.png){kind=icon}
.png){kind=icon}
Return
Speckle Noise Reduction in Ultrasound Medical Images via Multimodule Denoising Autoencoder
.png){kind=icon}
PRE
AI
F
M
F
DOI：10.1002/ima.70403.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Speckle noise in ultrasound imaging obscures anatomical boundaries and reduces image contrast, thereby limiting diagnostic accuracy and the performance of computer-aided diagnosis (CAD) systems. Existing deep learning methods achieve noise reduction but often apply spatially uniform processing, resulting in oversmoothing of fine structures or residual artifacts in heterogeneous tissue regions. This work presents a Multimodule Denoising Autoencoder (MM-DAE) that integrates three complementary processing modules within a unified encoder–decoder architecture. The Adaptive Filtering Module (AFM) generates spatially adaptive convolutional kernels guided by learned spatial descriptors. The Noise Pattern Memory Module (NPMM) maintains a learnable prototype bank and performs similarity- and confidence-weighted suppression to attenuate noise-like components. The Multiscale Graph Attention Module (MSGAM) leverages learnable graph attention across dilated multiscale features to preserve structural consistency. Evaluation on thyroid (TNUS), nerve (UNS), and fetal (HC18) datasets with synthetic speckle noise achieves peak signal-to-noise ratio (PSNR) gains of 2.4–2.7 dB over recent methods, with a structural similarity index (SSIM) of 0.96 and an edge preservation index (EPI) of 0.66, indicating maintained structural fidelity. On clinical breast ultrasound, contrast-to-noise ratio (CNR) and SNR increase by 0.60 and 1.68 dB respectively, indicating enhanced lesion-to-background separation with preserved internal echo patterns. The complementary modules outperform the individual components, particularly in boundary regions and low-contrast tissue.
Keywords:
adaptive filtering
deep learning
denoising autoencoder
multimodule architecture
multiscale processing
speckle noise reduction
ultrasound imaging
Journal
IF:
2.5
Papers:
2.1K
Citations:
2.3K
Organization
Cited Papers