Self-trained deep convolutional neural network for noise reduction in CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep convolutional neural networks (CNNs) are used for clinical CT noise reduction.
• Training these networks requires large, paired datasets, which are difficult to obtain and may lack generalizability.

Purpose of the Study:

• To propose a self-trained deep CNN (ST_CNN) for CT image noise reduction that eliminates the need for pre-existing training datasets.
• To develop a method that is generalizable across different imaging conditions.

Main Methods:

• Extensive data augmentation in the projection domain, including noise insertion and rotation augmentation.
• Self-training approach where the network learns from the patient's own data.
• Reconstruction of paired low- and high-quality images from the same patient for training.

Main Results:

• ST_CNN demonstrated comparable peak signal-to-noise ratio and structural similarity index measure to conventional CNNs.
• ST_CNN showed improved noise texture, homogeneity in liver parenchyma, and better subjective visualization of liver lesions.
• Slight sacrifice in vessel sharpness was observed but did not impact diagnostic visibility.

Conclusions:

• The self-trained deep CNN (ST_CNN) method can achieve similar image quality to conventional deep CNN denoising methods.
• ST_CNN offers a viable alternative for CT noise reduction, particularly when external training datasets are unavailable or unsuitable.