# Evaluating the impact of deep learning‐based image denoising on low‐dose CT for lung cancer screening

## TL;DR

This study evaluates how deep learning-based image denoising improves low-dose CT scans for lung cancer screening by reducing noise and enhancing image quality.

## Contribution

The study introduces a comprehensive evaluation of seven deep learning denoising methods on low-dose CT images using both image quality metrics and nodule-specific features.

## Key findings

- Denoising improved image quality metrics like SSIM, RMSE, and PSNR for both solid and subsolid nodules.
- Denoising reduced percent differences in nodule size, volume, and density measurements.
- Denoising increased Lung-RADS classification accuracy for solid nodules but had less impact on subsolid nodules.

## Abstract

Low‐dose CT (LDCT) is increasingly being adopted as a preferred method for lung cancer screening. However, the accompanying rise in image noise necessitates robust denoising strategies. Therefore, this study compared LDCT images with their denoised counterparts using objective image quality metrics and key nodule‐related features.

The dataset utilized in this study was chest CT scans for lung cancer screening, sourced from the LDCT and Projection Data collection. Seven deep learning‐based image denoising methods were used in this work. The denoising performance was evaluated using root‐mean‐square error (RMSE), peak signal‐to‐noise ratio (PSNR), structural similarity index measure (SSIM), nodule size, CT density, and Lung‐RADS classification.

For solid nodules, denoising improved SSIM from 51% to 60%–64%, reduced RMSE from 137.13 HU to 62.40–78.30 HU, and increased PSNR from 23.91 dB to 28.59–30.51 dB. It also reduced the percent difference in diameter (PDdia) from 2.05% to 1.44%–1.52%, in volume (PDvol) from 5.95% to 4.43%–4.70%, in mean HU value (PDHU) from 24.40% to 8.54%–15.33%. For subsolid nodules, denoising improved SSIM from 47% to 57%–61%, reduced RMSE from 110.87 HU to 54.62–63.96 HU, and increased PSNR from 25.78 dB to 30.53–31.61 dB. Before denoising, the PDdia, PDvol and PDHU were 15.41%, 40.16% and 10.69%, respectively, which were 7.54%–15.94%, 17.54%–29.29%, and 6.10%–8.25% after denoising. These improvements led to higher Lung‐RADS categorization accuracy for solid nodules, while subsolid nodules remained more affected by noise and denoising‐induced bias.

The integration of denoising techniques into LDCT workflows could potentially enhance early lung cancer detection without increasing radiation exposure. Nonetheless, validating their influence on diagnostic performance remains crucial for clinical adoption.