Random Search as a Neural Network Optimization Strategy for Convolutional-Neural-Network (CNN)-based Noise Reduction...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning image processing algorithms require extensive parameter tuning for optimal performance.
• Existing methods for optimizing these algorithms can be time-consuming and may not explore the full parameter space effectively.

Purpose of the Study:

• To systematically optimize deep learning-based image processing algorithms using random search.
• To demonstrate the generalizability of this optimization technique for various deep learning image processing applications, including noise reduction.

Main Methods:

• Employed random search to explore a parameter space including convolutional layers, filters, kernel size, loss functions, and network architectures (U-Net, ResNet).
• Conducted 100 network model examinations (50 random search, 50 ablation experiments) on a phantom-based noise reduction framework.
• Utilized a weighted feature reconstruction loss (0.2×VGG + 0.8×MSE) with a U-Net architecture.

Main Results:

• Random search identified near-optimal settings, with ablation experiments yielding only minor performance improvements.
• The top-performing model, a U-Net, achieved 90% noise reduction, 34% RMSE reduction, and 76% SSIM increase on test data.
• Significantly improved visualization of hepatic and bone lesions in low-dose input images.

Conclusions:

• Random search provides an efficient and effective systematic approach for optimizing deep learning image processing algorithms.
• The developed deep learning model demonstrates substantial improvements in image quality and lesion visualization for medical imaging applications.