WO2021115084A1 - 一种基于结构磁共振影像的大脑年龄深度学习预测系统 - Google Patents

一种基于结构磁共振影像的大脑年龄深度学习预测系统 Download PDF

Info

Publication number
WO2021115084A1
WO2021115084A1 PCT/CN2020/130073 CN2020130073W WO2021115084A1 WO 2021115084 A1 WO2021115084 A1 WO 2021115084A1 CN 2020130073 W CN2020130073 W CN 2020130073W WO 2021115084 A1 WO2021115084 A1 WO 2021115084A1
Authority
WO
WIPO (PCT)
Prior art keywords
brain
age
magnetic resonance
image
neural network
Prior art date
Application number
PCT/CN2020/130073
Other languages
English (en)
French (fr)
Inventor
刘涛
程健
刘子阳
Original Assignee
北京航空航天大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京航空航天大学 filed Critical 北京航空航天大学
Publication of WO2021115084A1 publication Critical patent/WO2021115084A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves 
    • A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves  involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/0033—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room
    • A61B5/004—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room adapted for image acquisition of a particular organ or body part
    • A61B5/0042—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room adapted for image acquisition of a particular organ or body part for the brain
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B2503/00—Evaluating a particular growth phase or type of persons or animals
    • A61B2503/08—Elderly
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B2576/00—Medical imaging apparatus involving image processing or analysis
    • A61B2576/02—Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part
    • A61B2576/026—Medical imaging apparatus involving image processing or analysis specially adapted for a particular organ or body part for the brain

