WO2021129234A1 - 心房间隔闭塞患者心脏医学影像分割方法及系统 - Google Patents

心房间隔闭塞患者心脏医学影像分割方法及系统 Download PDF

Info

Publication number
WO2021129234A1
WO2021129234A1 PCT/CN2020/129400 CN2020129400W WO2021129234A1 WO 2021129234 A1 WO2021129234 A1 WO 2021129234A1 CN 2020129400 W CN2020129400 W CN 2020129400W WO 2021129234 A1 WO2021129234 A1 WO 2021129234A1
Authority
WO
WIPO (PCT)
Prior art keywords
data set
mri
image segmentation
medical image
segmentation
Prior art date
Application number
PCT/CN2020/129400
Other languages
English (en)
French (fr)
Inventor
黄建龙
吴剑煌
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2021129234A1 publication Critical patent/WO2021129234A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/174—Segmentation; Edge detection involving the use of two or more images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10088—Magnetic resonance imaging [MRI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30048—Heart; Cardiac
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • the invention relates to a method and system for segmenting cardiac medical images of patients with atrial septal occlusion.
  • Atrial septal defect In the medical image analysis of atrial septal defect (ASD), the right atrium shows severe swelling due to the blood inflow caused by the septal defect from the left atrium and causes the blood volume imbalance in the two atria. Magnetic Resonance Imaging (MRI) is usually used to diagnose this kind of heart disease.
  • MRI Magnetic Resonance Imaging
  • the introduction of the metal atrial septal occluder causes a ghosting effect in the area where it is placed, resulting in incorrect active contour segmentation.
  • Kucera et al. have achieved reliable active contour 3D models on the short-axis and long-axis views of the heart. He proposed a region-based external force to segment the left ventricle. Sarti et al. proposed a region-based segmentation model method to realize the prior knowledge of the gray-level statistical distribution. They used the level set method to drive the curve evolution to obtain the maximum likelihood segmentation of the target relative to the statistical distribution of image pixels. . Boukerroui et al. proposed another region-based segmentation method, based on an adaptive segmentation algorithm, in which the weighting function considers local and global statistics. Mishra et al.
  • the classic snake model was originally proposed by Kass, Witkin, and Terzopolous to synthesize the noise filter response generated by the edge detector into a coherent depiction of the perceptual edge in the image. In this way, a boundary separating two image regions with different gray-level characteristics can be established.
  • the semi-automatic segmentation method is used to segment heart chambers based on MR images, and is implemented by the Kass snake algorithm, which involves a region-based method for segmentation. It can overcome the common problem of segmented objects with poor boundary definition in ultrasound imaging, but it cannot complete the segmentation of MRI images for segmented objects with inconspicuous boundary definitions.
  • traditional methods active contour models
  • traditional methods cannot better solve the relatively small size of training data in the research.
  • the present invention provides a cardiac medical image segmentation method for patients with atrial septal occlusion.
  • the method includes the following steps: a. Collect a cardiac MRI data set of patients with atrial septal occlusion, and process the MRI data set using a spectral analysis method; b. Process the MRI data set by a spectral analysis method Data enhancement is performed on the MRI data set after data enhancement, and the MRI data set after data enhancement is segmented by binary classification to obtain a correctly divided MRI data set; c. For the correctly divided MRI training data set, the transfer learning method is adopted Fine-tune the convolutional neural network model to extract features useful for subsequent medical image segmentation; d. Use the extracted features useful for medical image segmentation to design the U-Net architecture, and use the U-Net architecture to complete end-to-end pixel-to-pixel Medical image segmentation.
  • step b specifically includes:
  • the MRI data set after data enhancement is segmented, and the segmentation is regarded as a binary classification, that is, 0 and 1, 1 means that it is divided correctly, and 0 means that it is divided incorrectly.
  • Said step c specifically includes:
  • the step d specifically includes the following steps:
  • the left half of the U-Net architecture is the encoder part, and the encoder captures the contraction path of the context and performs feature extraction;
  • the right half of the U-Net architecture is the decoder part, and the decoder performs precise positioning of the symmetric extension path;
  • the segmentation results include:
  • the present invention provides a cardiac medical image segmentation system for patients with atrial septal occlusion.
  • the system includes an acquisition module, a data set partitioning module, a fine-tuning module, and an image segmentation module.
  • the acquisition module is used to acquire cardiac MRI data sets for patients with atrial septal occlusion.
  • the data set dividing module is used for data enhancement of the MRI data set processed by the spectral analysis method, and the MRI data set after the data enhancement is segmented into binary classification to obtain The correctly divided MRI data set;
  • the fine-tuning module adopts the transfer learning method to fine-tune the convolutional neural network model according to the correctly divided MRI training data set to extract the features useful for subsequent medical image segmentation;
  • the image segmentation module It is used to design the U-Net architecture using extracted features useful for medical image segmentation, and use the U-Net architecture to complete end-to-end pixel-to-pixel medical image segmentation.
  • the data set dividing module is specifically used for:
  • the MRI data set after data enhancement is segmented, and the segmentation is regarded as a binary classification, that is, 0 and 1, 1 means that it is divided correctly, and 0 means that it is divided incorrectly.
  • the fine-tuning module is specifically used for:
  • the image segmentation module is specifically used for:
