Learning CT projection denoising from adjacent views.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Learning-based low-dose (LD) computed tomography (CT) methods often need large paired datasets, which are scarce in clinical settings.
• Models trained on simulated data frequently fail to generalize to real clinical data.

Purpose of the Study:

• To develop an unsupervised learning technique for reconstructing clean CT projection data from adjacent LD projections.

Main Methods:

• An unsupervised approach was used, treating the middle projection of a sequential LD set as the target and others as input.
• The model was trained using mean absolute error with an inter-view gradient constraint to reduce outliers and preserve edges.
• Experiments utilized simulated (CT Grand Challenge) and real (torso phantom) datasets, evaluating with Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).

Main Results:

• Visual comparisons showed the proposed method outperformed other unsupervised and supervised methods in both image and projection domains on simulated and real data.
• Numerically, the method achieved higher SSIMs than other unsupervised methods at quarter and eighth dose levels.
• PSNR results were higher at eighth dose but lower at quarter dose compared to other unsupervised methods; supervised models generally performed better on simulated data.

Conclusions:

• The developed method effectively reduces noise in CT projections.
• This technique presents a promising tool for practical low-dose CT pre-processing.