WO2018103538A1 - 一种分析高维医疗数据的深度学习方法和装置 - Google Patents

一种分析高维医疗数据的深度学习方法和装置 Download PDF

Info

Publication number
WO2018103538A1
WO2018103538A1 PCT/CN2017/112606 CN2017112606W WO2018103538A1 WO 2018103538 A1 WO2018103538 A1 WO 2018103538A1 CN 2017112606 W CN2017112606 W CN 2017112606W WO 2018103538 A1 WO2018103538 A1 WO 2018103538A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
feature extraction
deep learning
dimensional medical
medical data
Prior art date
Application number
PCT/CN2017/112606
Other languages
English (en)
French (fr)
Inventor
张荣国
陈宽
Original Assignee
北京推想科技有限公司
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 北京推想科技有限公司 filed Critical 北京推想科技有限公司
Publication of WO2018103538A1 publication Critical patent/WO2018103538A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16ZINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
    • G16Z99/00Subject matter not provided for in other main groups of this subclass
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Definitions

  • the present invention relates to the field of medical artificial intelligence and big data processing, and in particular to a method and apparatus for analyzing high dimensional medical data.
  • the deep learning method is an artificial intelligence method for constructing deep network structure for multi-level feature learning, and has been widely and effectively used in the fields of image recognition and speech recognition.
  • imageNet large-scale image classification
  • the medical field dimension is more than the general application scenario.
  • the diagnosis and treatment data of each patient is complicated, and the deep learning and data analysis models are also larger and more complex than the general ones.
  • the training cost is very high, and the traditional method is deep. Learning and training with big data and machine learning models will consume a lot of manpower and material resources, greatly reducing the economic viability of the application.
  • High-dimensional medical imaging involves multidimensional data such as CT, PET, SPECT, MRI, and fMRI. Even for a single patient, the amount of data is very large. It is time-consuming and laborious for doctors to process such image data. Therefore, the analysis method based on deep learning can be It greatly reduces the workload of doctors and assists doctors in completing patient diagnosis more effectively.
  • the present invention provides a deep learning method and apparatus for analyzing high-dimensional medical data, which can effectively solve the problem of analyzing and processing high-dimensional medical data, so as to improve the practical application of deep learning in the field of processing medical data. effect.
  • a deep learning method for analyzing high-dimensional medical data includes the following steps Step:
  • S1 reading high-dimensional medical data, and preprocessing the high-dimensional medical data into a data format that the feature extraction model can receive by analyzing the data attributes;
  • step S1 the manner of preprocessing the high-dimensional data includes: a fixed length interval extraction method, a fixed length overlap extraction method, and a variable length sequence processing method.
  • the method for extracting the fixed length interval comprises: selecting data of different lengths according to a certain interval, and finally forming a data format of a certain length for use in a subsequent feature extraction model.
  • the fixed length overlap extraction method comprises: for different lengths of data, extracting a certain number of frames from one end of the data as the first group of data, and then extracting the same number of frames as the second group of data, two groups
  • the data partially overlaps and eventually forms a data format of a certain length for use in subsequent feature extraction models.
  • variable length sequence processing method comprises: adding data to a predetermined length of data format for data of different lengths, if the length of the data itself is less than the fixed length, the following data is complemented by 0 to be aligned, Used for subsequent feature extraction models.
  • step S2 for the data format obtained by the fixed length interval extraction method and the variable length sequence processing method, a 2d convolutional neural network or a 2d cyclic neural network is selected, and the data obtained by the fixed length overlapping extraction method is processed.
  • the format selects a 3d convolutional neural network or a 3d cyclic neural network.
