Phantom-Guided Adaptive Denoising of Medical Images Using Enhanced U-Net Architecture

Abstract

In this paper, we propose an approach to medical image denoising that combines a deep neural
network of the U-Net type with ASPP modules, coordinate layers, and adaptive modulation based on
statistical noise parameters obtained from preliminary phantom analysis. This architecture
qualitatively considers spatial patterns of noise and models its intensity in different areas of the image,
which in turn allows for better preservation of various small structures and contours even in areas
with low signal. To improve the reconstruction quality, a combined loss function was additionally
used, which includes gradient loss, SSIM, and MAE. Verifications on real data showed the superiority
of the proposed method over state-of-the-art approaches, such as DnCNN, FFDNet, and BM3D, both
in terms of PSNR, SSIM, EPI, and visual reconstruction quality metrics. The results demonstrate the
potential of this method for clinical application and automatic pre-enhancement of DICOM images.