Performance of a deep learning-based CT image denoising method: Generalizability over dose, reconstruction kernel,...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning (DL) shows promise for low-dose CT image denoising, potentially improving image quality over traditional filtered back projection (FBP).
• However, the generalizability of DL denoising methods across diverse CT acquisition parameters remains incompletely understood.

Purpose of the Study:

• To investigate the performance generalizability of a low-dose CT denoising neural network (REDCNN) under varying reconstruction kernel, slice thickness, and dose levels.
• To identify CT scan data properties that predict the generalizability of DL denoising networks.

Main Methods:

• Trained a residual encoder-decoder convolutional neural network (REDCNN) on CT datasets, varying one parameter (reconstruction kernel, slice thickness, or dose) at a time.
• Evaluated denoising performance using metrics like MSE, MTF, pNPS, and LCD on datasets with matching and mismatching parameters.

Main Results:

• REDCNN exhibited larger MSE when testing data reconstruction kernels differed from training data.
• Performance was robust to slice thickness variations but showed slightly worse MSE for higher-dose images when trained on quarter-dose data.
• Mixed-dose training improved low-contrast resolution preservation; smooth-kernel training struggled with sharp-kernel noise removal.

Conclusions:

• REDCNN demonstrates poor generalizability between reconstruction kernels but is robust to dose variations when trained with mixed-dose data.
• Slice thickness does not significantly impact network performance.
• Generalizability may correlate with the pixel-level noise power spectrum (pNPS) similarity between training and testing data.