Definitions

  • the present invention relates to the field of computer and neuroscience, in particular to a brain age prediction system based on structural magnetic resonance images.
  • Artificial intelligence methods can build a predictive model of brain aging by using magnetic resonance images of the brain structure to predict the age of the elderly.
  • the age predicted by the model is called "brain age".
  • the age of the brain can indicate the current aging stage of the brain and even predict the risk of related diseases in the future.
  • the prediction model established by the image data of healthy elderly actually describes a normal elderly brain aging trajectory, and the difference between the age of the brain and its true age can reflect the degree to which a person deviates from the aging trajectory of the healthy brain.
  • the degree of early or late aging of the brain It has been proven that the greater the difference between the brain age of the elderly and the real age, the higher the risk of mental or physical problems and the more likely to die early. In clinical practice, doctors can use this indicator to assess the degree of brain aging in the elderly and take corresponding intervention measures.
  • the method of predicting the age of the brain with structural magnetic resonance images is mainly based on traditional machine learning methods.
  • a large amount of preprocessing and feature extraction are required for the MRI images of the brain structure.
  • the features commonly used for brain age prediction include: Gray Matter Density Map (GMD), White Matter Density Map (WMD), White Matter Volume, Cortical Thickness, Network Feature Parameters, etc.
  • the extracted feature dimensions may range from tens of dimensions to hundreds of dimensions. Therefore, in the process of using traditional machine learning methods to establish a brain age prediction model, it is often necessary to perform feature selection or feature dimensionality reduction steps.
  • GMD-based models can be used to predict brain age, but because GMD-based models have the disadvantage of high feature dimensions, the constructed models are prone to overfitting and poor generalization ability.
  • extraction of image features is easily affected by many factors. For example, optional parameters such as image smoothing and voxel size are involved in the process of GMD feature generation. These parameters have a great impact on the results of feature extraction and even brain age prediction. influences.
  • convolutional neural networks are being tried to predict the age of the brain, but unlike natural images, structural magnetic resonance imaging is a three-dimensional image, and the training sample size is much smaller than natural images, so this is very important for convolutional neural network structures.
  • the structure of MRI presents a higher challenge, requiring the convolutional neural network to fully extract and utilize the information contained in the magnetic resonance image.
  • the current convolutional neural networks applied to the brain age prediction task are too traditional and only use the structure of convolution + pooling + full connection, and the efficiency of feature extraction and utilization needs to be improved.
  • the purpose of the present invention is to provide a brain age prediction system based on structured magnetic resonance images, which predicts the brain age of the elderly based on structured magnetic resonance images through convolutional neural networks.
  • the present invention regards the convolutional network as a high-dimensional feature extractor, and performs hierarchical characterization and description of the feature information of the magnetic resonance image of the complex brain structure, first expresses the underlying features, and then combines these low-level features to obtain a more detailed Rich high-level features are expressed, and effective features are efficiently extracted to evaluate the current aging stage of the brain, that is, "brain age”.
  • the present invention proposes an AGE-DenseNet network structure on the basis of a traditional convolutional network, which can greatly improve the utilization efficiency of features and obtain an excellent brain age prediction effect.
  • the invention has fast analysis speed, strong generalization ability, practicability and ease of use.
  • the present invention provides a brain age deep learning prediction system based on structural magnetic resonance images, including:
  • the training set building module is used to construct the training set using the original structure magnetic resonance images from different sources;
  • the data preprocessing module is used to perform preprocessing operations on the magnetic resonance images of each original structure.
  • the preprocessing operations include image registration, skull stripping and image data standardization;
  • the deep learning model building module is used to build a convolutional neural network model AGE-DenseNet based on the Keras framework and the DenseNet idea;
  • the deep learning model training module is used to use the constructed training set to perform supervised learning on the constructed convolutional neural network model AGE-DenseNet through back propagation and gradient descent algorithms;
  • the brain age prediction module is used to send the preprocessed MRI image of the original structure to the convolutional neural network model AGE-DenseNet to predict the age of the brain, obtain the predicted value of the brain age, and calculate the predicted value of the brain age and the actual age The difference between, assess the degree of brain aging.
  • the convolutional neural network model AGE-DenseNet includes five repeated convolution blocks, and each convolution block includes two identical convolution units and a 2 ⁇ 2 ⁇ 2 maximum pooling layer with a step size of 2.
  • Each convolution unit includes a 3 ⁇ 3 ⁇ 3 convolution layer with a step length of 1, a ReLU activation and a 3D batch normalization layer,
  • the maximum pooling method When learning and training, use the maximum pooling method to downsample the feature maps output by each convolution block, change its size, and then concatenate it with the feature maps output by other convolution blocks into a single tensor as the current The input of the convolution block; and after the last convolution block, the global average pooling layer (Global average pooling) is used to vectorize the feature map into a feature vector.
  • the global average pooling layer (Global average pooling) is used to vectorize the feature map into a feature vector.
  • a unary fully connected layer plus the ReLU activation function is used to map the feature vector obtained by the global average pooling to a single output value.
  • the image registration includes performing a non-linear registration operation on the original structure magnetic resonance image
  • the skull dissection includes acquiring a registered structural magnetic resonance image of the skull dissection image through a preset threshold;
  • the image data normalization includes calculating the average value and standard deviation of the voxels in the brain contour after skull stripping, and performing Gaussian normalization on the voxels in the brain contour.
  • the present invention has the following advantages:
  • the present invention adopts the convolutional neural network structure based on DenseNet idea-AGE-DenseNet. Compared with the traditional machine learning prediction model, the present invention can directly analyze and process the whole brain structure magnetic resonance image without manually extracting the structural features of the brain (such as gray density map, white matter density map, white matter volume, cortical thickness, etc.). It can avoid the tedious process of feature extraction and feature selection/dimensionality reduction.
  • the present invention adopts a "convolution + dense connection (Dense connection) + global average pooling" structure in the deep learning model design process, that is, adding between convolutional layers
  • the tight short-circuit connection can effectively improve the utilization efficiency of features, and obviously alleviate the problem of gradient disappearance during the gradient backpropagation process.
  • the global average pooling layer is used to replace the traditional fully connected layer, which saves more space At the same time of information, it slows down the phenomenon of over-fitting to a certain extent.
  • the present invention can accurately, efficiently and quickly predict the age of the brain of the elderly through the magnetic resonance imaging of the brain structure, and quantify the degree of deviation of the brain from the aging track of the healthy brain, especially for the early diagnosis and prevention of diseases of the elderly. Provide an important basis.
  • Figure 1 is a flow chart of the brain age prediction method based on structured magnetic resonance imaging of the present invention
  • Figure 2 is a structural diagram of the convolutional neural network model AGE-DenseNet of the present invention.
  • FIG. 3 is a schematic diagram of deep learning prediction of brain age based on structural magnetic resonance images according to an embodiment of the present invention.
  • Brain aging is a universal biological phenomenon that appears with age, which can lead to subtle changes in the structure of the brain. Even healthy elderly people will have a certain degree of decline in their cognitive ability as they age, and the risk of disease will increase accordingly. This is a normal aging trajectory. However, some people deviate from this normal aging trajectory due to genes, environment, or disease.
  • the present invention aims to quantify the early or delayed aging of the brain, and use this to predict the future development trajectory of individuals and the subsequent The risk of aging-related health deterioration.
  • the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on it.
  • the present invention is preferentially applicable to MRI images of brain structures that have undergone non-linear registration and skull stripping, and the preprocessing process uses the FSL5.0 fsl_anat command.
  • the present invention first preprocesses the image to meet the input requirements of the convolutional neural network model AGE-DenseNet of the present invention, such as image size, etc.; and then convolves and pools the image through the convolutional neural network model AGE-DenseNet And other feature extraction operations; finally, the extracted high-dimensional features are fused to obtain the predicted brain age.
  • the method for predicting brain age based on structural magnetic resonance imaging of the present invention includes the following steps:
  • the original input image is a T1 magnetic resonance image of the brain.
  • the preprocessing operations include image non-linear registration, skull stripping, and image data normalization.
  • image registration includes: performing a non-linear registration operation on the input image; skull peeling includes: obtaining a skull peeled image of the registered T1 image through a preset threshold; image data standardization includes: calculating the skull peeling For the average value and standard deviation of voxels in the brain contour, Gaussian normalization is performed on the voxels in the brain contour, and then the background outside the brain contour is set to 0.
  • Preprocessing the original input image can improve the prediction accuracy of the depth method to predict the age of the brain on the one hand, and accelerate the processing and analysis speed on the other hand, with high efficiency and ease of use.
  • the convolutional neural network architecture of the present invention uses fixed-size magnetic resonance image data of the three-dimensional brain T1 structure as input.
  • This CNN architecture contains five repeated Convolutional Blocks, and each Convolutional Block contains two identical Convolutional Units and a 2 ⁇ 2 ⁇ 2 maximum pooling layer with a step size of 2.
  • the convolution unit includes a 3 ⁇ 3 ⁇ 3 convolution layer with a step length of 1, a ReLU activation and a 3D batch normalization layer.
  • the number of feature channels is set to 8, and it is doubled after entering the next convolution block to infer a sufficiently rich representation of brain information.
  • the present invention is based on DenseNet, which connects the feature maps learned by different convolution blocks in series, increases the input variables of subsequent layers, and improves the feature utilization efficiency and learning of the network. effect.
  • DenseNet which connects the feature maps learned by different convolution blocks in series, increases the input variables of subsequent layers, and improves the feature utilization efficiency and learning of the network. effect.
  • Each convolution block will integrate all the previous convolution block output feature mapping information as input. Since the size of the output feature maps of different convolution blocks is different, first use the maximum pooling method to downsample the feature maps, change their size, and then concatenate them with the feature maps output by other convolution blocks into a single The tensor of is used as the input of the current convolution block.
  • the global average pooling layer (Global average pooling) is used to vectorize the feature map into a feature vector.
  • the final age prediction is to use a unary fully connected layer plus the ReLU activation function, which maps the feature vector obtained by the global average pooling to a single output value.
  • the structure of the convolutional neural network model AGE-DenseNet of the present invention is shown in FIG. 2.
  • S4 Use the training set to train and learn the constructed convolutional neural network model AGE-DenseNet through backpropagation and gradient descent algorithms, and select a model with high prediction accuracy and strong generalization performance for storage, which is convenient for users to call.
  • S5 Use the cross-validation method to verify the convolutional neural network model AGE-DenseNet, and adjust the hyperparameters in the training of the convolutional neural network model AGE-DenseNet.
  • S6 Send the MRI image of the original structure after the preprocessing operation to the convolutional neural network model AGE-DenseNet, which is verified to be accurate, to predict the age of the brain, obtain the predicted value of the brain age, and calculate the difference between the predicted value of the brain age and the actual age. Assess how far the current brain deviates from the aging trajectory of a healthy brain.
  • the skull in the registered image is stripped, and only the brain tissue is retained;
  • this scheme first obtains the rough brain contour by setting the threshold, and then calculates the average and standard deviation of the voxels in the brain contour. Gaussian normalization of brain voxels;