  • the invention also relates to a deep learning device for analyzing high-dimensional medical data, comprising: a data pre-processing module, which reads high-dimensional medical data, and preprocesses high-dimensional medical data into a feature extraction model to receive by analyzing data attributes; Data format; the feature extraction module receives the preprocessed data and performs feature extraction to obtain a corresponding feature vector, wherein the feature extraction model is selected according to different preprocessing methods; the model training module receives the corresponding feature Vector and training to get the final depth for medical data analysis Learning model; high-dimensional data for preprocessing methods include: fixed-length interval extraction, fixed-length overlap extraction, and variable-length sequence processing.
  • a 2d convolutional neural network or a 2d cyclic neural network is selected as the feature extraction module, and the data format obtained by the fixed length overlap extraction method is processed.
  • a 3d convolutional neural network or a 3d cyclic neural network is selected as the feature extraction module.
  • the beneficial effects brought by the technical solution provided by the invention are as follows: the accuracy of high-dimensional medical data analysis based on the deep learning model can be improved, and the invention is used for the analysis of medical images, so that the medical imaging intelligent diagnosis has a good application. value.
  • FIG. 1 is a schematic diagram of a deep learning method for analyzing high-dimensional medical data according to an embodiment of the present invention
  • FIG. 2 is a flow chart of a deep learning method for analyzing high dimensional medical data according to an embodiment of the present invention
  • FIG. 3 is a block diagram of a deep learning device that analyzes high dimensional medical data in accordance with an embodiment of the present invention.
  • the present invention provides a deep learning method for analyzing high-dimensional medical data, as shown in FIG. 1, which is a schematic diagram of a deep learning method for analyzing high-dimensional medical data according to an embodiment of the present invention.
  • the high-dimensional medical data is read, and the high-dimensional data is preprocessed into the data format that the feature extraction model can receive by analyzing the data attributes; the feature extraction model extracts the pre-processed data to obtain the feature vector; Come to train to get the final Deep learning model for analysis of medical data.
  • a deep learning method for analyzing high-dimensional medical data of the present invention includes the following steps:
  • S1 Read high-dimensional medical data, and analyze the data attributes to preprocess the high-dimensional medical data into a data format that the feature extraction model can receive.
  • step S1 the manner of preprocessing the high-dimensional data includes: a fixed length interval extraction method, a fixed length overlap extraction method, and a variable length sequence processing method.
  • the method for extracting the fixed length interval comprises: taking different lengths of data, such as CT slice data of different layers, taking the intermediate layer as a reference, and selecting slice data to the two sides at a certain interval, and finally forming a specific length (for example, 100)
  • the data format (100, 1, 512, 512) of the slice layer is used for subsequent feature extraction.
  • the fixed length overlapping extraction method includes: for different lengths of data, extracting a certain number of frames from one end of the data as the first group of data, and then extracting the same number of frames as the second group of data, and the two sets of data have a certain amount.
  • the overlap eventually forming a specific length of data format (100, 50, 512, 512) for subsequent feature extraction.
  • variable length sequence processing method includes: adding, for different lengths of data, all slice layers of the data to a predetermined length of the data format. If the length of the data itself is less than the fixed length, the subsequent data is complemented by 0 for alignment.
  • the final data format is (500, 1, 512, 512) for use in subsequent feature extraction.
  • the feature extraction model is selected, and the feature extraction is performed to obtain the corresponding feature vector.
  • the data format obtained by the fixed length interval extraction method is (100, 1, 512, 512), and the processing method for the variable length sequence processing is performed.
  • the obtained data format is (500, 1, 512, 512), and 2d convolutional neural network (CNN) or 2d cyclic (recursive) neural network (RNN) is selected for feature extraction to obtain a feature vector, such as (100). , 4096) or (500, 4096).
  • CNN convolutional neural network
  • RNN 2d cyclic neural network