  • the left half of the U-Net architecture is the encoder part, and the encoder captures the contraction path of the context and performs feature extraction;
  • the right half of the U-Net architecture is the decoder part, and the decoder performs precise positioning of the symmetric extension path;
  • the segmentation results include:
  • the invention can improve the efficiency of diagnosis in cardiovascular MRI examination, treat medical image segmentation as a binary classification problem, and use transfer learning to solve the problem of insufficient training data in the training phase of the convolutional neural network when the training data scale is relatively small.
  • the complete convolutional network of the Net framework can accurately segment the above-mentioned special MRI images, and can achieve end-to-end classification of the target and background of cardiac MRI images more efficiently.
  • Fig. 1 is a flowchart of a method for segmenting a cardiac medical image of a patient with atrial septal occlusion according to the present invention
  • FIG. 2 is a schematic diagram of processing a cardiac MRI data set of patients with atrial septal occlusion by using a spectral analysis method according to an embodiment of the present invention
  • FIG. 3 is a schematic diagram of a deep learning convolutional neural network VGG 16 used in migration learning according to an embodiment of the present invention
  • VGG 16 is a schematic diagram of pre-training results of the deep learning convolutional neural network VGG 16 provided by an embodiment of the present invention
  • FIG. 5 is a schematic diagram of the structure of a convolutional neural network based on the U-Net framework provided by an embodiment of the present invention
  • Fig. 6 is a hardware architecture diagram of a cardiac medical image segmentation system for patients with atrial septal occlusion according to the present invention.
  • FIG. 1 it is a flowchart of a preferred embodiment of a method for segmenting a cardiac medical image of a patient with atrial septal occlusion according to the present invention.
  • Step S1 Collect a cardiac MRI data set of patients with atrial septal occlusion, process the MRI data set using a spectral analysis method (see also Figure 2), and divide the MRI data set into a training set and a test set.
  • a spectral analysis method see also Figure 2
  • step S2 data enhancement is performed on the MRI data set processed by the spectral analysis method, and the MRI data set after the data enhancement is segmented by binary classification to obtain a correctly divided MRI data set and an incorrectly divided MRI data set.
  • step S21 data enhancement is performed on the MRI data set by a data enhancement method, horizontal and vertical sliding, random cropping, and color jitter and Gaussian noise are added. Including the following steps:
  • Step S211 sliding the data of the MRI data set horizontally and vertically;
  • Step S212 randomly crop the data of the MRI data set
  • the slice image is randomly cropped using the random crop function tf.random_crop in tensorflow, and the slice image is cropped to 2/3 of its size, size, width, and height.
  • step S22 the MRI data set after the data enhancement is segmented, and the segmentation is regarded as a binary classification, that is, 0 and 1, where 1 represents that it is divided correctly, and 0 represents that it is divided incorrectly.
  • Step S3 for the correctly divided MRI training data set, the transfer learning method is used to fine-tune the convolutional neural network model to extract features useful for subsequent medical image segmentation.
  • the transfer learning method is used to fine-tune the convolutional neural network model to extract features useful for subsequent medical image segmentation.
  • Step S31 using the transfer learning method to select the pre-trained model of the deep learning convolutional neural network VGG 16 (see also Figure 3) as the encoder of the U-Net network;
  • Step S32 using the pre-training model of the deep learning convolutional neural network VGG 16 to initialize the ImageNet weights;
  • step S33 the method of fine-tuning is adopted in the training process to modify the output category of the last layer of the deep learning convolutional neural network VGG 16 pre-training model, and speed up the parameter learning rate of the last layer; adjust the configuration parameters of Solver, which is in this embodiment
  • the pre-training results of the deep learning convolutional neural network VGG16 are shown in Figure 4:
  • One of the fine-tuning is a deep learning method, which is to continuously adjust the parameters of the network to maximize the performance of the convolutional network. Because the prerequisite for fine-tuning is the weight of the pre-trained model with meaningful values. When the learning rate is large, the weights will be updated quickly and destroy the original training network structure learning rate. In this embodiment, the learning rate is set to 1 ⁇ 10 -4 .
  • Step S4 using the extracted features useful for medical image segmentation to design a U-Net architecture (please also refer to Figure 5), and using the U-Net architecture to complete end-to-end pixel-to-pixel medical image segmentation. in particular:
  • Step S41 the left half of the U-Net architecture is the encoder part.
  • the encoder captures the contraction path of the context and performs feature extraction, which specifically includes:
  • Step S411 the network of the encoder adopts the deep learning convolutional neural network VGG 16 for feature extraction;
  • Step S412 the encoder partially removes the fully connected layer and replaces it with a single convolutional layer of 512 channels;
  • Step S42 the right half of the U-Net architecture is the decoder part, and the decoder performs precise positioning of the symmetric extension path, which specifically includes:
  • Step S421 the decoder part uses the transposed convolutional layer to construct, so that the size of the feature map is doubled, and the number of channels is reduced by half at the same time;
  • Step S422 connecting the output of the transposed convolution to the decoder for output;
  • Step S423 the up-sampling process is repeated 5 times to match the 5 pools with the largest output feature map size; the custom loss function is:
