PocketNet: A Smaller Neural Network for Medical Image Analysis

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Deep learning models for medical imaging are computationally intensive.
• Training and evaluating these models require specialized hardware.
• Model size and complexity pose challenges for deployment in resource-limited environments.

Purpose of the Study:

• To introduce the PocketNet paradigm for reducing the size of deep learning models.
• To enable efficient training and deployment of medical imaging models in resource-constrained settings.
• To maintain performance while significantly decreasing model complexity.

Main Methods:

• Proposing the PocketNet paradigm to control channel growth in convolutional neural networks.
• Applying PocketNet architectures to various medical imaging segmentation and classification tasks.
• Evaluating model performance, parameter count, GPU memory usage, and training time.

Main Results:

• PocketNet architectures achieve results comparable to conventional neural networks.
• Significant reduction in the number of parameters (orders of magnitude).
• Reduced GPU memory usage (up to 90%) and accelerated training times (up to 40%).

Conclusions:

• PocketNet offers an effective solution for creating smaller, more efficient deep learning models for medical imaging.
• The paradigm facilitates the use of advanced deep learning in settings with limited computational resources.
• PocketNet demonstrates the feasibility of deploying high-performance models under hardware constraints.