SGU-Net: Shape-Guided Ultralight Network for Abdominal Image Segmentation.

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer-aided diagnosis

Background:

• Convolutional neural networks (CNNs) excel in medical image segmentation but require substantial parameters, hindering deployment on low-resource hardware.
• Existing compact models often compromise segmentation accuracy for reduced computational cost.

Purpose of the Study:

• To propose a shape-guided ultralight network (SGU-Net) for efficient and accurate medical image segmentation.
• To address the limitations of current CNNs in terms of parameter count and deployment on embedded systems.

Main Methods:

• Introduced an ultralight convolution combining asymmetric and depthwise separable convolutions to reduce parameters and enhance robustness.
• Incorporated an adversarial shape-constraint for self-supervised learning of target shape representations.
• Evaluated SGU-Net on LiTS, CHAOS, NIH-TCIA, and 3Dircbdb datasets.

Main Results:

• SGU-Net demonstrated superior segmentation accuracy with significantly lower memory usage compared to state-of-the-art networks.
• The ultralight convolution was successfully applied to 3D segmentation, achieving comparable performance with reduced parameters and memory footprint.
• Achieved high segmentation accuracy for abdomen medical images through self-supervised shape representation learning.

Conclusions:

• SGU-Net offers an effective solution for deploying accurate medical image segmentation on resource-constrained devices.
• The proposed ultralight convolution and shape-guided approach represent a significant advancement in efficient deep learning for medical imaging.
• SGU-Net provides a promising direction for developing practical AI tools in medical diagnostics.