LOW-DOSE CT MEDICAL IMAGE DENOISING: NLM AND SIMPLE CNN PERFORMANCE COMPARISON

Ismail^1,2, Susilo Widodo¹, Brazovskiy K.S.²

 

1 – Research Center for Safety, Metrology, and Nuclear Quality Technology, Research Organization for Nuclear Energy, National Research and Innovation Agency. South Tangerang, Indonesia.

2 – National Research Tomsk Polytechnic University. Tomsk, Russia.

To reduce the noise level (denoising) of low-dose CT medical images using Non-Local Means (NLM) and Convolutional Neural Networks (CNN) and then compare the performance of both quantitatively and qualitatively.

Materials and Methods. A total of 856 image pairs (low-dose and full-dose) of DICOM images from patients with code L219, image size 512 × 512 pixels, and slice thickness of 1 mm, were downloaded from the American Association of Physicists in Medicine (AAPM) open dataset. The denoising methods used were NLM and CNN. The quantitative evaluation metrics used were Root Mean Square Error (RMSE), Peak-Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Mahalanobis Distance. Qualitative evaluation is done by visual observation using IndoQCT software.

Results. The image quality resulting from the denoising process using CNN is better than NLM based on the results of calculating RMSE, PSNR, SSIM, and Mahalanobis Distance values.

Discussion. The denoising process using CNN provides better results. CNN is one of the deep learning models widely used for the denoising process. This model learns clear image patterns as ground truth and then applies the denoising process to noisy images. In carrying out the denoising process, patch-based NLM does not read the clear image pattern but replaces the pixel value with a weighted average of other pixels throughout the image itself. This weight is determined by the similarity of the patches so that detail and texture are still well maintained.

Conclusion. The CNN model successfully performed better on the denoising process, demonstrating the superiority of deep learning over classical denoising methods such as NLM. The simple CNN model proposed in this study also ran well on a standard-spec computer, opening up significant opportunities for researchers or students interested in pursuing this research topic.

 

Keywords: low-dose, computed tomography, medical image, denoising, NLM, CNN.

 

Corresponding author: Ismail, e-mail: Этот e-mail адрес защищен от спам-ботов, для его просмотра у Вас должен быть включен Javascript , Этот e-mail адрес защищен от спам-ботов, для его просмотра у Вас должен быть включен Javascript

For citation: Ismail, Susilo Widodo, Brazovskiy K.S. Low-dose CT medical image denoising: NLM and simple CNN performance comparison. REJR 2026; 16(1):158-164. DOI: 10.21569/2222-7415-2026-16-1-158-164.

Received: 18.10.25 Accepted:            12.11.25