  • the feature vector is obtained by a network or 3d loop (recursive) neural network, and its format is (100, 4096).
  • the present invention also relates to a deep learning device for analyzing high-dimensional medical data, the device comprising:
  • the data preprocessing module reads the high dimensional medical data and preprocesses the high dimensional medical data into a data format that the feature extraction model can receive by analyzing the data attributes.
  • the feature extraction module selects different feature extraction models according to different preprocessing methods, and extracts the pre-processed data into the selected feature extraction model for feature extraction to obtain corresponding feature vectors.
  • the model training module inputs the corresponding feature vector into the cyclic neural network model for training, and obtains a final deep learning model for medical data analysis.
  • the preprocessing manner includes: a fixed length interval extraction method, a fixed length overlap extraction method, and a variable length sequence processing method.
  • different feature extraction models are selected for feature extraction according to different preprocessing methods.
  • Deep learning of high-dimensional medical data of lung CT images is performed using the depth learning device proposed by the present invention.
  • Pre-processing module It reads high-dimensional medical data information of CT images of the lungs. Because the CT scan spacing of the lungs is different, the number of all slices per CT is also different. For the CT of (300, 1, 512, 512), the slice size is 512 ⁇ 512, and the number of slices is 300. If 100 frames are to be extracted for analysis, the fixed-length interval extraction method is adopted: one frame is extracted every three frames. For analysis, that is, the fixed interval is 2 frames, so that 100 frames of image data of the feature to be extracted are obtained;
  • Feature extraction module For the fixed-length interval extraction method, select the 2d feature extraction model, and input the pre-processed (100, 1, 512, 512) into the 2d feature extraction model to obtain the feature vector in the (100, 4096) format;
  • Model training module The eigenvectors (N, 100, 4096) with the number of training sets N are input into the cyclic neural network for training to obtain the final deep learning model.
  • Second Embodiment Deep learning is performed on high-dimensional medical data of a head MRI image using the depth learning device proposed by the present invention.
  • Feature extraction module For the data processed by the fixed length overlap extraction method, select the 3d feature extraction model, and input the preprocessed (100, 20, 512, 512) into the 3d feature extraction model to obtain the (100, 4096) format.
  • Feature vector For the data processed by the fixed length overlap extraction method, select the 3d feature extraction model, and input the preprocessed (100, 20, 512, 512) into the 3d feature extraction model to obtain the (100, 4096) format.
  • Model training module The feature vector (N, 100, 4096) with the number of training sets N is input into the cyclic neural network for training to obtain the final deep learning model.
  • Deep learning is performed on high-dimensional medical data of lung CT images using the depth learning device proposed by the present invention.
  • Data processing module It reads high-dimensional medical data information of CT images of the lungs. Because the CT scan spacing of the lungs is different, the number of slices per CT is also different. For CTs of different lengths such as (300, 1, 512, 512), (416, 1, 512, 512), (200, 1, 512, 512), the slice size is 512 ⁇ 512, and the number of slices is different.
  • the variable length sequence processing method performs processing: 500 is used as the length of different length CT data processing, less than 500 frames are complemented by 0 alignment, and finally, the different lengths of CT processing are obtained in (500, 1, 512, 512) format to be extracted. Characteristic data
  • Feature extraction module For the data processed by the variable length sequence processing method, select the 2d feature extraction model, and input the preprocessed (500, 1, 512, 512) into the 2d feature extraction model to obtain the (500, 4096) format.
  • Feature vector For the data processed by the variable length sequence processing method, select the 2d feature extraction model, and input the preprocessed (500, 1, 512, 512) into the 2d feature extraction model to obtain the (500, 4096) format.
  • Model training module The feature vector (N, 500, 4096) with the number of training sets N is input into the cyclic neural network for training to obtain the final deep learning model.

