Comparative Study of Training Strategies for Projection-Domain Denoising in Multislice Helical CT

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) is crucial for minimizing radiation exposure in patients.
• LDCT imaging is often degraded by significant noise and streak artifacts, impacting diagnostic accuracy.
• Current LDCT denoising research predominantly uses image-domain methods, leaving projection-domain denoising less explored due to its complexity and limited access to raw data.

Purpose of the Study:

• To evaluate and compare the performance of various deep learning-based denoising models specifically for 2D cone-beam CT projections.
• To investigate the impact of different training strategies on the generalizability and robustness of these denoising models.
• To assess the effectiveness of these models using both simulated and real 2D cone-beam projection data.

Main Methods:

• Four deep learning architectures were investigated: U-Net, EDCNN, GAN, and a diffusion-based model.
• Training was performed exclusively on realistic mathematical phantoms, forward-projected using cone-beam geometry and incorporating a validated noise model (Poisson and Gaussian).
• Comparative analysis was conducted on images reconstructed via filtered back projection (FBP) on PI-lines, using both phantom and real projection data for evaluation.

Main Results:

• The study demonstrated that deep learning models can be applied to projection-domain denoising for LDCT.
• Performance varied across different models and training strategies, underscoring the sensitivity of these algorithms to training data and methodology.
• The results indicated that careful consideration of physical modeling and data consistency during training is critical for achieving robust and generalizable denoising outcomes.

Conclusions:

• Deep learning-based denoising in the projection domain is feasible and holds potential for improving LDCT quality.
• The choice of training strategy, alongside accurate physical modeling and data consistency, significantly influences the success of LDCT denoising algorithms.
• This research underscores the need for advanced, physically informed deep learning approaches for reliable low-dose CT imaging.