  • y i is the correct answer of the i-th data in a batch
  • y i ' is the predicted value obtained by the neural network
  • x is the actual value
  • y is the predicted value
  • a and b are constants.
  • step S43 the medical image is segmented using the encoder part and the decoder part of the U-Net architecture, and the segmentation result obtained specifically includes:
  • TP True positives
  • TN True negatives
  • the similarity of the segmented images is evaluated by the following commonly used metrics in medical image segmentation:
  • the metrics include: dice index, accuracy, and Jaccard similarity coefficient.
  • the Jaccard similarity coefficient is used to compare the similarity and difference between the limited sample sets. The larger the Jaccard coefficient value, the higher the sample similarity.
  • FIG. 6 is a hardware architecture diagram of the cardiac medical image segmentation system 10 for patients with atrial septal occlusion according to the present invention.
  • the system includes: an acquisition module 101, a data set division module 102, a fine-tuning module 103, and an image segmentation module 104.
  • the acquisition module 101 is used to acquire a cardiac MRI data set of patients with atrial septal occlusion, process the MRI data set using a spectral analysis method (see also FIG. 2), and divide the MRI data set into a training set and a test set. in particular:
  • the data set dividing module 102 is used for data enhancement of the MRI data set processed by the spectral analysis method, and the MRI data set after the data enhancement is divided into binary classification, so as to obtain a correctly divided MRI data set and a wrong MRI data set.
  • Divided MRI data set in particular:
  • the data set dividing module 102 adopts a data enhancement method to perform data enhancement on the MRI data set, sliding horizontally and vertically, cutting randomly, and adding color jitter and Gaussian noise.
  • the slice image is randomly cropped using the random crop function tf.random_crop in tensorflow, and the slice image is cropped to 2/3 of its size, size, width, and height;
  • the data set dividing module 102 divides the data-enhanced MRI data set, and regards the division as a binary classification, that is, 0 and 1, where 1 represents the correct division, and 0 represents the wrong division.
  • the fine-tuning module 103 is used to fine-tune the convolutional neural network model for the correctly divided MRI training data set by using a transfer learning method to extract features useful for subsequent medical image segmentation. in particular:
  • the fine-tuning module 103 uses the transfer learning method to select the pre-trained model of the deep learning convolutional neural network VGG 16 (see also Figure 3) as the encoder of the U-Net network;
  • the fine-tuning module 103 uses the pre-training model of the deep learning convolutional neural network VGG 16 to initialize the ImageNet weights;
  • the fine-tuning module 103 adopts a fine-tuning method during the training process to modify the output category of the last layer of the deep learning convolutional neural network VGG 16 pre-training model, and speed up the parameter learning rate of the last layer; adjust the configuration parameters of Solver, this implementation
  • the pre-training results of the deep learning convolutional neural network VGG 16 of the example are shown in Figure 4:
  • One of the fine-tuning is a deep learning method, which is to continuously adjust the parameters of the network to maximize the performance of the convolutional network. Because the prerequisite for fine-tuning is the weight of the pre-trained model with meaningful values. When the learning rate is large, the weights will be updated quickly and destroy the original training network structure learning rate. In this embodiment, the learning rate is set to 1 ⁇ 10 -4 .
  • the image segmentation module 104 is used to design a U-Net architecture (please also refer to FIG. 5) using the extracted features useful for medical image segmentation, and use the U-Net architecture to complete end-to-end pixel-to-pixel medical image segmentation. in particular:
  • the image segmentation module 104 performs feature extraction, that is, the left half of the U-Net architecture is the encoder part.
  • the encoder captures the contraction path of the context and performs feature extraction, which specifically includes:
  • the network of the encoder adopts the deep learning convolutional neural network VGG 16 for feature extraction
  • the encoder partly removed the fully connected layer and replaced it with a single convolutional layer of 512 channels;
  • the image segmentation module 104 constructs the decoder part of the U-Net architecture.
  • the right half of the U-Net architecture is the decoder part.
  • the decoder performs precise positioning of the symmetric extension path, which specifically includes:
  • the image segmentation module 104 uses the transposed convolutional layer to construct the decoder part, so that the size of the feature map is doubled and the number of channels is reduced by half;
  • the image segmentation module 104 connects the output of the transposed convolution to the decoder for output;
  • the image segmentation module 104 repeats the up-sampling process 5 times to match 5 pools with the largest output feature map size; the custom loss function is:
  • y i is the correct answer of the i-th data in a batch
  • y i ' is the predicted value obtained by the neural network
  • x is the actual value
  • y is the predicted value
  • a and b are constants.
  • the image segmentation module 104 uses the encoder part and the decoder part of the U-Net architecture to segment the medical image, and the segmentation result obtained specifically includes:
  • TP True positives
  • TN True negatives
  • the similarity of the segmented images is evaluated by the following commonly used metrics in medical image segmentation:
  • the metrics include: dice index, accuracy, and Jaccard similarity coefficient.
  • the Jaccard similarity coefficient is used to compare the similarity and difference between the limited sample sets. The larger the Jaccard coefficient value, the higher the sample similarity.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Image Analysis (AREA)

