Scholarlink

In this paper, we address the problem of Superresolution magnetic resonance imaging (SR-MRI) to reconstruct a high-resolution (HR) image from a low-resolution (LR) scan using progressive blocks of hybrid domain network. Each block consists of cascaded stages of an encoder-decoder based network model in the frequency domain (FDN) and spatial domain (SDN). We perform the SR-MRI reconstruction by estimating the signal components encoded into the higher frequencies sequentially using FDN-SDN pairs with progressively increasing k-space coverage. In this scheme, each block of hybrid-domain encoder-decoder type convolutional neural network (CNN) is sequentially trained to estimate the unknown signal values at k-space locations within the immediate neighborhood of the region covered by the previous block. A key advantage is the consistent PSNR improvement exceeding 3.2 dB at 4-fold down-sampling, achieved without employing multi-contrast training samples or multistreaming architecture.

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

A Hybrid Domain Encoder-Decoder Network for Progressive k-space Extrapolation in Super-Resolution MRI

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Vazim Ibrahim,
  • Joseph Suresh Paul

摘要

In this paper, we address the problem of Superresolution magnetic resonance imaging (SR-MRI) to reconstruct a high-resolution (HR) image from a low-resolution (LR) scan using progressive blocks of hybrid domain network. Each block consists of cascaded stages of an encoder-decoder based network model in the frequency domain (FDN) and spatial domain (SDN). We perform the SR-MRI reconstruction by estimating the signal components encoded into the higher frequencies sequentially using FDN-SDN pairs with progressively increasing k-space coverage. In this scheme, each block of hybrid-domain encoder-decoder type convolutional neural network (CNN) is sequentially trained to estimate the unknown signal values at k-space locations within the immediate neighborhood of the region covered by the previous block. A key advantage is the consistent PSNR improvement exceeding 3.2 dB at 4-fold down-sampling, achieved without employing multi-contrast training samples or multistreaming architecture.