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Animal Behavior & Ethology (AREA)
  • Public Health (AREA)
  • Pathology (AREA)
  • Radiology & Medical Imaging (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Veterinary Medicine (AREA)
  • High Energy & Nuclear Physics (AREA)
  • Neurology (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Signal Processing (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

一种基于结构磁共振影像的大脑年龄深度学习预测系统，包括：训练集构建模块；数据预处理模块；深度学习模型构建模块，用于利用Keras框架以及基于DenseNet思想，构建卷积神经网络模型AGE-DenseNet；深度学习模型训练模块；模型验证模块；大脑年龄预测模块。采用基于DenseNet思想的卷积神经网络结构AGE-DenseNet，能够从复杂的大脑结构磁共振影像中提炼出高维复杂特征，准确、高效、快速地预测出大脑年龄，并以此量化大脑偏离健康大脑衰老轨迹的程度。

Description

一种基于结构磁共振影像的大脑年龄深度学习预测系统
本申请要求于2019年12月11日提交中国专利局、申请号为201911266178.X、发明名称为“一种基于结构磁共振影像的大脑年龄深度学习预测系统”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及计算机及神经科学领域，特别涉及一种基于结构磁共振影像的大脑年龄深度学习预测系统。
背景技术
随着全球人口老龄化问题的日益严重，与老化有关的大脑疾病正在给社会造成越来越大的负担。而人类的大脑会随着年龄的增长而在结构上发生一些微妙的变化，这些变化会导致大脑在正常功能上产生退化，并与神经退行性等脑部疾病呈现出显著的相关性。基因、环境、疾病或受伤等原因可能会导致大脑的老化速率有显著的加快，需要有方法来量化这种异常的大脑衰老速度，评估当前大脑所处的衰老阶段。
人工智能方法可以利用大脑结构磁共振影像建立一个大脑老化的预测模型，以此对老年人的年龄进行预测，由模型预测出的年龄我们称之为“大脑年龄”。大脑年龄可以表明当前大脑所处的老化阶段，甚至预测未来的相关疾病的风险。由健康老年人影像数据所建立的预测模型，实际上描述了一种正常的老年人大脑衰老轨迹，而利用大脑年龄与其真实年龄的差值，可以反映一个人偏离健康大脑老化轨迹的程度，反映了其大脑衰老的提前或延后程度。现已被证明，老年人大脑年龄与真实年龄差距越大，则其精神或身体出现问题的风险越高，也更容易早逝。在临床上，医生可使用这一指标，评估老年人的大脑衰老程度，并采取相应的干预措施。
目前以结构磁共振影像预测大脑年龄的方法，主要以传统机器学习方法为主。在建立预测模型时，需要对大脑结构磁共振影像进行大量的预处理及特征提取工作。目前，常用于大脑年龄预测的特征有：灰度密度图(Grey matter density map,GMD)、白质密度图(White matter density  map,WMD)、白质体积、皮层厚度、网络特征参数等。提取的特征维度可能从数十维到上百维。因此，在利用传统机器学习方法建立大脑年龄预测模型过程中，往往需要进行特征选择或特征降维等步骤。例如，基于GMD特征，高斯回归过程、支持向量机等机器学习模型可用于预测大脑年龄，但由于基于GMD的模型具有特征维度高的缺点，所构建的模型容易过拟合，泛化能力差。另外，图像特征的提取易受到诸多因素的影响，例如：在GMD特征生成过程中就涉及图像平滑、体素大小等可选参数，这些参数对特征提取乃至大脑年龄预测的结果都有很大的影响。
此外，卷积神经网络正被尝试用于大脑年龄预测中，但不同于自然图像，结构磁共振影像是一种三维的图像，而且训练样本量远小于自然图像，因此这对于卷积神经网络结构的结构提出了较高的挑战，要求卷积神经网络能对磁共振影像中所包含的信息进行充分的提取和利用。而目前应用于大脑年龄预测任务的卷积神经网络均过于传统，仅使用了卷积+池化+全连接的结构，对特征的提取及利用效率都有待提高。
发明内容
本发明的目的是提供一种基于结构磁共振影像的大脑年龄深度学习预测系统，通过卷积神经网络，基于结构磁共振影像对老年人大脑年龄进行预测。本发明将卷积网络视作一种高维特征提取器，对复杂的大脑结构磁共振影像的特征信息进行分层次表征描述，先表示出底层的特征，然后将这些低层特征组合得到更为细致丰富的高层特征进行表达，高效地提炼出有效特征，以此评估当前大脑所处的衰老阶段，即“大脑年龄”。本发明在传统卷积网络的基础上提出AGE-DenseNet网络结构，能够大大提高特征的利用效率，获得优异的大脑年龄预测效果。本发明分析速度快，泛化能力强，且具有实用性、易用性。
本发明提供了一种基于结构磁共振影像的大脑年龄深度学习预测系统，包括：
训练集构建模块，用于利用不同来源的原始结构磁共振影像构建训练集；
数据预处理模块，用于对各原始结构磁共振影像进行预处理操作，预 处理操作包括图像配准、头骨剥离以及图像数据标准化；
深度学习模型构建模块，用于利用Keras框架以及基于DenseNet思想，构建卷积神经网络模型AGE-DenseNet；
深度学习模型训练模块，用于利用所构建的训练集，通过反向传播和梯度下降算法对构建的卷积神经网络模型AGE-DenseNet进行监督学习；
模型验证模块，用于利用交叉验证的方法对卷积神经网络模型AGE-DenseNet进行验证，调整卷积神经网络模型AGE-DenseNet训练中的超参数；
大脑年龄预测模块，用于将预处理操作后的原始结构磁共振影像送入验证准确的卷积神经网络模型AGE-DenseNet进行大脑年龄预测，得到大脑年龄预测值，计算大脑年龄预测值与实际年龄的差值，评估大脑衰老程度。
进一步，所述卷积神经网络模型AGE-DenseNet包括五个重复的卷积块，各卷积块包括两个完全相同的卷积单元和一个步长为2的2×2×2最大池化层，各卷积单元包括步长为1的3×3×3卷积层、一个ReLU激活以及一个3D批标准化层，
学习训练时，用最大池化的方式将各卷积块输出的特征映射进行下采样，改变其尺寸，然后再将其与其他卷积块输出的特征映射串联成一个单独的张量，作为当前卷积块的输入；并在最后一个卷积块结束后，使用全局平均池化层(Global average pooling)，将特征映射矢量化为一个特征向量。大脑年龄预测时，使用一个一元全连接层加上ReLU激活函数，将全局平均池化得到的特征向量映射到一个单独的输出值。
