Multislice input for 2D and 3D residual convolutional neural network noise reduction in CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep convolutional neural networks (CNNs) are increasingly applied for noise reduction in computed tomography (CT) imaging.
• Existing CNN denoising methods utilize either 2D or 3D architectures with single- or multi-slice inputs.

Purpose of the Study:

• To evaluate the impact of multi-slice input on CNN denoising performance compared to single-slice input.
• To determine if 3D CNN architectures offer advantages over 2D CNNs when processing multi-slice CT data.

Main Methods:

• Compared 2D CNN models with one, three, and seven input slices.
• Evaluated two 3D CNN models using seven input slices (one or three output slices).
• Assessed performance on liver CT images using quantitative metrics and radiologist visual assessment.

Main Results:

• Increasing the number of input slices for 2D CNN models showed a trend of improved denoising performance.
• The 3D CNN model with a single-slice output outperformed other models in noise texture, homogeneity, and vessel visualization.

Conclusions:

• Multi-slice input is an effective strategy for enhancing the performance of 2D deep CNN denoising models.
• While 3D CNNs show potential for improved axial slice continuity, performance differences compared to 2D CNNs with equivalent slice inputs were not statistically significant.