Scholarlink

<p>Medical image denoising is fundamentally constrained by a conflict between aggressive noise suppression, detail preservation, and the instability of manually tuned filtering parameters. To address this limitation, we propose an adaptive coarse-to-fine denoising framework in which noise control, feature refinement, and reconstruction are jointly designed rather than sequentially combined. The core innovation is an automatically parameterized fast bilateral filtering mechanism, where wavelet-domain statistics guide spatial and range parameter adaptation in a data-driven manner, eliminating manual tuning. This adaptive filtering stage defines a structured prior that is further refined by a heterogeneous residual CNN, explicitly designed for multi-scale feature correction, while a variational autoencoder regularizes reconstruction consistency and suppresses over-smoothing artifacts. An adaptive loss weighting strategy stabilizes optimization across heterogeneous objectives. Experiments on MRI (T1/T2) and low-dose CT datasets demonstrate consistent and significant improvements, achieving up to 37.06&#xa0;dB/0.89 SSIM, 38.09&#xa0;dB/0.79 SSIM, and 36.92&#xa0;dB/0.98 SSIM, outperforming state-of-the-art denoising methods.</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

Adaptive medical image denoising via optimized fast bilateral filtering, wavelet-guided CNNs, and variational autoencoders

  • 期刊论文

  • 发表日期：

    2026年5月24日

查看全文
  • Omid Darvishi ghaleh,
  • Vafa Maihami,
  • Keyhan Khamforoosh

摘要

Medical image denoising is fundamentally constrained by a conflict between aggressive noise suppression, detail preservation, and the instability of manually tuned filtering parameters. To address this limitation, we propose an adaptive coarse-to-fine denoising framework in which noise control, feature refinement, and reconstruction are jointly designed rather than sequentially combined. The core innovation is an automatically parameterized fast bilateral filtering mechanism, where wavelet-domain statistics guide spatial and range parameter adaptation in a data-driven manner, eliminating manual tuning. This adaptive filtering stage defines a structured prior that is further refined by a heterogeneous residual CNN, explicitly designed for multi-scale feature correction, while a variational autoencoder regularizes reconstruction consistency and suppresses over-smoothing artifacts. An adaptive loss weighting strategy stabilizes optimization across heterogeneous objectives. Experiments on MRI (T1/T2) and low-dose CT datasets demonstrate consistent and significant improvements, achieving up to 37.06 dB/0.89 SSIM, 38.09 dB/0.79 SSIM, and 36.92 dB/0.98 SSIM, outperforming state-of-the-art denoising methods.