Low-Memory CNNs Enabling Real-Time Ultrasound Segmentation Towards Mobile Deployment

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) are state-of-the-art for image analysis but are not optimized for ultrasound's portability and real-time use.
• Standard CNNs have large memory footprints and high computational demands, hindering deployment on mobile devices and leading to slow CPU inference.

Purpose of the Study:

• To develop efficient CNNs for real-time CPU operation in clinical ultrasound settings.
• To achieve low memory footprints and minimal accuracy compromise for medical image analysis.

Main Methods:

• Training 'thin' CNNs with reduced feature channels for faster medical image segmentation.
• Utilizing separable convolutions to enhance inference speed, decrease parameter count, and enable mobile deployment.
• Implementing a novel knowledge distillation technique to improve the accuracy of lightweight models without sacrificing speed.

Main Results:

• The proposed efficient CNN models achieve real-time performance on CPUs with significantly reduced memory requirements.
• The final model processes images at 30 frames per second on a CPU, 9x faster than standard U-Net.
• The model requires 420x less memory compared to standard U-Net, with a negligible loss in Dice performance for nerve segmentation.

Conclusions:

• Efficient CNNs can be trained for real-time ultrasound analysis on CPUs, overcoming limitations of standard models.
• The developed methods enable the deployment of advanced AI in portable, resource-constrained clinical environments.
• This research facilitates the practical application of AI in ultrasound diagnostics, enhancing both speed and accessibility.