深度学习提升磁共振成像 ROI 欠采样重建精度

118 浏览量更新于2024-08-26 收藏 2.32MB PDF 举报

本文是一篇研究论文，标题为《感兴趣区域欠采样MRI重建：一种深度卷积神经网络方法》（Region-of-interest undersampled MRI reconstruction: A deep convolutional neural network approach），发表在知名期刊《磁共振成像》（Magnetic Resonance Imaging）上，该期刊的网址为www.elsevier.com/locate/mri。作者包括李艳、张伟帆、邢浩(*为通讯作者)和叶虹，分别来自厦门大学智慧城市感知与计算实验室和美国哥伦比亚大学电气工程系。文章关注的是压缩感知在磁共振成像（MRI）重建中的应用，这种方法能够利用欠采样的k空间数据实现快速成像。然而，目前大多数MRI重建模型并未针对性地处理特定的组织区域，这可能导致对诊断关键和感兴趣的区域（ROI）的重建精度不足。传统的ROI重建策略通常依赖于人工提取ROI，这既耗时又依赖于专家经验，效率不高且可能影响结果的准确性。本文创新之处在于提出了一种深度卷积神经网络（Deep Convolutional Neural Network, DCNN）的方法，用于解决感兴趣区域的MRI欠采样重建问题。DCNN作为机器学习技术的一种，具有自动特征学习和适应性，能够学习并优化从欠采样数据恢复ROI图像的复杂映射关系。这种方法有望提高ROI的重建质量，减少手动操作的工作量，并可能改善医生在诊断过程中的准确性和效率。具体而言，研究可能包括以下步骤： 1. **数据预处理**：收集和准备MRI欠采样数据集，确保包含各种类型的ROI样本。 2. **网络架构设计**：构建深度卷积神经网络，可能包含编码器-解码器结构，以及多尺度或注意力机制，以便更好地捕捉ROI的细节。 3. **训练与优化**：使用有监督学习，通过损失函数（如均方误差或结构相似性指数）来调整网络参数，使其能有效恢复ROI的图像质量。 4. **性能评估**：在测试集上验证模型性能，比较其与传统方法在ROI重建上的改进，可能包括定量指标如PSNR、SSIM等。 5. **实际应用**：探讨将该方法应用于临床实践的可能性，如实时成像、病灶检测或定量分析。总结，本文的研究工作旨在填补现有MRI重建技术在关注ROI精确度方面的空白，利用深度学习的力量提升在医学图像处理领域的应用水平。这对于加快MRI扫描速度，降低辐射暴露，以及提高疾病诊断的精确性具有潜在的临床价值。

Contents lists available at ScienceDirect

Magnetic Resonance Imaging

journal homepage: www.elsevier.com/locate/mri

Original contribution

Region-of-interest undersampled MRI reconstruction: A deep convolutional

neural network approach

Liyan Sun

, Zhiwen Fan

, Xinghao Ding

*, Yue Huang

, John Paisley

Fujian Key Laboratory of Sensing and Computing for Smart City, Xiamen University, Fujian, China

Department of Electrical Engineering, Columbia University, New York, NY, USA

ARTICLE INFO

Keywords:

Deep convolutional neural network

Magnetic resonance imaging

Image reconstruction

Region of interest

ABSTRACT

Compressive sensing enables fast magnetic resonance imaging (MRI) reconstruction with undersampled k-space

data. However, in most existing MRI reconstruction models, the whole MR image is targeted and reconstructed

without taking speciﬁc tissue regions into consideration. This may fails to emphasize the reconstruction accuracy

on important and region-of-interest (ROI) tissues for diagnosis. In some ROI-based MRI reconstruction models,

the ROI mask is extracted by human experts in advance, which is laborious when the MRI datasets are too large.

