Scholarlink

<p>The existence of speckle noise in medical ultrasound (US) images greatly diminishes image quality and decreases diagnostic accuracy. The recent models for de-noising US images suffer from poor generalizability, a loss of texture information, and over-smoothing of the US images, especially at high noise levels. This study suggests an efficient and robust de-speckling filter for US images based on an attentional auto-encoder. This filter incorporates two distinct and efficient attention levels. The channel attention level (CAL) and large kernel attention level (KAL). These levels are integrated into the down-sampling and up-sampling steps, respectively. A skip link is added before the up-sampling stage to avoid the gradient vanishing during training. The performance of the introduced model was assessed using subjective visual evaluations and objective determination indices. The outcomes of the suggested approach on online databases achieved the minimum execution time (AET), the maximum values for structural similarity (SSIM), and peak signal-to-noise ratio (PSNR) indices over all levels of noise, exceeding other recent algorithms. For instance, at a high level of noise, the presented technique based on the breast database achieves the following mean values: 0.005, 0.81, and 23.71&#xa0;dB for the AET, SSIM, and PSNR, respectively. Also, in clinical settings, the introduced model attained the highest values for the equivalent number of looks (ENL = 18.78), contrast-to-noise ratio (CNR = 5.12), and highest resolution (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2024_3777_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="23" /> </InlineMediaObject> <EquationSource Format="TEX">\(\alpha^{\sim }\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>α</mi> <mo>∼</mo> </msup> </math></EquationSource> </InlineEquation> = 0.28). These findings prove that the presented architecture is more efficient and robust for rejecting noise while maintaining image details than other recent techniques.</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

Efficient and robust de-speckling filter for ultrasound images based on an attentional auto-encoder

  • 期刊论文

  • 发表日期：

    2025年1月3日

查看全文
  • Anas Fouad Ahmed

摘要

The existence of speckle noise in medical ultrasound (US) images greatly diminishes image quality and decreases diagnostic accuracy. The recent models for de-noising US images suffer from poor generalizability, a loss of texture information, and over-smoothing of the US images, especially at high noise levels. This study suggests an efficient and robust de-speckling filter for US images based on an attentional auto-encoder. This filter incorporates two distinct and efficient attention levels. The channel attention level (CAL) and large kernel attention level (KAL). These levels are integrated into the down-sampling and up-sampling steps, respectively. A skip link is added before the up-sampling stage to avoid the gradient vanishing during training. The performance of the introduced model was assessed using subjective visual evaluations and objective determination indices. The outcomes of the suggested approach on online databases achieved the minimum execution time (AET), the maximum values for structural similarity (SSIM), and peak signal-to-noise ratio (PSNR) indices over all levels of noise, exceeding other recent algorithms. For instance, at a high level of noise, the presented technique based on the breast database achieves the following mean values: 0.005, 0.81, and 23.71 dB for the AET, SSIM, and PSNR, respectively. Also, in clinical settings, the introduced model attained the highest values for the equivalent number of looks (ENL = 18.78), contrast-to-noise ratio (CNR = 5.12), and highest resolution ( \(\alpha^{\sim }\) α ∼ = 0.28). These findings prove that the presented architecture is more efficient and robust for rejecting noise while maintaining image details than other recent techniques.