进一步，所述图像配准包括将原始结构磁共振影像执行非线性配准操作；
所述头骨剥离包括通过预先设定阈值获取已配准后的结构磁共振影像的头骨剥离图像；
所述图像数据标准化包括计算经过头骨剥离后的大脑轮廓内体素的平均值和标准差，将大脑轮廓内体素进行高斯标准化。
本发明与现有技术相比，其优点是：
1)本发明采用了基于DenseNet思想的卷积神经网络结构 —AGE-DenseNet。相比于传统机器学习预测模型，本发明可以直接对全脑结构磁共振影像进行分析处理，无需手动提取大脑的结构特征(如灰度密度图、白质密度图、白质体积、皮层厚度等)，可以免去繁琐的特征提取及特征选择/降维等过程。另外，相比于一般的卷积神经网络，本发明在深度学习模型设计过程中，采用“卷积+紧密连接(Dense connection)+全局平均池化”结构，即，在卷积层之间添加紧密的短路连接，能够有效地提高特征的利用效率，并在梯度反向传播过程中明显缓解了梯度消失问题的发生，同时使用全局平均池化层代替传统的全连接层，在保留更多空间信息的同时，一定程度上减缓过拟合现象。
2)本发明能通过大脑结构磁共振影像准确、高效、快速地预测出老年人的大脑年龄，并以此量化其大脑偏离健康大脑衰老轨迹的程度，为尤其是老年人疾病的早期诊断与防治提供了重要依据。
说明书附图
下面结合附图对本发明作进一步说明：
图1为本发明的基于结构磁共振影像的大脑年龄深度学习预测方法流程图；
图2为本发明的卷积神经网络模型AGE-DenseNet的结构图；
图3为本发明实施例的基于结构磁共振影像的大脑年龄深度学习预测示意图。
具体实施方式
下面结合本发明实施例中的附图，对本发明实施例中技术方案进行详细的描述，显然，所描述的实施例仅是本发明一部分实施例，而不是全部的实施例；基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例都属于本发明保护的范围。
大脑衰老是随着年龄增长而呈现的一种普遍生物学现象，其会导致大脑结构上发生微妙的改变。即使是健康老年人也会随着年龄的不断增长，认知能力有一定程度的下降，患病风险随之增加，这是一种正常的衰老轨迹。但存在部分人群因受到基因、环境或患病而偏离这种正常的衰老轨迹，本发明旨在量化大脑提前或延后的老化，并以此来预测个人未来发展轨迹 和随之而来的与老化相关的健康恶化风险。下面结合附图和实施例进一步描述本发明，应该理解，以下所述实施例旨在便于对本发明的理解，而对其不起任何限定作用。
首先需要说明的是，本发明优先适用于经过非线性配准及剥离头骨的大脑结构磁共振影像，预处理过程使用FSL5.0 fsl_anat命令。本发明首先将图像进行预处理，使其满足本发明的卷积神经网络模型AGE-DenseNet的输入需求，例如图像尺寸等；再通过卷积神经网络模型AGE-DenseNet对图像进行卷积、池化等特征提取操作；最后将提取出的高维特征进行融合，得到预测出的大脑年龄。
如图1所示，本发明的基于结构磁共振影像的大脑年龄深度学习预测方法，包括如下步骤：
S1:构建原始输入图像；
原始输入图像为大脑T1磁共振影像。
S2：对原始输入图像进行预处理操作并利用预处理后的原始输入图像构建训练集；
所述预处理操作包括图像非线性配准，头骨剥离以及图像数据标准化。具体的，图像配准包括：将输入影像执行非线性配准操作；头骨剥离包括：通过预先设定阈值获取已配准后T1图像的头骨剥离图像；图像数据标准化包括：计算经过头骨剥离后的大脑轮廓内体素的平均值和标准差，将大脑轮廓内体素进行高斯标准化，再将大脑轮廓外背景设置为0。对原始输入图像进行预处理操作，一方面可以提高深度学方法预测大脑年龄的预测精度，另一方面加快了处理分析的速度，具有高效性与易用性。
S3:构建卷积神经网络模型AGE-DenseNet；
本发明的卷积神经网络架构使用固定大小的三维大脑T1结构磁共振影像数据作为输入。这个CNN架构包含五个重复的卷积块(Convolutional Block)，每个卷积块包含两个完全相同的卷积单元(Convolutional Unit)和一个步长为2的2×2×2最大池化层。卷积单元包含步长为1的3×3×3卷积层，一个ReLU激活以及一个3D批标准化层。在第一个卷积块中，特征通道的数量被设置为8，并在进入到下一个卷积块之后对其进行加倍，以推断出一个足够丰富的大脑信息表征。
另外，为缓解梯度消失问题，提高特征的利用效率，本发明基于DenseNet，将不同卷积块所学习到的特征映射串联起来，增加了后续层输入的变量，提高了网络的特征利用效率及学习效果。每个卷积块都将综合之前所有卷积块输出特征映射的信息以作输入。由于不同卷积块输出特征映射的尺寸是不同的，所以首先使用最大池化的方式将特征映射进行下采样，改变其尺寸，然后再将其与其他卷积块输出的特征映射串联成一个单独的张量，作为当前卷积块的输入。并在最后一个卷积块结束后，使用全局平均池化层(Globalaverage pooling)，将特征映射矢量化为一个特征向量。最后的年龄预测是使用一个一元全连接层加上ReLU激活函数，它将全局平均池化得到的特征向量映射到一个单独的输出值。本发明的卷积神经网络模型AGE-DenseNet结构如图2所示。
S4:利用训练集，通过反向传播和梯度下降算法对构建的卷积神经网络模型AGE-DenseNet进行训练学习，选取预测精度高、泛化性能强的模型进行保存，方便用户调用。
S5:利用交叉验证的方法对卷积神经网络模型AGE-DenseNet进行验证，调整卷积神经网络模型AGE-DenseNet训练中的超参数。
S6:将预处理操作后的原始结构磁共振影像送入验证准确的卷积神经网络模型AGE-DenseNet进行大脑年龄预测，得到大脑年龄预测值，计算大脑年龄预测值与实际年龄的差值，以评估当前大脑偏离健康大脑衰老轨迹的程度。
下面利用具体实施例来更加清晰准确地理解并应用本发明。
如图3所示，首先构建原始输入，即大脑T1磁共振影像作为原始输入。再进行图像预处理，将原始输入图像经非线性配准至MNI152-2mm模板中，配准后数据尺寸为91×109×91；
根据设定好的阈值，对配准后的影像中的头骨进行剥离，仅保留脑组织；
进行数据标准化，由于MRI图像存在着大量的黑色背景，会极大影响数据标准化效果，因此本方案首先通过设定阈值得到粗糙的大脑轮廓，然后计算大脑轮廓内体素的平均值和标准差，将脑部体素进行高斯标准化；
选取前期已训练好的卷积神经网络模型AGE-DenseNet，根据输入图像进行大脑年龄预测，得到大脑年龄预测值，将大脑年龄与实际年龄做差，即可得到当前大脑偏离健康衰老轨迹的量化值。
上面结合附图对本发明的实施方式作了详细说明，但是本发明并不限于上述实施方式，在所属技术领域普通技术人员所具备的知识范围内，还可以在不脱离本发明宗旨的前提下做出各种变化。