Abstract

本申请公开了一种分析高维医疗数据的深度学习方法和装置。所述方法包括:读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式;根据不同的预处理的方式,来选择不同的特征提取模型,将预处理后的数据输入选择的特征提取模型进行特征提取,得到相应的特征向量;将相应的特征向量输入至循环神经网络模型进行训练,得到用于医疗数据分析的最终深度学习模型。该方法能够提高基于深度学习模型进行高维医疗数据分析的准确率。本发明还包括一种分析高维医疗数据的深度学习装置,其包括:预处理模块、特征提取模块和模型训练模块。

Description

一种分析高维医疗数据的深度学习方法和装置 技术领域
本发明涉及医疗人工智能与大数据处理领域,特别涉及分析高维医疗数据的方法和装置。
背景技术
近年来人工智能技术蓬勃发展,随着以深度学习框架为内核的新人工智能技术强势崛起,在各个领域都获得了长足的发展与推进,AlphaGo、无人驾驶车、语音识别等人们期盼多年的技术也都在很短的时间内获得了突破。在可见的未来当中,深度学习也将推动医疗行业的大数据分析与人工智能应用的发展,深度学习方法在改变医疗健康方面有很大的潜力。深度学习方法是通过构建深层网络结构进行多层次特征学习的人工智能方法,已经广泛而有效地用于图像识别、语音识别等领域。随着深度学习方法在大规模图片分类(ImageNet)上取得突破性进展,深度学习在各个方面都受到了极大的关注,在图像识别与语音识别领域都取得了很大的成功。
具体到医疗行业,医疗领域维度本来就比一般应用场景要多,每一个病人的诊疗数据纷繁复杂,深度学习和数据分析模型也比一般的要庞大和复杂,训练成本非常高,传统方法对深度学习和大数据、机器学习模型进行学习和训练将耗费大量的人力与物力,大大降低该应用的经济可行性。高维医疗影像涉及到CT、PET、SPECT、MRI、fMRI等多维数据,即使对单个病人来说,数据量就非常大,医生对处理这样的影像数据费时费力,因此基于深度学习的分析方法可以大大减少医生的工作量,辅助医生更有效地完成病人的诊断。
发明内容
为克服现有技术的不足,本发明提供了一种分析高维医疗数据的深度学习方法和装置,其可以有效解决高维医疗数据的分析处理问题,以提升深度学习在处理医疗数据领域的实用效果。
本发明的一种分析高维医疗数据的深度学习方法,其包括如下步 骤:
S1:读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式;
S2:根据不同的预处理的方式,来选择不同的特征提取模型,将预处理后的数据通过选择的特征提取模型进行特征提取,得到相应的特征向量;
S3:对相应的特征向量进行训练,得到用于医疗数据分析的最终深度学习模型;
在步骤S1中,将高维数据进行预处理的方式包括:定长间隔抽取方式、定长重叠抽取方式以及变长序列加工方式。
优选地,所述定长间隔抽取方式包括:将不同长度的数据,按一定的间隔选取数据,最终组成一定长度的数据格式,以备后续的特征提取模型所用。
优选地,所述定长重叠抽取方式包括:对于不同长度的数据,从数据的一端开始抽取一定数量的帧数作为第一组数据,然后抽取相同数量的帧数作为第二组数据,两组数据有部分重叠,最终组成一定长度的数据格式,以备后续的特征提取模型所用。
优选地,所述变长序列处理方式包括:对于不同长度的数据,将数据都加入到事先确定的一定长度的数据格式里面,如果数据本身长度小于该定长,后面数据补0进行对齐,以备后续的特征提取模型所用。
优选地,在步骤S2中,对于定长间隔抽取方式和变长序列处理方式处理得到的数据格式,选择2d的卷积神经网络或者2d的循环神经网络,对于定长重叠抽取方式处理得到的数据格式选择3d的卷积神经网络或者3d的循环神经网络。
