Transferring U-Net between low-dose CT denoising tasks: a validation study with varied spatial resolutions.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning, particularly U-Net, is increasingly used for low-dose computed tomography (LDCT) to enhance image quality at reduced radiation doses.
• Validating the generalizability of trained deep learning models across diverse LDCT datasets from different imaging systems is crucial for clinical adoption.

Purpose of the Study:

• To assess the reproducibility of a pre-trained U-Net's denoising performance on LDCT images with varying spatial resolutions.
• To investigate the impact of spatial resolution differences on U-Net denoising transferability and identify potential artifacts.

Main Methods:

• A U-Net model was trained on LDCT images and then validated across datasets with six different spatial resolutions (62.5–625 µm).
• Quantitative metrics including residual variance, PSNR, NRMSE, and SSIM were used for performance evaluation.
• Network retraining with a subset of data was explored to mitigate cross-resolution artifacts.

Main Results:

• The U-Net demonstrated effectiveness in denoising LDCT images across different spatial resolutions, though cross-validation introduced image artifacts.
• Artifacts were more pronounced with increasing spatial resolution inconsistency, particularly at object margins and centers.
• Retraining the U-Net with approximately 20% of the original data significantly reduced artifacts (NRMSE decreased from 0.1898 to 0.1263, SSIM increased from 0.7558 to 0.8036).

Conclusions:

• Transferring a U-Net trained for LDCT denoising to datasets with different spatial resolutions can lead to artifact generation.
• Retraining the U-Net with a small, targeted dataset of the desired spatial resolution is recommended to maintain optimal denoising performance and minimize artifacts.