Scholarlink

Abstract <p>In this paper, we propose an approach to medical image denoising that combines a deep neural network of the U-Net type with atrous spatial pyramid pooling (ASPP) modules, coordinate layers, and adaptive modulation based on statistical noise parameters obtained from preliminary phantom analysis. This architecture qualitatively accounts for spatial noise patterns and models its intensity across different image regions, thereby enabling better preservation of small structures and contours, even in low-signal areas. To improve reconstruction quality, a combined loss function was also used, including gradient loss, structural similarity index (SSIM), and mean absolute error (MAE). Verifications on real data showed the superiority of the proposed method over state-of-the-art approaches, such as denoising convolutional neural network (DnCNN), flexible feature denoising network (FFDNet), and block-matching and 3D filtering (BM3D), both in terms of quantitative metrics (PSNR increased from 28 to 41 dB, SSIM from 0.5 to 0.97, edge preservation index (EPI) decreased from the range of 1–6.5 to a stable 1, indicating a decrease in structure errors), and in terms of visual reconstruction indicators. The results demonstrate the potential of this method for clinical application and automatic pre-enhancement of digital imaging and communications in medicine (DICOM) images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 发表日期：

    2025年2月1日

查看全文
  • D. G. Sliusarenko,
  • A. V. Netreba

摘要

Abstract

In this paper, we propose an approach to medical image denoising that combines a deep neural network of the U-Net type with atrous spatial pyramid pooling (ASPP) modules, coordinate layers, and adaptive modulation based on statistical noise parameters obtained from preliminary phantom analysis. This architecture qualitatively accounts for spatial noise patterns and models its intensity across different image regions, thereby enabling better preservation of small structures and contours, even in low-signal areas. To improve reconstruction quality, a combined loss function was also used, including gradient loss, structural similarity index (SSIM), and mean absolute error (MAE). Verifications on real data showed the superiority of the proposed method over state-of-the-art approaches, such as denoising convolutional neural network (DnCNN), flexible feature denoising network (FFDNet), and block-matching and 3D filtering (BM3D), both in terms of quantitative metrics (PSNR increased from 28 to 41 dB, SSIM from 0.5 to 0.97, edge preservation index (EPI) decreased from the range of 1–6.5 to a stable 1, indicating a decrease in structure errors), and in terms of visual reconstruction indicators. The results demonstrate the potential of this method for clinical application and automatic pre-enhancement of digital imaging and communications in medicine (DICOM) images.