Deep learning-based low-dose CT simulator for non-linear reconstruction methods.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Clinical-dose computed tomography (CT) simulations often require projection data and specific reconstruction methods, limiting their use in research.
• Non-linear reconstruction methods (iterative, deep learning) complicate traditional image-domain noise simulation, making analytical noise texture determination intractable.

Purpose of the Study:

• To develop a deep learning-based, image-domain method for generating low-dose CT (LDCT) images from clinical-dose CT (CDCT) images.
• To enable LDCT image synthesis compatible with non-linear reconstruction techniques.

Main Methods:

• A three-stage convolutional neural network (CNN) approach was employed: denoising CDCT images, predicting LDCT standard deviation maps, and generating local noise power spectra (NPS).
• U-net architectures, partly or fully 3D, were utilized for all CNN models.
• Paired brain CT scans were used, with registration for motion correction, and data partitioned for training, validation, and testing.

Main Results:

• The denoising network achieved a median noise reduction factor of 4.5 in cerebrospinal fluid with minimal bias.
• The standard deviation map estimation network showed a median error of 2.1 HU.
• The NPS network accurately captured anisotropic and shift-variant noise characteristics, showing good agreement with real LDCT noise.

Conclusions:

• The proposed deep learning method successfully generates synthetic LDCT images from CDCT images without needing projection data or reconstruction algorithms.
• This technique is valuable for validation, optimization, and repeatability studies of image-processing algorithms by enabling multiple LDCT image realizations from a single CDCT image.