to enhance local semantic correlations and feature separability.
• Utilizing inter-pixel semantic correlations to strengthen feature representation.

Main Results:

• The proposed method demonstrated superior performance on public datasets (ISIC 2018, ISIC2016, PH2).
• Achieved better segmentation accuracy compared to existing state-of-the-art methods.
• Effectively addressed challenges of lesion shape, size, and boundary variations.

Conclusions:

• The novel method effectively models inter-pixel correlations for improved skin lesion segmentation.
• The combination of PTIC and LNML modules enhances feature representation, leading to higher diagnostic accuracy.
• This approach offers a promising advancement for computer-aided melanoma diagnosis.