本发明还涉及一种分析高维医疗数据的深度学习装置,其包括:数据预处理模块,其读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式;特征提取模块,其接收预处理后的数据并进行特征提取,得到相应的特征向量,其中,特征提取模型根据不同的预处理的方式来选择;模型训练模块,其接收相应的特征向量并进行训练,得到用于医疗数据分析的最终深度学 习模型;高维数据进行预处理的方式包括:定长间隔抽取方式、定长重叠抽取方式以及变长序列加工方式。
优选地,对于定长间隔抽取方式和变长序列处理方式处理得到的数据格式,选择2d的卷积神经网络或者2d的循环神经网络作为特征提取模块,对于定长重叠抽取方式处理得到的数据格式选择3d的卷积神经网络或者3d的循环神经网络作为特征提取模块。
本发明提供的技术方案带来的有益效果体现在:可以提高基于深度学习模型进行高维医疗数据分析的准确率,从而用于医疗影像的分析,使得在医学影像智能诊断方面具有很好的应用价值。
附图说明
为了更清楚地说明本发明实施例中的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍。显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来说,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1是根据本发明实施例的分析高维医疗数据的深度学习方法的示意图;
图2是根据本发明实施例的分析高维医疗数据的深度学习方法的流程图;
图3是根据本发明实施例的分析高维医疗数据的深度学习装置的框图。
具体实施方式
为使本发明的目的、技术方案和优点更加清楚,下面将结合附图对本发明的实施方式做进一步地详细描述。
本发明提供了一种分析高维医疗数据的深度学习方法,如图1所示,其为根据本发明实施例的分析高维医疗数据的深度学习方法的示意图。
首先读取高维医疗数据,通过分析数据属性而将高维数据预处理为特征提取模型能够接收的数据格式;特征提取模型对预处理后的数据进行特征提取,从而得到特征向量;用特征向量来训练得到最终的 深度学习模型,从而用于医疗数据的分析。
如图2所示,本发明的一种分析高维医疗数据的深度学习方法包括如下步骤:
S1:读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式。
S2:根据不同的预处理的方式,来选择不同的特征提取模型,将预处理后的数据输入选择的特征提取模型进行特征提取,得到相应的特征向量。
S3:将相应的特征向量输入至循环神经网络模型进行训练,得到用于医疗数据分析的最终深度学习模型。
在步骤S1中,将高维数据进行预处理的方式包括:定长间隔抽取方式、定长重叠抽取方式以及变长序列加工方式。
所述定长间隔抽取方式包括:将不同长度的数据,如不同层数的CT切片数据,取中间层为基准,按一定的间隔向两侧分别选取切片数据,最终组成特定长度(例如100个切片层)的数据格式(100,1,512,512),以备后续的特征提取所用。
所述定长重叠抽取方式包括:对于不同长度的数据,从数据的一端开始抽取一定数量的帧数作为第一组数据,然后抽取相同数量的帧数作为第二组数据,两组数据有一定量的重叠,最终组成特定长度的数据格式(100,50,512,512),以备后续的特征提取所用。
所述变长序列加工方式包括:对于不同长度的数据,将数据所有切片层都加入到事先确定的一定长度的数据格式里面,如果数据本身长度小于该定长,后面数据补0进行对齐。最终的数据格式为(500,1,512,512),以备后续的特征提取所用。
根据不同的数据格式选择特征提取模型,进行特征提取得到相应的特征向量,其中,对于定长间隔抽取方式处理得到的数据格式如(100,1,512,512),对于变长序列加工方式处理得到的数据格式如(500,1,512,512),选择2d的卷积神经网络(CNN)或者2d的循环(递归)神经网络(RNN)等进行特征提取得到特征向量,其格式如(100,4096)或者(500,4096)。对于定长重叠抽取方式整理得到的数据格式如(100,50,512,512),选择3d的卷积神经网 络或者3d的循环(递归)神经网络进行特征抽取得到特征向量,其格式如(100,4096)。