Abstract

一种心房间隔闭塞患者心脏医学影像分割方法及系统，包括：采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析方法处理该MRI数据集（S1）；将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得到被正确划分的MRI数据集及被错误划分的MRI数据集（S2）；对被正确划分的MRI训练数据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征（S3）；利用提取的对医学影像分割有用的特征设计U-Net架构，并利用U-Net架构完成端到端的像素到像素的医学影像分割（S4）。该方法能够提高心血管MRI检查中的诊断效率，准确分割心脏医学影像，高效实现端到端的对心脏MRI图像的目标和背景分类。

Description

心房间隔闭塞患者心脏医学影像分割方法及系统 技术领域
本发明涉及一种心房间隔闭塞患者心脏医学影像分割方法及系统。
背景技术
在房间隔缺损(ASD)医学图像分析中，由于来自左心房的隔膜缺损引起的血液流入，右心房显示出严重肿胀，并导致两个心房中的血容量不平衡。磁共振成像(Magnetic Resonance Imaging，MRI)通常用于诊断这种心脏病，然而，金属房间隔封堵器的引入在其放置的区域造成了重影效应，导致不正确的活动轮廓分割。
Kucera等人已经实现了在心脏的短轴和长轴视图上可靠的活动轮廓3D模型，他提出了一种基于区域的外力来分割左心室。Sarti等提出了一种基于区域的分割模型方法，实现了灰度级统计分布的先验知识，他们使用水平集方法驱动曲线演化，以便相对于图像像素的统计分布规律获得目标的最大似然分割。Boukerroui等人提出了另一种基于区域的分割方法，基于自适应分割算法，其中加权函数考虑本地和全局统计。Mishra等人提出了基于遗传算法(GA)求解优化问题的短轴视图分割方法中左心室的另一种主动轮廓分割模型。随后，Mignotte和Meunier提出了多尺度方法进行轮廓优化。Mitchell等人在瞬态超声图像中执行三维主动外观模型(AAM)分割。Bosch等人提出了一种基于其前身 AAM的主动外观运动模型(AAMM)，并开发用于在完整的心动周期中分割左心室，其他完善的分割方法涉及人工神经网络，模糊多尺度边缘检测器和基于卡尔曼滤波器的跟踪方法。
经典的snake模型最初是由Kass、Witkin和Terzopolous提出的，用于将由边缘检测器产生的噪声滤波器响应合成为图像中感知边缘的连贯描绘。这样，可以建立分离具有不同灰度级特性的两个图像区域的边界。半自动分割方法用于基于MR图像分割心脏腔室，通过Kass蛇算法实现，该算法涉及用于分割的基于区域的方法。其可以克服在超声成像中常见的具有差的边界清晰度的分割物体的问题，但对于边界清晰度不明显的分割物体没办法完成MRI图像的分割。同时传统的方法(活动轮廓模型)不能够准确分割一些特殊的MRI图像，比如做过心脏手术并安装过金属支架或金属网的患者，他们的心脏MRI图像中金属物体会以阴影的形式呈现。同时传统的方法，没办法更好地解决研究中训练数据规模相对较小的情况。
发明内容
有鉴于此，有必要提供一种心房间隔闭塞患者心脏医学影像分割方法及系统。
本发明提供一种心房间隔闭塞患者心脏医学影像分割方法，该方法包括如下步骤：a.采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析方法处理该MRI数据集；b.将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得到被正确划分的MRI数据集；c.对被正确划分的MRI训练数 据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征；d.利用提取的对医学影像分割有用的特征设计U-Net架构，并利用U-Net架构完成端到端的像素到像素的医学影像分割。
其中，所述的步骤b具体包括：
采用数据增强的方法对MRI数据集进行数据增强，水平和垂直滑动，随机裁切，增加颜色抖动和高斯噪声；
对数据增强后的MRI数据集进行分割，将分割视为二元分类，即0和1，1代表被正确地划分，0代表被错误地划分。
所述的步骤c具体包括：
利用迁移学习方法选用深度学习卷积神经网络VGG 16的预训练模型作为U-Net网络的编码器；
利用深度学习卷积神经网络VGG 16的预训练模型对ImageNet权值进行初始化；
采用微调的方法修改深度学习卷积神经网络VGG 16预训练模型最后一层的输出类别，并且加快最后一层的参数学习速率；调整Solver的配置参数。
所述的步骤d具体包括如下步骤：
U-Net架构的左半部分是编码器部分，所述编码器捕获上下文的收缩路径，进行特征提取；
U-Net架构的右半部分是解码器部分，所述解码器进行精确定位对称扩展路径；
利用U-Net架构的编码器部分和解码器部分对医学影像进行分割并 得到分割结果。
所述的分割结果包括：
True positives：被正确地划分为正例的个数；
False positives：被错误地划分为正例的个数；
False negatives：被错误地划分为负例的个数；
True negatives：被正确地划分为负例的个数。
本发明提供一种心房间隔闭塞患者心脏医学影像分割系统，该系统包括采集模块、数据集划分模块、微调模块以及影像分割模块，其中：所述采集模块用于采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析方法处理该MRI数据集；所述数据集划分模块用于将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得到被正确划分的MRI数据集；所述微调模块根据对被正确划分的MRI训练数据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征；所述影像分割模块用于利用提取的对医学影像分割有用的特征设计U-Net架构，并利用U-Net架构完成端到端的像素到像素的医学影像分割。
其中，所述的数据集划分模块具体用于：
采用数据增强的方法对MRI数据集进行数据增强，水平和垂直滑动，随机裁切，增加颜色抖动和高斯噪声；
对数据增强后的MRI数据集进行分割，将分割视为二元分类，即0和1，1代表被正确地划分，0代表被错误地划分。
所述的微调模块具体用于：
利用迁移学习方法选用深度学习卷积神经网络VGG 16的预训练模 型作为U-Net网络的编码器；
