RGShuffleNet: An Efficient Design for Medical Image Segmentation on Portable Devices.

Area of Science:

• Medical image analysis
• Computer-assisted diagnosis
• Mobile health technology

Background:

• Medical image segmentation is vital for diagnosis and surgical planning.
• UNet and its variants dominate medical image segmentation but require substantial computational resources.
• Current models are difficult to deploy on resource-limited portable devices, hindering in-home (Homelab) applications.

Purpose of the Study:

• To develop a lightweight medical image segmentation model for resource-constrained mobile devices.
• To address the computational cost limitations of existing segmentation models.
• To enable rapid and efficient medical image segmentation outside of hospital settings.

Main Methods:

• Introduction of RGShuffleNet, a novel lightweight model for medical image segmentation.
• Development of Reshaped Group Convolution to reduce parameters and computational complexity by restructuring feature groups.
• Implementation of the MSC-Shuffle block for enhanced information flow across channel and spatial dimensions between feature groups.

Main Results:

• RGShuffleNet demonstrated superior segmentation performance compared to state-of-the-art methods on cardiac ultrasound and chest CT datasets.
• The model achieved significantly lower computational complexity, making it suitable for mobile deployment.
• Successful deployment of RGShuffleNet on portable devices was achieved.

Conclusions:

• RGShuffleNet offers an effective solution for lightweight medical image segmentation on mobile devices.
• The proposed Reshaped Group Convolution and MSC-Shuffle block contribute to improved efficiency and performance.
• This advancement facilitates advanced medical image analysis in portable and resource-limited environments.