Gradient-based Optimization Algorithm for Hybrid Loss Function in Low-dose CT Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose computed tomography (CT) imaging reduces patient radiation exposure but often results in noisy images.
• Deep learning (DL) based post-processing methods are used for denoising CT images, but performance can be further enhanced.
• The choice of objective functions significantly impacts the parameter optimization and denoising quality in DL networks.

Purpose of the Study:

• To identify and analyze the most effective and efficient hybrid loss functions for DL-based CT image denoising.
• To investigate the impact of different combinations of common objective functions on denoising performance.
• To introduce a hyperparameter learning algorithm for optimizing hybrid loss function scalability.

Main Methods:

• Examined combinations of L1 loss, per-pixel loss, perceptual loss, and structural dissimilarity loss.
• Utilized a hyperparameter learning algorithm to determine optimal scaling factors for hybrid loss functions.
• Employed the RED-CNN architecture to demonstrate and evaluate the performance of various loss function combinations.

Main Results:

• Hybrid loss functions, when balanced via gradient-based optimization, show potential for improved CT image denoising.
• The study identified specific combinations and scaling factors that enhance denoising effectiveness.
• The RED-CNN model effectively showcased the performance differences influenced by the loss functions.

Conclusions:

• Balancing hybrid loss functions is crucial for enhancing the generalizability of future CT denoising networks.
• The findings provide insights into optimizing DL models for high-quality, low-dose CT imaging.
• This research contributes to the development of safer and more effective medical imaging techniques.