PCF-Net: Position and context information fusion attention convolutional neural network for skin lesion segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate skin lesion segmentation is vital for effective skin cancer diagnosis and treatment.
• Variations in lesion characteristics (position, shape, size, edges) present significant challenges for automated segmentation in dermoscopic images.

Purpose of the Study:

• To develop an advanced convolutional neural network, PCF-Net, for improved skin lesion segmentation.
• To address the challenges posed by lesion variability through a novel attention mechanism and feature extraction modules.

Main Methods:

• Utilized UNet as a baseline architecture.
• Introduced a novel Position and Context Information Aggregation Attention Module (PCFAM) for fusing spatial and contextual data.
• Incorporated a Global Context Information Complementary Module (GCCM) to capture long-range dependencies.
• Developed a Multi-scale Grouped Dilated Convolution Feature Extraction Module (MSEM) for multi-scale feature analysis.

Main Results:

• Ablation experiments on the ISIC2018 dataset validated the effectiveness of PCF-Net with the integrated modules (PCFAM, GCCM, MSEM).
• PCF-Net demonstrated superior performance in dermoscopic image segmentation compared to existing state-of-the-art methods.
• Achieved competitive results across all evaluated metrics, highlighting its robustness and accuracy.

Conclusions:

• PCF-Net represents a significant advancement in automated skin lesion segmentation.
• The proposed attention and feature extraction modules effectively enhance segmentation accuracy by addressing lesion variability.
• PCF-Net offers a promising tool for improving the accuracy and efficiency of skin cancer diagnosis.