利用深度学习卷积神经网络VGG 16的预训练模型对ImageNet权值进行初始化；
采用微调的方法修改深度学习卷积神经网络VGG 16预训练模型最后一层的输出类别，并且加快最后一层的参数学习速率；调整Solver的配置参数。
所述的影像分割模块具体用于：
U-Net架构的左半部分是编码器部分，所述编码器捕获上下文的收缩路径，进行特征提取；
U-Net架构的右半部分是解码器部分，所述解码器进行精确定位对称扩展路径；
利用U-Net架构的编码器部分和解码器部分对医学影像进行分割并得到分割结果。
所述的分割结果包括：
True positives：被正确地划分为正例的个数；
False positives：被错误地划分为正例的个数；
False negatives：被错误地划分为负例的个数；
True negatives：被正确地划分为负例的个数。
本发明能够提高心血管MRI检查中诊断的效率，将医学图像分割视为二元分类问题，在面临训练数据规模相对较小的情况时，利用转移学习解决卷积神经网络在训练阶段训练数据不足导致过度拟合的问题；对传统的方法如活动轮廓模型，不能够准确分割特殊 的MRI图像，如做过心脏手术并安装过金属支架或金属网患者的心脏MRI图像，本申请构建一个基于U-Net框架的完整卷积网络，准确分割上述特殊的MRI图像，并能够更高效的实现端到端的对心脏MRI图像的目标和背景分类。
附图说明
图1为本发明心房间隔闭塞患者心脏医学影像分割方法的流程图；
图2为本发明实施例提供的利用光谱分析方法处理心房间隔闭塞患者心脏MRI数据集的示意图；
图3为本发明实施例提供的迁移学习采用的深度学习卷积神经网络VGG 16的示意图；
图4为本发明实施例提供的深度学习卷积神经网络VGG 16预训练结果示意图；
图5为本发明实施例提供的基于U-Net框架的卷积神经网络结构示意图；
图6为本发明心房间隔闭塞患者心脏医学影像分割系统的硬件架构图。
具体实施方式
下面结合附图及具体实施例对本发明作进一步详细的说明。
参阅图1所示，是本发明心房间隔闭塞患者心脏医学影像分割方法较佳实施例的作业流程图。
步骤S1，采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析 方法处理该MRI数据集(请同时参阅图2)，并将所述MRI数据集分为训练集和测试集。具体而言：
本实施例在实验过程中招募了200例心房间隔闭塞患者；
利用西门子1.5T磁共振系统(MRI)MAGNETOM Avanto 1.5T扫描仪和Numaris-4软件进行心房间隔闭塞患者术前和术后MRI数据集的采集；
MRI数据集采集的图像均采用回顾性门控和25个时间帧指数(从nt＝1至25)获得单个切片图像，采集参数包括：在256×256像素的矩阵处TR＝47.1ms，TE＝1.6ms，FOV＝298×340mm 2；
将采集到的550张心房间隔闭塞患者心脏MRI数据集中，80％的数据作为训练集，其余20％的数据作为测试集。
步骤S2，将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得到被正确划分的MRI数据集及被错误划分的MRI数据集。具体而言：
步骤S21，采用数据增强的方法对MRI数据集进行数据增强，水平和垂直滑动，随机裁切，增加颜色抖动和高斯噪声。包括如下步骤：
步骤S211，将MRI数据集的数据进行水平和垂直滑动；
将MRI数据集的数据水平翻转或者垂直翻转，利用工具包任意命令进行翻转，对切片图像进行随机任意角度(0～360度)旋转。
步骤S212，对MRI数据集的数据进行随机裁切；
在本实施例中，利用tensorflow中的随机裁剪函数tf.random_crop随机裁剪切片图像，将所述切片图像裁剪至其大、小、宽、高的2/3。
步骤S213，将MRI数据集的数据增加颜色抖动和高斯噪声；
对切片图像进行颜色抖动，调整图像的饱和度，调整图像亮度，调整图像对比度，调整图像锐度，同时对图像进行高斯噪声处理。
步骤S22，对数据增强后的MRI数据集进行分割，将分割视为二元分类，即0和1，1代表被正确地划分，0代表被错误地划分。
步骤S3，对被正确划分的MRI训练数据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征。具体而言：
步骤S31，利用迁移学习方法选用深度学习卷积神经网络VGG 16(请同时参阅图3)的预训练模型作为U-Net网络的编码器；
步骤S32，利用深度学习卷积神经网络VGG 16的预训练模型对ImageNet权值进行初始化；
步骤S33，训练过程中采用微调的方法，修改深度学习卷积神经网络VGG 16预训练模型最后一层的输出类别，并且加快最后一层的参数学习速率；调整Solver的配置参数，本实施例的深度学习卷积神经网络VGG16预训练结果如图4所示：
所述微调一种是深度学习的方法，就是通过不断的调整网络的参数，使卷积网络性能达到最好。因为微调的先决条件是具有有意义值的预训练模型的权重。当学习速率大时，权重会快速更新，并破坏原来的训练网络结构学习速率，在本实施例中，所述学习速率设置为1×10 -4。
步骤S4，利用提取的对医学影像分割有用的特征设计U-Net架构(请同时参阅图5)，并利用U-Net架构完成端到端的像素到像 素的医学影像分割。具体而言：
步骤S41，U-Net架构的左半部分是编码器部分，所述编码器捕获上下文的收缩路径，进行特征提取，具体包括：
步骤S411，所述编码器的网络采用深度学习卷积神经网络VGG 16进行特征提取；
步骤S412，编码器部分移除了完全连接的层，并用512个通道的单个卷积层进行替换；
步骤S42，U-Net架构的右半部分是解码器部分，所述解码器进行精确定位对称扩展路径，具体包括：
步骤S421，解码器部分使用转置卷积层进行构造，使得特征映射的大小加倍，同时将通道数减少一半；
步骤S422，将转置卷积的输出连接到解码器进行输出；
步骤S423，上采样过程重复5次，以配对5个输出特征图尺寸最大的池；自定义损失函数为：
Figure PCTCN2020129400-appb-000001
其中，y i为一个batch中第i个数据的正确答案，y i'为神经网络得到的预测值，x为实际值，y为预测值，a和b是常量。
步骤S43，利用U-Net架构的编码器部分和解码器部分对医学影像进行分割，得到的分割结果具体包括：
True positives(TP)：被正确地划分为正例的个数，即实际为正例且被分类器划分为正例的实例数(样本数)；