如图2所示,本发明还涉及一种分析高维医疗数据的深度学习装置,所述装置包括:
数据预处理模块,其读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式。
特征提取模块,其根据不同的预处理的方式,来选择不同的特征提取模型,将预处理后的数据输入选择的特征提取模型进行特征提取,得到相应的特征向量。
模型训练模块,其将相应的特征向量输入至循环神经网络模型进行训练,得到用于医疗数据分析的最终深度学习模型。
在数据预处理模块中,所述预处理的方式包括:定长间隔抽取方式、定长重叠抽取方式以及变长序列加工方式。
在特征提取模块中,根据不同的预处理的方式,而选择不同的特征提取模型进行特征提取。
下面以示例来进行说明。
第一实施例:利用本发明提出的深度学习装置来对肺部CT图像的高维医疗数据进行深度学习。
预处理模块:其读取肺部CT图像的高维医疗数据信息,由于肺部CT扫描间距不同,因此每隔CT所有的切片数目也有所不同。对于(300,1,512,512)的CT,切片大小是512×512,切片数目为300,如果要抽取其中的100帧做分析,采用定长间隔抽取方式进行处理:每3帧抽取1帧用来分析,即固定间隔为2帧,这样得到待提取特征的100帧图像数据;
特征提取模块:对于定长间隔抽取方式,选择2d的特征提取模型,将预处理得到的(100,1,512,512)输入2d的特征提取模型,得到(100,4096)格式的特征向量;
模型训练模块:将训练集数目为N的特征向量(N,100,4096)输入到循环神经网络中进行训练得到最终的深度学习模型。
第二实施例:利用本发明提出的深度学习装置来对头部MRI图像的高维医疗数据进行深度学习。
数据加工模块:其读取头部MRI图像的高维医疗数据信息;由于扫描层间距的不同,每例MRI的切片数目也有所不同,例如对于(416,1,512,512)的MRI,切片大小是512*512,切片数目为416,采用定长重叠抽取方式进行处理:从数据一端开始抽取20帧作为第一组数据,移动步长4,取20帧作为第二组数据,即每组数据重叠16帧,这样处理得到(416-20)/4+1=100组待提取特征的数据;
特征提取模块:对于定长重叠抽取方式处理的数据,选择3d的特征提取模型,将预处理得到的(100,20,512,512)输入3d的特征提取模型,得到(100,4096)格式的特征向量;
模型训练模块:将训练集数目为N的特征向量(N,100,4096)输入到循环神经网络中进行训练得到最终深度学习模型。
第三实施例:利用本发明提出的深度学习装置来对肺部CT图像的高维医疗数据进行深度学习。
数据加工模块:其读取肺部CT图像的高维医疗数据信息,由于肺部CT扫描间距不同,因此每隔CT所有的切片数目也有所不同。对于(300,1,512,512)、(416,1,512,512)、(200,1,512,512)等不同长度的CT,切片大小是512×512,切片数目各不相同,用变长序列处理方式进行处理:将500作为不同长度CT数据处理的长度,不足500帧的进行补0对齐,最后将上述不同长度的CT处理得到(500,1,512,512)格式的待提取特征的数据;
特征提取模块:对于变长序列处理方式处理的数据,选择2d的特征提取模型,将预处理得到的(500,1,512,512)输入2d的特征提取模型中,得到(500,4096)格式的特征向量;
模型训练模块:将训练集数目为N的特征向量(N,500,4096)输入到循环神经网络中进行训练得到最终深度学习模型。
以上详细描述了本发明的具体实施例,但可以理解,在不脱离本发明的精神下可以对其做出修改。本发明的权利要求旨在覆盖这些修改,以保证其落入本发明的真实范围和精神内。

