Reducing the U-Net size for practical scenarios: Virus recognition in electron microscopy images

Area of Science:

• Computer Vision
• Machine Learning
• Biomedical Imaging

Background:

• Convolutional Neural Networks (CNNs) achieve expert-level performance but require substantial hardware.
• High hardware demands limit CNNs in real-time applications like clinical and biomedical analysis.
• This study focuses on reducing the computational cost of CNNs, specifically U-Net.

Purpose of the Study:

• Investigate methods to create lighter and more efficient CNN architectures.
• Reduce the number of trainable weights in the U-Net model.
• Enable practical, on-the-fly analysis in resource-constrained environments.

Main Methods:

• Explored U-Net hyper-parameters including feature map count, multiplier, and network depth.
• Applied pixel-wise virus classification in Transmission Electron Microscopy images.
• Utilized minimal annotations (virus particle centers or centerlines) for training.

Main Results:

• Architecture hyper-parameters are critical for reducing trainable weights effectively.
• A U-Net model achieved 82.2% accuracy with ~4x fewer weights (7.8M vs 31.0M).
• A network with <2M weights achieved 76.4% accuracy, demonstrating significant weight reduction.

Conclusions:

• U-Net hyper-parameter optimization enables efficient CNN design for reduced weight usage.
• Lighter CNNs are essential for practical implementation in various applications.
• Optimized networks are faster, less prone to overfitting, and generalize better.