False positives(FP)：被错误地划分为正例的个数，即实际为负例但被分类器划分为正例的实例数；
False negatives(FN)：被错误地划分为负例的个数，即实际为正例但被分类器划分为负例的实例数；
True negatives(TN)：被正确地划分为负例的个数，即实际为负例且被分类器划分为负例的实例数；
通过医学图像分割中下述常用的度量指标分别评估分割的图像的相似性：
所述度量指标包括：dice指数、准确度和Jaccard相似系数。其中，所述Jaccard相似系数(Jaccard similarity coefficient)用于比较有限样本集之间的相似性与差异性，Jaccard系数值越大，样本相似度越高。
参阅图6所示，是本发明心房间隔闭塞患者心脏医学影像分割系统10的硬件架构图。该系统包括：采集模块101、数据集划分模块102、微调模块103以及影像分割模块104。
所述采集模块101用于采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析方法处理该MRI数据集(请同时参阅图2)，并将所述MRI数据集分为训练集和测试集。具体而言：
本实施例在实验过程中招募了200例心房间隔闭塞患者；
利用西门子1.5T磁共振系统(MRI)MAGNETOM Avanto 1.5T扫描仪和Numaris-4软件进行心房间隔闭塞患者术前和术后MRI数据集的采集；
MRI数据集采集的图像均采用回顾性门控和25个时间帧指数(从nt＝1至25)获得单个切片图像，采集参数包括：在256×256像素的矩阵处TR＝47.1ms，TE＝1.6ms，FOV＝298×340mm 2；
将采集到的550张心房间隔闭塞患者心脏MRI数据集中，80％的数据作为训练集，其余20％的数据作为测试集。
所述数据集划分模块102用于将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得到被正确划分的MRI数据集及被错误划分的MRI数据集。具体而言：
所述数据集划分模块102采用数据增强的方法对MRI数据集进行数据增强，水平和垂直滑动，随机裁切，增加颜色抖动和高斯噪声。包括：
将MRI数据集的数据进行水平和垂直滑动：
将MRI数据集的数据水平翻转或者垂直翻转，利用工具包任意命令进行翻转，对切片图像进行随机任意角度(0～360度)旋转；
对MRI数据集的数据进行随机裁切：
在本实施例中，利用tensorflow中的随机裁剪函数tf.random_crop随机裁剪切片图像，将所述切片图像裁剪至其大、小、宽、高的2/3；
将MRI数据集的数据增加颜色抖动和高斯噪声：
对切片图像进行颜色抖动，调整图像的饱和度，调整图像亮度，调整图像对比度，调整图像锐度，同时对图像进行高斯噪声处理。
所述数据集划分模块102对数据增强后的MRI数据集进行分割，将 分割视为二元分类，即0和1，1代表被正确地划分，0代表被错误地划分。
所述微调模块103用于对被正确划分的MRI训练数据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征。具体而言：
所述微调模块103利用迁移学习方法选用深度学习卷积神经网络VGG 16(请同时参阅图3)的预训练模型作为U-Net网络的编码器；
所述微调模块103利用深度学习卷积神经网络VGG 16的预训练模型对ImageNet权值进行初始化；
所述微调模块103训练过程中采用微调的方法，修改深度学习卷积神经网络VGG 16预训练模型最后一层的输出类别，并且加快最后一层的参数学习速率；调整Solver的配置参数，本实施例的深度学习卷积神经网络VGG 16预训练结果如图4所示：
所述微调一种是深度学习的方法，就是通过不断的调整网络的参数，使卷积网络性能达到最好。因为微调的先决条件是具有有意义值的预训练模型的权重。当学习速率大时，权重会快速更新，并破坏原来的训练网络结构学习速率，在本实施例中，所述学习速率设置为1×10 -4。
所述影像分割模块104用于利用提取的对医学影像分割有用的特征设计U-Net架构(请同时参阅图5)，并利用U-Net架构完成端到端的像素到像素的医学影像分割。具体而言：
所述影像分割模块104进行特征提取，也即是，U-Net架构的左半部分是编码器部分，所述编码器捕获上下文的收缩路径，进行特征提取，具体包括：
所述编码器的网络采用深度学习卷积神经网络VGG 16进行特征提取；
编码器部分移除了完全连接的层，并用512个通道的单个卷积层进行替换；
所述影像分割模块104构建U-Net架构的解码器部分，U-Net架构的右半部分是解码器部分，所述解码器进行精确定位对称扩展路径，具体包括：
所述影像分割模块104使用转置卷积层进行构造解码器部分，使得特征映射的大小加倍，同时将通道数减少一半；
所述影像分割模块104将转置卷积的输出连接到解码器进行输出；
所述影像分割模块104重复上采样过程5次，以配对5个输出特征图尺寸最大的池；自定义损失函数为：
Figure PCTCN2020129400-appb-000002
其中，y i为一个batch中第i个数据的正确答案，y i'为神经网络得到的预测值，x为实际值，y为预测值，a和b是常量。
所述影像分割模块104利用U-Net架构的编码器部分和解码器部分对医学影像进行分割，得到的分割结果具体包括：
True positives(TP)：被正确地划分为正例的个数，即实际为正例且被分类器划分为正例的实例数(样本数)；
False positives(FP)：被错误地划分为正例的个数，即实际为负例但被分类器划分为正例的实例数；
False negatives(FN)：被错误地划分为负例的个数，即实际为正例但被分类器划分为负例的实例数；
True negatives(TN):被正确地划分为负例的个数，即实际为负例且被分类器划分为负例的实例数；
通过医学图像分割中下述常用的度量指标分别评估分割的图像的相似性：
所述度量指标包括：dice指数、准确度和Jaccard相似系数。其中，所述Jaccard相似系数(Jaccard similarity coefficient)用于比较有限样本集之间的相似性与差异性，Jaccard系数值越大，样本相似度越高。
虽然本发明参照当前的较佳实施方式进行了描述，但本领域的技术人员应能理解，上述较佳实施方式仅用来说明本发明，并非用来限定本发明的保护范围，任何在本发明的精神和原则范围之内，所做的任何修饰、等效替换、改进等，均应包含在本发明的权利保护范围之内。

