A Comparative Analysis of Enhanced Residual U-Net and DnCNN Architectures for Blind Medical Image Denoising

Abstract

High resolution CT imaging plays a sigificant role in today's diagnostic practices; yet, the images may be degraded by quantum noise and acquisition artifacts, which hide important anatomical features. Although there have been many studies on denoising using deep learning methods including DnCNN and U-Net models showing outstanding performance, the traditional approaches lack the ability to deal with different noise conditions and preserve fine structures in blind denoising. This paper compares two modified versions of deep learning networks: ER-U-Net and ER-DnCNN.
For the purposes of architectural fairness, both models employ the same optimization techniques, namely optimized local residual blocks, dynamic stochastic noise augmentation pipeline with σ noise values between 0.05 and 0.50, and a combined loss function based on the use of Mean Squared Error (MSE) (L2 loss) and Mean Absolute Error (MAE) (L1 loss) to enhance numerical and structural preservation of images. The experimental analysis of the two models conducted on a publicly available brain CT dataset has shown that both of them demonstrate strong capabilities of blind denoising at all noise levels. Specifically, the ER-U-Net outperforms in PSNR and SSIM scores, especially in conditions of severe noise, due to the presence of multi-scale feature extraction and skip connection techniques in its architecture. On the other hand, the ER-DnCNN proves to be a good alternative of high quality with simple architecture for medical imaging.