Claims (8)

  1. 一种基于结构磁共振影像的大脑年龄深度学习预测系统，其特征在于，包括：
    训练集构建模块，用于利用不同来源的原始结构磁共振影像构建训练集；
    数据预处理模块，用于对各原始结构磁共振影像进行预处理操作，预处理操作包括图像配准、头骨剥离以及图像数据标准化；
    深度学习模型构建模块，用于利用Keras框架以及基于DenseNet思想，构建卷积神经网络模型AGE-DenseNet；
    深度学习模型训练模块，用于利用所构建的训练集，通过反向传播和梯度下降算法对构建的卷积神经网络模型AGE-DenseNet进行监督学习；
    模型验证模块，用于利用交叉验证的方法对卷积神经网络模型AGE-DenseNet进行验证，调整卷积神经网络模型AGE-DenseNet训练中的超参数；
    大脑年龄预测模块，用于将预处理操作后的原始结构磁共振影像送入验证准确的卷积神经网络模型AGE-DenseNet进行大脑年龄预测，得到大脑年龄预测值，计算大脑年龄预测值与实际年龄的差值，评估大脑衰老程度。
  2. 根据权利要求1所述的系统，其特征在于，所述卷积神经网络模型AGE-DenseNet包括五个重复的卷积块，各卷积块包括两个完全相同的卷积单元和一个步长为2的2×2×2最大池化层，各卷积单元包括步长为1的3×3×3卷积层、一个ReLU激活以及一个3D批标准化层；
    学习训练时，用最大池化的方式将各卷积块输出的特征映射进行下采样，改变其尺寸，然后再将其与其他卷积块输出的特征映射串联成一个单独的张量，作为当前卷积块的输入，并在最后一个卷积块结束后，使用全局平均池化层，将特征映射矢量化为一个特征向量；
    大脑年龄预测时，使用一个一元全连接层加上ReLU激活函数，将全局平均池化得到的特征向量映射到一个单独的输出值。
  3. 根据权利要求1或2所述的系统，其特征在于，所述图像配准包括将原始结构磁共振影像执行非线性配准操作；
    所述头骨剥离包括通过预先设定阈值获取已配准后的结构磁共振影 像的头骨剥离图像；
    所述图像数据标准化包括计算经过头骨剥离后的大脑轮廓内体素的平均值和标准差，将大脑轮廓内体素进行高斯标准化。
  4. 一种基于结构磁共振影像的大脑年龄预测方法，其特征在于，包括：
    构建原始输入图像；所述原始输入图像为大脑磁共振影像；
    根据所述原始输入图像构建训练集；
    构建卷积神经网络模型；
    利用所述训练集，通过反向传播和梯度下降算法对构建的卷积神经网络模型进行训练，得到训练好的模型；
    将待预测的大脑磁共振影像输入所述训练好的模型，得到大脑年龄预测值。
  5. 根据权利要求4所述的基于结构磁共振影像的大脑年龄预测方法，其特征在于，所述根据所述原始输入图像构建训练集，具体包括：
    将所述原始输入图像执行非线性配准操作；
    通过预先设定阈值对已配准后的图像进行头骨剥离，得到头骨剥离图像；
    计算头骨剥离图像的大脑轮廓内体素的平均值和标准差，将大脑轮廓内体素进行高斯标准化，再将大脑轮廓外背景设置为0，得到图像数据标准化的结果；
    根据图像数据标准化的结果构建训练集。
  6. 根据权利要求4所述的基于结构磁共振影像的大脑年龄预测方法，其特征在于，所述卷积神经网络模型的架构包含五个重复的卷积块，每个卷积块包含两个完全相同的卷积单元和一个步长为2的2×2×2最大池化层；每个卷积单元包含步长为1的3×3×3卷积层、一个ReLU激活以及一个3D批标准化层；不同卷积块学习到的特征映射基于DenseNe串联。
  7. 根据权利要求4所述的基于结构磁共振影像的大脑年龄预测方法，其特征在于，所述利用所述训练集，通过反向传播和梯度下降算法对构建的卷积神经网络模型进行训练，得到训练好的模型，具体包括：
    利用所述训练集，通过反向传播和梯度下降算法对构建的卷积神经网络模型进行训练；
    利用交叉验证的方法对每次训练后的卷积神经网络模型进行验证，调整所述卷积神经网络模型训练中的超参数，进入下一次训练；
    当到达训练停止条件时，得到训练好的模型。
  8. 根据权利要求4所述的基于结构磁共振影像的大脑年龄预测方法，其特征在于，所述将待预测的大脑磁共振影像输入所述训练好的模型，得到大脑年龄预测值，之后还包括：
    计算大脑年龄预测值与实际年龄的差值，评估当前大脑偏离健康大脑衰老轨迹的程度。
PCT/CN2020/130073 2019-12-11 2020-11-19 一种基于结构磁共振影像的大脑年龄深度学习预测系统 WO2021115084A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911266178.XA CN110859624A (zh) 2019-12-11 2019-12-11 一种基于结构磁共振影像的大脑年龄深度学习预测系统
CN201911266178.X 2019-12-11

