WENet: a lightweight dermoscopic image segmentation network with wide edge assistance generated by morphological...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Dermoscopic image segmentation is vital for computer-aided diagnosis (CAD) of skin lesions.
• Accurate segmentation aids objective clinical decision-making.
• Existing deep learning models struggle with edge segmentation and computational complexity, limiting edge device deployment.

Purpose of the Study:

• To design a dermoscopic image segmentation model addressing edge segmentation challenges.
• To develop a model with low computational complexity for resource-constrained devices.

Main Methods:

• Developed WENet, a wide edge-assisted lightweight dermoscopic image segmentation network.
• Utilized a squeeze dual-path convolution (SDPC) encoder for efficient feature extraction and reduced complexity.
• Incorporated a wide-boundary generator (WBG) and deep supervision to enhance edge segmentation.
• Employed prediction information fusion decoding layer (PFDL) and progressive multi-scale feature fusion segmentation head (PMSSH) for refined segmentation.

Main Results:

• Achieved high mean intersection over union (mIoU) scores: 80.37% (ISIC2017), 81.34% (ISIC2018), and 85.98% (PH2).
• Obtained excellent specificity (Spe) values: 98.37% (ISIC2017), 97.39% (ISIC2018), and 96.23% (PH2).
• Maintained a model size under 15KB with significantly fewer parameters than state-of-the-art models.

Conclusions:

• WENet offers an accurate and computationally efficient solution for dermoscopic image segmentation.
• The model outperforms existing methods in model compactness and boundary segmentation precision.
• WENet is suitable for deployment on resource-constrained edge devices for skin lesion analysis.