Claims (10)

  1. 一种心房间隔闭塞患者心脏医学影像分割方法，其特征在于，该方法包括如下步骤：
    a.采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析方法处理该MRI数据集；
    b.将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得到被正确划分的MRI数据集；
    c.对被正确划分的MRI训练数据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征；
    d.利用提取的对医学影像分割有用的特征设计U-Net架构，并利用U-Net架构完成端到端的像素到像素的医学影像分割。
  2. 如权利要求1所述的方法，其特征在于，所述的步骤b具体包括：
    采用数据增强的方法对MRI数据集进行数据增强，水平和垂直滑动，随机裁切，增加颜色抖动和高斯噪声；
    对数据增强后的MRI数据集进行分割，将分割视为二元分类，即0和1，1代表被正确地划分，0代表被错误地划分。
  3. 如权利要求2所述的方法，其特征在于，所述的步骤c具体包括：
    利用迁移学习方法选用深度学习卷积神经网络VGG 16的预训练模型作为U-Net网络的编码器；
    利用深度学习卷积神经网络VGG 16的预训练模型对ImageNet权值进行初始化；
    采用微调的方法修改深度学习卷积神经网络VGG 16预训练模型最后一层的输出类别，并且加快最后一层的参数学习速率；调整Solver的配置参数。
  4. 如权利要求3所述的方法，其特征在于，所述的步骤d具体包括如下步骤：
    U-Net架构的左半部分是编码器部分，所述编码器捕获上下文的收缩路径，进行特征提取；
    U-Net架构的右半部分是解码器部分，所述解码器进行精确定位对称扩展路径；
    利用U-Net架构的编码器部分和解码器部分对医学影像进行分割并得到分割结果。
  5. 如权利要求4所述的方法，其特征在于，所述的分割结果包括：
    True positives：被正确地划分为正例的个数；
    False positives：被错误地划分为正例的个数；
    False negatives：被错误地划分为负例的个数；
    True negatives：被正确地划分为负例的个数。
  6. 一种心房间隔闭塞患者心脏医学影像分割系统，其特征在于，该系统包括采集模块、数据集划分模块、微调模块以及影像分割模块，其中：
    所述采集模块用于采集心房间隔闭塞患者心脏MRI数据集，利用光谱分析方法处理该MRI数据集；
    所述数据集划分模块用于将由光谱分析方法处理过的MRI数据集进行数据增强，并将数据增强后的MRI数据集进行二元分类的分割，得 到被正确划分的MRI数据集；
    所述微调模块根据对被正确划分的MRI训练数据集，采用迁移学习方法微调卷积神经网络模型，以提取对后续医学影像分割有用的特征；
    所述影像分割模块用于利用提取的对医学影像分割有用的特征设计U-Net架构，并利用U-Net架构完成端到端的像素到像素的医学影像分割。
  7. 如权利要求6所述的系统，其特征在于，所述的数据集划分模块具体用于：
    采用数据增强的方法对MRI数据集进行数据增强，水平和垂直滑动，随机裁切，增加颜色抖动和高斯噪声；
    对数据增强后的MRI数据集进行分割，将分割视为二元分类，即0和1，1代表被正确地划分，0代表被错误地划分。
  8. 如权利要求7所述的系统，其特征在于，所述的微调模块具体用于：
    利用迁移学习方法选用深度学习卷积神经网络VGG 16的预训练模型作为U-Net网络的编码器；
    利用深度学习卷积神经网络VGG 16的预训练模型对ImageNet权值进行初始化；
    采用微调的方法修改深度学习卷积神经网络VGG 16预训练模型最后一层的输出类别，并且加快最后一层的参数学习速率；调整Solver的配置参数。
  9. 如权利要求8所述的系统，其特征在于，所述的影像分割模块具体用于：
    U-Net架构的左半部分是编码器部分，所述编码器捕获上下文的收缩路径，进行特征提取；
    U-Net架构的右半部分是解码器部分，所述解码器进行精确定位对称扩展路径；
    利用U-Net架构的编码器部分和解码器部分对医学影像进行分割并得到分割结果。
  10. 如权利要求9所述的系统，其特征在于，所述的分割结果包括：
    True positives：被正确地划分为正例的个数；
    False positives：被错误地划分为正例的个数；
    False negatives：被错误地划分为负例的个数；
    True negatives：被正确地划分为负例的个数。
PCT/CN2020/129400 2019-12-28 2020-11-17 心房间隔闭塞患者心脏医学影像分割方法及系统 WO2021129234A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911389822.2 2019-12-28
CN201911389822.2A CN111127504B (zh) 2019-12-28 2019-12-28 心房间隔闭塞患者心脏医学影像分割方法及系统

Publications (1)

Publication Number Publication Date
WO2021129234A1 true WO2021129234A1 (zh) 2021-07-01

Family

ID=70504584

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/129400 WO2021129234A1 (zh) 2019-12-28 2020-11-17 心房间隔闭塞患者心脏医学影像分割方法及系统

Country Status (2)

Country Link
CN (1) CN111127504B (zh)
WO (1) WO2021129234A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111127504B (zh) * 2019-12-28 2024-02-09 中国科学院深圳先进技术研究院 心房间隔闭塞患者心脏医学影像分割方法及系统
CN111584093B (zh) * 2020-05-12 2021-04-30 鲁东大学 可注射水凝胶疗效评估的左心室几何模型构造方法和装置
CN111739000B (zh) * 2020-06-16 2022-09-13 山东大学 一种提高多个心脏视图左心室分割精确度的系统及装置
CN111915557A (zh) * 2020-06-23 2020-11-10 杭州深睿博联科技有限公司 一种深度学习的房间隔缺损检测方法及装置
CN113205528B (zh) * 2021-04-02 2023-07-07 上海慧虎信息科技有限公司 医学影像分割模型训练方法、分割方法及装置
CN113379682B (zh) * 2021-05-21 2022-10-04 郑州大学 一种心脏mri影像耦合水平集分割方法及系统
CN113555089A (zh) * 2021-07-14 2021-10-26 江苏宏创信息科技有限公司 一种应用于临床影像的人工智能医学影像质控方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109215035A (zh) * 2018-07-16 2019-01-15 江南大学 一种基于深度学习的脑部mri海马体三维分割方法
CN110619641A (zh) * 2019-09-02 2019-12-27 南京信息工程大学 一种基于深度学习的三维乳腺癌核磁共振图像肿瘤区域的自动分割方法
CN111127504A (zh) * 2019-12-28 2020-05-08 中国科学院深圳先进技术研究院 心房间隔闭塞患者心脏医学影像分割方法及系统

