QGD-Net: A Lightweight Model Utilizing Pixels of Affinity in Feature Layer for Dermoscopic Lesion Segmentation

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Pixels with location affinity contain similar semantic information, crucial for model performance.
• Existing group and dilated convolutions have limitations in leveraging these affinities across or within layers.
• Addressing these limitations is key to advancing deep learning in image analysis.

Purpose of the Study:

• To develop novel convolutional methods that effectively utilize location-affinity pixels between and within layers.
• To introduce quaternion group convolution and a quaternion sawtooth wave-like dilated convolutions module (QS module).
• To design and evaluate the quaternion group dilated neutral network (QGD-Net) for improved segmentation accuracy.

Main Methods:

• Proposed quaternion group convolution to enhance inter-channel communication and utilize affinities.
• Developed the QS module using quaternion convolution with sawtooth wave-like dilated rates for expanded receptive fields.
• Integrated these components into the quaternion group dilated neutral network (QGD-Net).

Main Results:

• QGD-Net significantly reduced model parameters compared to existing methods.
• The proposed method highly promoted precision in Dermoscopic Lesion Segmentation on ISIC datasets.
• Demonstrated generalizability in retinal vessel segmentation tasks.

Conclusions:

• The novel quaternion-based convolutional approaches effectively leverage location-affinity pixels.
• QGD-Net offers a parameter-efficient and high-precision solution for medical image segmentation.
• The method shows promise for various biomedical image analysis applications.