Publications (1)

Publication Number Publication Date
WO2021115084A1 true WO2021115084A1 (zh) 2021-06-17

Family

ID=69656759

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/130073 WO2021115084A1 (zh) 2019-12-11 2020-11-19 一种基于结构磁共振影像的大脑年龄深度学习预测系统

Country Status (2)

Country Link
CN (1) CN110859624A (zh)
WO (1) WO2021115084A1 (zh)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110859624A (zh) * 2019-12-11 2020-03-06 北京航空航天大学 一种基于结构磁共振影像的大脑年龄深度学习预测系统
CN111582215A (zh) * 2020-05-17 2020-08-25 华中科技大学同济医学院附属协和医院 一种胆胰系统正常解剖结构的扫查识别系统和方法
CN111863247B (zh) * 2020-08-03 2022-04-15 北京航空航天大学 基于结构磁共振影像的大脑年龄级联精炼预测方法及系统
CN111968113B (zh) * 2020-09-02 2024-01-19 中国人民解放军国防科技大学 一种基于最优传输映射的脑影像二维卷积深度学习方法
CN112690774B (zh) * 2020-09-29 2022-07-19 首都医科大学附属北京天坛医院 一种基于磁共振影像的卒中复发预测方法和系统
CN112568872B (zh) * 2020-12-30 2021-11-02 深圳大学 基于mri影像和血液生化指标的脑龄融合预测方法
CN113158913A (zh) * 2021-04-25 2021-07-23 安徽科大擎天科技有限公司 一种人脸口罩佩戴识别方法、系统及终端
CN113378898B (zh) * 2021-05-28 2023-05-19 南通大学 基于相对熵损失函数卷积神经网络的脑龄预测方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107067395A (zh) * 2017-04-26 2017-08-18 中国人民解放军总医院 一种基于卷积神经网络的核磁共振图像处理装置与方法
CN109222902A (zh) * 2018-08-27 2019-01-18 上海铱硙医疗科技有限公司 基于核磁共振的帕金森预测方法、系统、存储介质及设备
CN109859189A (zh) * 2019-01-31 2019-06-07 长安大学 一种基于深度学习的年龄估计方法
WO2019125026A1 (ko) * 2017-12-20 2019-06-27 주식회사 메디웨일 심혈관 질병 진단 보조 방법 및 장치
CN109993210A (zh) * 2019-03-05 2019-07-09 北京工业大学 一种基于神经影像的大脑年龄估计方法
CN110859624A (zh) * 2019-12-11 2020-03-06 北京航空航天大学 一种基于结构磁共振影像的大脑年龄深度学习预测系统

