NCRNet: Neighborhood Context Refinement Network for skin lesion segmentation.

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Dermatology

Background:

• Accurate skin lesion segmentation is crucial for computer-aided melanoma analysis.
• Existing Fully Convolutional Network (FCN)-based methods show promise but struggle with variable lesion shapes, scales, noise, and ambiguous boundaries.
• Limitations in lesion location and boundary delineation persist in current segmentation techniques.

Purpose of the Study:

• To propose a novel Neighborhood Context Refinement Network (NCRNet) for accurate skin lesion segmentation.
• To overcome challenges in lesion location and boundary delineation posed by complex skin lesion characteristics.
• To enhance the capabilities of computer-aided melanoma analysis through improved segmentation.

Main Methods:

• A novel Neighborhood Context Refinement Network (NCRNet) employing a coarse-to-fine strategy.
• A Parallel Attention Decoder (PAD) for multi-level information fusion to locate lesions.
• A Neighborhood Context Refinement Decoder (NCRD) utilizing fine-grained neighborhood cues for boundary refinement.
• Neighborhood-based deep supervision to focus on boundary areas and promote convergence.

Main Results:

• The proposed NCRNet achieved state-of-the-art performance on the ISIC2017 skin lesion segmentation dataset.
• Achieved segmentation scores of 78.62% (Jaccard), 86.55% (Dice), and 94.01% (Accuracy).
• Outperformed nine other competitive methods in comprehensive experiments.

Conclusions:

• NCRNet effectively addresses the limitations of existing methods in skin lesion segmentation.
• The network's architecture, combining PAD and NCRD with deep supervision, enhances both lesion localization and boundary delineation.
• The results demonstrate NCRNet's potential to significantly advance computer-aided melanoma analysis.