Claims (7)

  1. 一种分析高维医疗数据的深度学习方法,其特征在于包括如下步骤:
    S1:读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式;
    S2:根据不同的预处理的方式,来选择不同的特征提取模型,将预处理后的数据通过选择的特征提取模型进行特征提取,得到相应的特征向量;
    S3:对相应的特征向量进行训练,得到用于医疗数据分析的最终深度学习模型;
    在步骤S1中,将高维数据进行预处理的方式包括:定长间隔抽取方式、定长重叠抽取方式以及变长序列加工方式。
  2. 根据权利要求1所述的分析高维医疗数据的深度学习方法,其特征在于,所述定长间隔抽取方式包括:
    将不同长度的数据,按一定的间隔选取数据,最终组成一定长度的数据格式,以备后续的特征提取模型所用。
  3. 根据权利要求1所述的分析高维医疗数据的深度学习方法,其特征在于,所述定长重叠抽取方式包括:
    对于不同长度的数据,从数据的一端开始抽取一定数量的帧数作为第一组数据,然后抽取相同数量的帧数作为第二组数据,两组数据有部分重叠,最终组成一定长度的数据格式,以备后续的特征提取模型所用。
  4. 根据权利要求1所述的分析高维医疗数据的深度学习方法,其特征在于,所述变长序列处理方式包括:
    对于不同长度的数据,将数据都加入到事先确定的一定长度的数据格式里面,如果数据本身长度小于该定长,后面数据补0进行对齐,以备后续的特征提取模型所用。
  5. 根据权利要求1所述的分析高维医疗数据的深度学习方法,其特征在于,在步骤S2中,对于定长间隔抽取方式和变长序列处理方式处理得到的数据格式,选择2d的卷积神经网络或者2d的循环神经网络,对于定长重叠抽取方式处理得到的数据格式选择3d的卷积神经网络或者3d的循环神经网络。
  6. 一种分析高维医疗数据的深度学习装置,其特征在于包括:
    数据预处理模块,其读取高维医疗数据,通过分析数据属性,将高维医疗数据预处理为特征提取模型能够接收的数据格式;
    特征提取模块,其接收预处理后的数据并进行特征提取,得到相应的特征向量,其中,特征提取模型根据不同的预处理的方式来选择;
    模型训练模块,其接收相应的特征向量并进行训练,得到用于医疗数据分析的最终深度学习模型;
    将高维数据进行预处理的方式包括:定长间隔抽取方式、定长重叠抽取方式以及变长序列加工方式。
  7. 根据权利要求6所述的分析高维医疗数据的深度学习装置,其中,对于定长间隔抽取方式和变长序列处理方式处理得到的数据格式,选择2d的卷积神经网络或者2d的循环神经网络作为特征提取模块,对于定长重叠抽取方式处理得到的数据格式选择3d的卷积神经网络或者3d的循环神经网络作为特征提取模块。
PCT/CN2017/112606 2016-12-08 2017-11-23 一种分析高维医疗数据的深度学习方法和装置 WO2018103538A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201611122716.4 2016-12-08
CN201611122716.4A CN108198625B (zh) 2016-12-08 2016-12-08 一种分析高维医疗数据的深度学习方法和装置

Publications (1)

Publication Number Publication Date
WO2018103538A1 true WO2018103538A1 (zh) 2018-06-14

Family

ID=62490657

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/112606 WO2018103538A1 (zh) 2016-12-08 2017-11-23 一种分析高维医疗数据的深度学习方法和装置

Country Status (2)

Country Link
CN (1) CN108198625B (zh)
WO (1) WO2018103538A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960422A (zh) * 2018-06-19 2018-12-07 河南工业大学 一种基于主成分分析的宽度学习方法
CN109902734A (zh) * 2019-02-22 2019-06-18 中国石油大学(华东) 一种基于yolo卷积神经网络的胆石病ct医疗图像快速识别方法
CN112330644A (zh) * 2020-11-11 2021-02-05 复旦大学 基于深度学习的医疗影像诊断系统
CN113156243A (zh) * 2021-04-09 2021-07-23 南方电网电动汽车服务有限公司 故障预测方法和预测系统

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109300531A (zh) * 2018-08-24 2019-02-01 深圳大学 一种脑疾病早期诊断方法和装置
CN109448855A (zh) * 2018-09-17 2019-03-08 大连大学 一种基于cnn和模型融合的糖尿病血糖预测方法
CN109620203A (zh) * 2018-12-18 2019-04-16 安徽心之声医疗科技有限公司 一种基于一维卷积神经网络的心电信号特征自动提取方法
CN110222087B (zh) * 2019-05-15 2023-10-17 平安科技(深圳)有限公司 特征提取方法、装置及计算机可读存储介质
CN110751183A (zh) * 2019-09-24 2020-02-04 东软集团股份有限公司 影像数据分类模型的生成方法、影像数据分类方法及装置
CN110826908A (zh) * 2019-11-05 2020-02-21 北京推想科技有限公司 人工智能预测的评估方法、装置、存储介质及电子设备
CN111243658B (zh) * 2020-01-07 2022-07-22 西南大学 一种基于深度学习的生物分子网络构建与优化方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5852684A (en) * 1993-10-15 1998-12-22 Panasonic Technologies, Inc. Multimedia rendering marker and method
CN105760507A (zh) * 2016-02-23 2016-07-13 复旦大学 基于深度学习的跨模态主题相关性建模方法
CN106096616A (zh) * 2016-06-08 2016-11-09 四川大学华西医院 一种基于深度学习的磁共振影像特征提取及分类方法
CN106097335A (zh) * 2016-06-08 2016-11-09 安翰光电技术(武汉)有限公司 消化道病灶图像识别系统及识别方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106033554A (zh) * 2015-03-13 2016-10-19 中国科学院声学研究所 一种基于分段的两级深度学习模型的大数据处理方法
CN113421652A (zh) * 2015-06-02 2021-09-21 推想医疗科技股份有限公司 对医疗数据进行分析的方法、训练模型的方法及分析仪
CN105718952B (zh) * 2016-01-22 2018-10-30 武汉科恩斯医疗科技有限公司 使用深度学习网络对断层医学影像进行病灶分类的系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5852684A (en) * 1993-10-15 1998-12-22 Panasonic Technologies, Inc. Multimedia rendering marker and method
CN105760507A (zh) * 2016-02-23 2016-07-13 复旦大学 基于深度学习的跨模态主题相关性建模方法
CN106096616A (zh) * 2016-06-08 2016-11-09 四川大学华西医院 一种基于深度学习的磁共振影像特征提取及分类方法
CN106097335A (zh) * 2016-06-08 2016-11-09 安翰光电技术(武汉)有限公司 消化道病灶图像识别系统及识别方法

