Performance evaluation of reduced complexity deep neural networks.

Area of Science:

• Computer Science
• Medical Imaging
• Artificial Intelligence

Background:

• Deep Neural Networks (DNNs) excel at medical image classification but are computationally intensive.
• Complexity reduction is crucial for deploying DNNs in low-power edge applications.
• Existing methods often involve time-consuming operations and performance trade-offs.

Purpose of the Study:

• To propose a simplified model complexity reduction technique for DNNs.
• To demonstrate complexity reduction for ResNet-50 integration in low-power devices.
• To evaluate performance for multiclass classification of Chest X-ray (CXR) images.

Main Methods:

• A novel channel reduction technique applied to DNNs.
• ResNet-50 model complexity reduction for low-power edge devices.
• Multiclass classification of CXR images (normal, pneumonia, COVID-19).
• Model generalization and Grad-CAM visualization for interpretability.
• Theoretical VLSI architecture design for optimized performance.

Main Results:

• Achieved successive size reductions of 75%, 87%, and 93%.
• Minimal classification performance reduction of 0.5%, 0.5%, and 0.8% respectively.
• Demonstrated acceptable model generalization and interpretable visualizations.
• Presented a theoretical VLSI architecture for the best performing model.

Conclusions:

• The proposed channel reduction technique effectively reduces DNN complexity for low-power medical imaging applications.
• Significant model size reduction is achievable with negligible impact on classification performance.
• The method offers a practical approach for deploying advanced AI models on edge devices for disease diagnosis.