In this paper, we propose a deep neural network architecture for ROI MRI reconstruction called ROIRecNet to

improve reconstruction accuracy of the ROI regions in under-sampled MRI. In the model, we obtain the ROI

masks by feeding an initially reconstructed MRI from a pre-trained MRI reconstruction network (RecNet) to a

pre-trained MRI segmentation network (ROINet). Then we ﬁne-tune the RecNet with a binary weighted ℓ

loss

function using the produced ROI mask. The resulting ROIRecNet can oﬀer more focus on the ROI. We test the

model on the MRBrainS13 dataset with diﬀerent brain tissues being ROIs. The experiment shows the proposed

ROIRecNet can signiﬁcantly improve the reconstruction quality of the region of interest.

1. Introduction

Magnetic resonance imaging (MRI) is a medical imaging modality

that generates high-resolution anatomical images with little radiation,

but is potentially limited by slow data acquisition speed [1]. Diﬀerent

acceleration techniques have been proposed such as fast pulse sequence

designs, parallel imaging, and applying compressed sensing theory.

Compressed sensing has attracted researchers' attention in general be-

cause of the wide applicability of its theory, which proves a signal can

be very accurately recovered with few measurements if it can be

sparsely represented [2]. Compressed sensing methods have been ex-

plored in fast MR imaging (CS-MRI) for acceleration and are recently

being employed by industry [3].

In conventional MRI reconstruction methods built upon compressed

sensing, a reconstructed image is produced based on interpolating the

unsampled Fourier measurements in k-space. The classic CS-MRI pro-

blem can be formulated as

= +x F x y xargmin ( ),

(1)

where

P 1

is the underlying P-dimensional vectorized MR image

and

Q 1

(Q ≪ P) is the under-sampled Q-dimensional k-space

measurements with much lower dimension. The undersamped Fourier

encoding matrix is denoted as

Q P

. By under-sampling k-space

measurements, the imaging can be signiﬁcantly accelerated. The ﬁrst

term is the data ﬁdelity and second term a regularization. In this ob-

jective, ρ(x) encodes the prior information on the desired property of

the MRI. The optimization of the problem seeks to minimize the data

ﬁdelity loss and regularization loss simultaneously. Conventionally, the

sparse priors are used to constrain the ill-posed problem like ℓ

norm [4], but recently, deep convolutional neural networks have been

utilized to address the problem [5, 6]. Wavelet sparsity and total var-

iation [1, 4] are common ﬁxed-basis regularizers (ρ(x) in Eq. (1)). Some

adaptive-basis variants of wavelets [7] or dictionary learning techni-

ques [8, 9] are also used. Recently, deep neural networks have been

applied to undersampled MRI reconstruction [10], achieving state-of-

the-art performance in both quality and eﬃciency.

From the objective in Eq. (1), it is worth noting that all the pixels in

the MR image x are typically weighted equally, regardless of the spe-

ciﬁc tissues. In other words, conventional MRI reconstruction methods

lack the ability to provide better reconstruction quality a region of in-

terest (ROI). However, in real application scenarios, diﬀerent tissues in

the same MRI represents diﬀerent biological information, and in some

cases the needs to focus on certain tissues over others exists. For

https://doi.org/10.1016/j.mri.2019.07.010

Received 23 March 2019; Received in revised form 6 July 2019; Accepted 14 July 2019

Corresponding author.

E-mail address: dxh@xmu.edu.cn (X. Ding).

Magnetic Resonance Imaging 63 (2019) 185–192

下载后可阅读完整内容，剩余7页未读，立即下载

weixin_38725450

粉丝: 1
资源: 913

深度学习提升磁共振成像 ROI 欠采样重建精度

边缘修正的多尺度卷积神经网络重建算法

LOUPE:基于学习的MRI欠采样模式优化

资源三号卫星影像云识别：多尺度膨胀卷积神经网络方法

基于NSST的迭代收缩阈值压缩感知欠采样MRI重建

CNN-3D-images-Tensorflow：使用CNN（卷积神经网络）进行3D图像分类

深度迭代卷积神经网络的快速脑部MRI重建算法.pdf

基于3D深度卷积神经网络依据MRI生成伪CT的研究.pdf

深度学习入门：李宏毅讲授CNN卷积神经网络

深度学习基础：LeNet-5卷积神经网络解析

小数据集MRI重建：神经网络优化算法探索

最新资源