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960422A (zh) * 2018-06-19 2018-12-07 河南工业大学 一种基于主成分分析的宽度学习方法
CN109902734A (zh) * 2019-02-22 2019-06-18 中国石油大学(华东) 一种基于yolo卷积神经网络的胆石病ct医疗图像快速识别方法
CN112330644A (zh) * 2020-11-11 2021-02-05 复旦大学 基于深度学习的医疗影像诊断系统
CN113156243A (zh) * 2021-04-09 2021-07-23 南方电网电动汽车服务有限公司 故障预测方法和预测系统

Also Published As

Publication number Publication date
CN108198625B (zh) 2021-07-20
CN108198625A (zh) 2018-06-22

Similar Documents

Publication Publication Date Title
WO2018103538A1 (zh) 一种分析高维医疗数据的深度学习方法和装置
Bhatt et al. CNN variants for computer vision: History, architecture, application, challenges and future scope
CN107273800B (zh) 一种基于注意机制的卷积递归神经网络的动作识别方法
CN110097131B (zh) 一种基于对抗协同训练的半监督医疗图像分割方法
CN110532900B (zh) 基于U-Net和LS-CNN的人脸表情识别方法
Zhou et al. Contextual ensemble network for semantic segmentation
WO2018024031A1 (zh) 一种对医疗图像进行转换学习的方法和装置
CN110443231A (zh) 一种基于人工智能的单手手指点读文字识别方法和系统
CN110992351B (zh) 基于多输入卷积神经网络的sMRI图像分类方法和装置
CN107239801A (zh) 视频属性表示学习方法及视频文字描述自动生成方法
CN109377520A (zh) 基于半监督循环gan的心脏图像配准系统及方法
CN108090403A (zh) 一种基于3d卷积神经网络的人脸动态识别方法及系统
CN104484886B (zh) 一种mr图像的分割方法及装置
CN108647599B (zh) 结合3d跃层连接和循环神经网络的人体行为识别方法
Sun et al. AMPNet: Average-and max-pool networks for salient object detection
CN112766220B (zh) 双通道微表情识别方法、系统、存储介质、计算机设备
CN107424145A (zh) 基于三维全卷积神经网络的核磁共振图像的分割方法
CN112307995A (zh) 一种基于特征解耦学习的半监督行人重识别方法
CN105095857A (zh) 基于关键点扰动技术的人脸数据增强方法
Varadarajan et al. An efficient deep convolutional neural network approach for object detection and recognition using a multi-scale anchor box in real-time
Li et al. End-to-end training for compound expression recognition
CN112508121A (zh) 一种工业机器人感知外界的方法和系统
Grigorescu et al. Interpretable convolutional neural networks for preterm birth classification
Zhang et al. Learn to walk across ages: Motion augmented multi-age group gait video translation
Miao et al. Bone age assessment based on SuperPoint features

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: 17879006

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: 17879006

Country of ref document: EP

Kind code of ref document: A1