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10521902B2 (en) * 2015-10-14 2019-12-31 The Regents Of The University Of California Automated segmentation of organ chambers using deep learning methods from medical imaging
CN109690554B (zh) * 2016-07-21 2023-12-05 西门子保健有限责任公司 用于基于人工智能的医学图像分割的方法和系统
CN108492286B (zh) * 2018-03-13 2020-05-05 成都大学 一种基于双通路u型卷积神经网络的医学图像分割方法
CN110570432A (zh) * 2019-08-23 2019-12-13 北京工业大学 一种基于深度学习的ct图像肝脏肿瘤分割方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109215035A (zh) * 2018-07-16 2019-01-15 江南大学 一种基于深度学习的脑部mri海马体三维分割方法
CN110619641A (zh) * 2019-09-02 2019-12-27 南京信息工程大学 一种基于深度学习的三维乳腺癌核磁共振图像肿瘤区域的自动分割方法
CN111127504A (zh) * 2019-12-28 2020-05-08 中国科学院深圳先进技术研究院 心房间隔闭塞患者心脏医学影像分割方法及系统

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
FAISAL MAHMOOD; RICHARD CHEN; SANDRA SUDARSKY; DAPHNE YU; NICHOLAS J. DURR: "Deep Learning with Cinematic Rendering - Fine-Tuning Deep Neural Networks Using Photorealistic Medical Images", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 22 May 2018 (2018-05-22), 201 Olin Library Cornell University Ithaca, NY 14853, XP080880850 *
MAAYAN FRID-ADAR; AVI BEN-COHEN; RULA AMER; HAYIT GREENSPAN: "Improving the Segmentation of Anatomical Structures in Chest Radiographs using U-Net with an ImageNet Pre-trained Encoder", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 4 October 2018 (2018-10-04), 201 Olin Library Cornell University Ithaca, NY 14853, XP080929686, DOI: 10.1007/978-3-030-00946-5_17 *
WANG C.; RAJCHL M.; CHAN A. D. C.; UKWATTA E.: "An ensemble of U-Net architecture variants for left atrial segmentation", PROGRESS IN BIOMEDICAL OPTICS AND IMAGING, SPIE - INTERNATIONAL SOCIETY FOR OPTICAL ENGINEERING, BELLINGHAM, WA, US, vol. 10950, 13 March 2019 (2019-03-13), BELLINGHAM, WA, US, pages 109500M - 109500M-7, XP060119895, ISSN: 1605-7422, ISBN: 978-1-5106-0027-0, DOI: 10.1117/12.2512905 *

Also Published As

Publication number Publication date
CN111127504A (zh) 2020-05-08
CN111127504B (zh) 2024-02-09

Similar Documents

Publication Publication Date Title
WO2021129234A1 (zh) 心房间隔闭塞患者心脏医学影像分割方法及系统
US11961226B2 (en) Medical image recognition method, model training method, and computer device
CN110930416B (zh) 一种基于u型网络的mri图像前列腺分割方法
CN107563434B (zh) 一种基于三维卷积神经网络的脑部mri图像分类方法、装置
CN111401480A (zh) 一种基于融合注意力机制的乳腺mri自动辅助诊断新方法
WO2018060723A1 (en) Image processing
CN111597946A (zh) 图像生成器的处理方法、图像生成方法及装置
JP2023159216A (ja) 画像処理装置、画像処理方法、プログラム
Lu et al. Cardiac chamber segmentation using deep learning on magnetic resonance images from patients before and after atrial septal occlusion surgery
Yang et al. A robust iris segmentation using fully convolutional network with dilated convolutions
CN110047075A (zh) 一种基于对抗网络的ct图像分割方法
CN113888520A (zh) 用于生成牛眼图的系统和方法
JP6041781B2 (ja) 医用画像処理装置およびその作動方法並びに医用画像処理プログラム
Bhuvaneswari et al. Contrast enhancement of retinal images using green plan masking and whale optimization algorithm
Kanse et al. HG-SVNN: harmonic genetic-based support vector neural network classifier for the glaucoma detection
Ataei et al. Low dose ct denoising using dilated residual learning with perceptual loss and structural dissimilarity
Guo et al. Thyroid nodule ultrasonic imaging segmentation based on a deep learning model and data augmentation
Bobkov et al. The application of ensemble learning for delineation of the left ventricle on echocardiographic records
Carmo et al. Extended 2d volumetric consensus hippocampus segmentation
Khan et al. A Computer-Aided Diagnostic System to Identify Diabetic Retinopathy, Utilizing a Modified Compact Convolutional Transformer and Low-Resolution Images to Reduce Computation Time. Biomedicines. 2023. No. 11. Art. 1566
El-Melegy et al. A combined fuzzy C-means and level set method for automatic DCE-MRI kidney segmentation using both population-based and patient-specific shape statistics
Shao et al. Semantic segmentation method of 3D liver image based on contextual attention model
Bian et al. Automatic optic disc/cup segmentation and glaucoma classification and fovea localization
Kumar et al. Feed forward neural network based automatic detection of liver in computer tomography images
Paul et al. Computer-Aided Diagnosis Using Hybrid Technique for Fastened and Accurate Analysis of Tuberculosis Detection with Adaboost and Learning Vector Quantization

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: 20908224

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: 20908224

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 1205A DATED 20/01/2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20908224

Country of ref document: EP

Kind code of ref document: A1