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108875624B (zh) * 2018-06-13 2022-03-25 华南理工大学 基于多尺度的级联稠密连接神经网络的人脸检测方法
CN109145920A (zh) * 2018-08-21 2019-01-04 电子科技大学 一种基于深度神经网络的图像语义分割方法
CN109523522B (zh) * 2018-10-30 2023-05-09 腾讯医疗健康(深圳)有限公司 内窥镜图像的处理方法、装置、系统及存储介质
CN109543623B (zh) * 2018-11-26 2021-01-29 微医云(杭州)控股有限公司 一种基于核磁共振成像的胎儿发育状况预测装置
CN109635862B (zh) * 2018-12-05 2021-08-24 合肥奥比斯科技有限公司 早产儿视网膜病plus病变分类方法
CN109712119B (zh) * 2018-12-13 2020-07-03 深圳先进技术研究院 一种磁共振成像及斑块识别方法和装置
CN109949824B (zh) * 2019-01-24 2021-08-03 江南大学 基于N-DenseNet和高维mfcc特征的城市声音事件分类方法
CN109816612A (zh) * 2019-02-18 2019-05-28 京东方科技集团股份有限公司 图像增强方法和装置、计算机可读存储介质
CN110555828A (zh) * 2019-08-08 2019-12-10 北京深睿博联科技有限责任公司 基于3d卷积神经网络的脑龄预测方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107067395A (zh) * 2017-04-26 2017-08-18 中国人民解放军总医院 一种基于卷积神经网络的核磁共振图像处理装置与方法
WO2019125026A1 (ko) * 2017-12-20 2019-06-27 주식회사 메디웨일 심혈관 질병 진단 보조 방법 및 장치
CN109222902A (zh) * 2018-08-27 2019-01-18 上海铱硙医疗科技有限公司 基于核磁共振的帕金森预测方法、系统、存储介质及设备
CN109859189A (zh) * 2019-01-31 2019-06-07 长安大学 一种基于深度学习的年龄估计方法
CN109993210A (zh) * 2019-03-05 2019-07-09 北京工业大学 一种基于神经影像的大脑年龄估计方法
CN110859624A (zh) * 2019-12-11 2020-03-06 北京航空航天大学 一种基于结构磁共振影像的大脑年龄深度学习预测系统

Also Published As

Publication number Publication date
CN110859624A (zh) 2020-03-06

Similar Documents

Publication Publication Date Title
WO2021115084A1 (zh) 一种基于结构磁共振影像的大脑年龄深度学习预测系统
US10706333B2 (en) Medical image analysis method, medical image analysis system and storage medium
JP6947759B2 (ja) 解剖学的対象物を自動的に検出、位置特定、及びセマンティックセグメンテーションするシステム及び方法
CN109272048B (zh) 一种基于深度卷积神经网络的模式识别方法
CN108898175B (zh) 基于深度学习胃癌病理切片的计算机辅助模型构建方法
JP6522161B2 (ja) ディープラーニングに基づく医療データ分析方法及びそのインテリジェントアナライザー
CN108898160B (zh) 基于cnn和影像组学特征融合的乳腺癌组织病理学分级方法
CN109003270B (zh) 一种图像处理方法、电子设备及存储介质
CN109544518B (zh) 一种应用于骨骼成熟度评估的方法及其系统
WO2020215676A1 (zh) 基于残差网络的图像识别方法、装置、设备及存储介质
CN111863247B (zh) 基于结构磁共振影像的大脑年龄级联精炼预测方法及系统
CN113782184A (zh) 一种基于面部关键点与特征预学习的脑卒中辅助评估系统
CN115147600A (zh) 基于分类器权重转换器的gbm多模态mr图像分割方法
CN115018809A (zh) Ct图像的靶区分割识别方法及系统
CN115294075A (zh) 一种基于注意力机制的octa图像视网膜血管分割方法
Sengan et al. Echocardiographic image segmentation for diagnosing fetal cardiac rhabdomyoma during pregnancy using deep learning
CN116205915B (zh) 一种基于掩码的大脑年龄评估方法、系统及电子设备
CN117036288A (zh) 一种面向全切片病理图像的肿瘤亚型诊断方法
CN109191425B (zh) 基于多层神经网络模型医学影像分析方法
CN111612739A (zh) 一种基于深度学习的脑梗死分类方法
CN115937590A (zh) 一种并联融合CNN和Transformer的皮肤病图像分类方法
CN110992309B (zh) 基于深层信息传递网络的眼底图像分割方法
CN114120035A (zh) 一种医学影像识别训练方法
CN113456031A (zh) 一种脑状态预测模型的训练装置、预测装置及电子设备
Perumprath et al. Deep Learning for Segmentation of Brain Tumors using MR Images based on U-Net Architecture

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20897900

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20897900

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205N DATED 06.02.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20897900

Country of ref document: EP

Kind code of ref document: A1