TY - GEN
T1 - Deep Residual Shrinkage Proximal-Point Mapping Network for Image Reconstruction
AU - Xu, Yuan
AU - Liao, Heng Wei
AU - Hao, Da Ben
AU - He, Yan Lin
AU - Zhu, Qun Xiong
AU - Yong, Tao
AU - Xiao, Zi Feng
AU - Zhang, Yang
AU - Zhang, Ming Qing
N1 - Publisher Copyright:
© 2025 IEEE.
PY - 2025
Y1 - 2025
N2 - Image processing procedures in medical imaging are highly complex and are often affected by artifacts and noise, which impact image quality and diagnostic accuracy. To address this issue, we propose a novel deep residual shrinkage proximal-point mapping network (RSPPM-Net) aimed at enhancing image reconstruction performance through feature denoising. The innovation of RSPPM-Net lies in its integration of stochastic gradient descent modules and localized residual shrinkage proximal mapping modules, effectively reducing the interference of redundant features. Additionally, the network introduces a cascading module called the Fourier residual shrinkage block (FRSBlock) to correct potential biases between the input and output signals. FRSBlock integrates a residual FFT block and two identical adaptive blocks, each formed by cascading a global average pooling layer, a fully connected layer, a batch normalization layer, and an activation function. Through this architecture, FRSBlock can adaptively generate and adjust denoising thresholds based on the input data, achieving localized multi-level denoising and precise signal correction. To validate the effectiveness of RSPPM-Net, we conducted extensive reconstruction experiments on orthopedic joint imaging datasets. Experimental results demonstrate that RSPPM-Net exhibits superior performance in image recovery quality, significantly improving the clarity and accuracy of reconstructed images.
AB - Image processing procedures in medical imaging are highly complex and are often affected by artifacts and noise, which impact image quality and diagnostic accuracy. To address this issue, we propose a novel deep residual shrinkage proximal-point mapping network (RSPPM-Net) aimed at enhancing image reconstruction performance through feature denoising. The innovation of RSPPM-Net lies in its integration of stochastic gradient descent modules and localized residual shrinkage proximal mapping modules, effectively reducing the interference of redundant features. Additionally, the network introduces a cascading module called the Fourier residual shrinkage block (FRSBlock) to correct potential biases between the input and output signals. FRSBlock integrates a residual FFT block and two identical adaptive blocks, each formed by cascading a global average pooling layer, a fully connected layer, a batch normalization layer, and an activation function. Through this architecture, FRSBlock can adaptively generate and adjust denoising thresholds based on the input data, achieving localized multi-level denoising and precise signal correction. To validate the effectiveness of RSPPM-Net, we conducted extensive reconstruction experiments on orthopedic joint imaging datasets. Experimental results demonstrate that RSPPM-Net exhibits superior performance in image recovery quality, significantly improving the clarity and accuracy of reconstructed images.
KW - Deep residual shrinkage proximal-point mapping network
KW - Fourier residual shrinkage block
KW - Image reconstruction
KW - Multi-stage cascading adaptive module
UR - https://www.scopus.com/pages/publications/105013958282
U2 - 10.1109/CCDC65474.2025.11090671
DO - 10.1109/CCDC65474.2025.11090671
M3 - Conference contribution
AN - SCOPUS:105013958282
T3 - Proceedings of the 37th Chinese Control and Decision Conference, CCDC 2025
SP - 3415
EP - 3420
BT - Proceedings of the 37th Chinese Control and Decision Conference, CCDC 2025
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 37th Chinese Control and Decision Conference, CCDC 2025
Y2 - 16 May 2025 through 19 May 2025
ER -