WO2020057074A1 - 用于斑块分割的模型训练方法、装置、设备及存储介质 - Google Patents

用于斑块分割的模型训练方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2020057074A1
WO2020057074A1 PCT/CN2019/078469 CN2019078469W WO2020057074A1 WO 2020057074 A1 WO2020057074 A1 WO 2020057074A1 CN 2019078469 W CN2019078469 W CN 2019078469W WO 2020057074 A1 WO2020057074 A1 WO 2020057074A1
Authority
WO
WIPO (PCT)
Prior art keywords
plaque
segmentation
image
training
fine
Prior art date
Application number
PCT/CN2019/078469
Other languages
English (en)
French (fr)
Inventor
郑海荣
刘新
胡战利
张娜
梁栋
杨永峰
Original Assignee
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 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Publication of WO2020057074A1 publication Critical patent/WO2020057074A1/zh

Links

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS

Definitions

  • the invention belongs to the field of medical technology, and particularly relates to a model training method, device, equipment and storage medium for plaque segmentation.
  • Stroke is the disease with the highest mortality and disability in China. Stroke includes hemorrhagic stroke and ischemic stroke. Stroke in China is mainly ischemic stroke, accounting for 79% of all stroke cases. Moreover, there is an increasing trend. Similar cardiovascular diseases seriously threaten human life and health. With the rapid development of science and technology, the early quantitative diagnosis and risk assessment of cardiovascular diseases play a key role in extending human life and health. Studies have shown that thrombosis caused by atherosclerotic plaque rupture is the main pathogenesis of ischemic stroke. Therefore, timely discovery of stroke-related vascular beds, including vulnerable plaques in the intracranial, carotid and thoracic aorta Or other wall lesions are the key to early prevention and accurate treatment of ischemic stroke.
  • the purpose of the present invention is to provide a model training method, device, equipment and storage medium for plaque segmentation, which aims to solve the problem that the plaque on the segmented blood vessel wall due to the inability to provide an effective plaque segmentation model in the prior art.
  • the present invention provides a model training method for plaque segmentation, which includes the following steps:
  • the fine-tuned deep learning network model for plaque segmentation after outputting iterative training is output.
  • the present invention provides a model training device for plaque segmentation.
  • the device includes:
  • An image acquisition unit configured to acquire an undivided plaque image of a first preset number of blood vessel walls and a divided plaque image corresponding to the undivided plaque image
  • a segmentation training unit configured to input the unsegmented plaque image to a deep learning network model for segmentation training to obtain a first segmentation result of the unsegmented plaque image
  • a fine-tuning training unit configured to fine-tune and iteratively train the deep learning network model after the segmentation training by using the first segmentation result and the segmented plaque image
  • a model output unit is configured to output the deep learning network model for plaque segmentation after fine-tuned iterative training.
  • the present invention also provides a model training device including a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • a model training device including a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • the processor executes the computer program, Implement the steps of the model training method for plaque segmentation as described above.
  • the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-mentioned model training method for plaque segmentation. A step of.
  • the present invention obtains a first preset number of undivided plaque images of a blood vessel wall and a divided plaque image corresponding to the undivided plaque image, inputs the undivided plaque image to a deep learning network model, and performs segmentation training to obtain an undivided plaque image. Segment the first segmentation result of the plaque image, and use the first segmentation result and the segmented plaque image to fine-tune the iterative training of the segmented training deep learning network model, and output the fine-tuned deep learning network for plaque segmentation after iterative training
  • the model thus realizes the automation of plaque segmentation on the blood vessel wall and improves the accuracy of plaque segmentation, thereby improving the efficiency of plaque segmentation and further improving the efficiency of plaque segmentation.
  • FIG. 1 is an implementation flowchart of a model training method for plaque segmentation provided by Embodiment 1 of the present invention
  • FIG. 2 is an implementation flowchart of a model training method for plaque segmentation provided by Embodiment 2 of the present invention
  • FIG. 3 is a schematic diagram of a fine-tuning process of a first segmentation result provided in Embodiment 2 of the present invention.
  • FIG. 4 is a schematic structural diagram of a model training device for plaque segmentation provided by Embodiment 3 of the present invention.
  • FIG. 5 is a schematic structural diagram of a model training device for plaque segmentation provided by Embodiment 4 of the present invention.
  • FIG. 6 is a schematic diagram of a preferred structure of an image fine-tuning unit provided in Embodiment 4 of the present invention.
  • FIG. 7 is a schematic structural diagram of a model training device according to a fifth embodiment of the present invention.
  • FIG. 1 shows the implementation process of the model training method for plaque segmentation provided by the first embodiment of the present invention. For convenience of explanation, only the parts related to the embodiment of the present invention are shown, and the details are as follows:
  • step S101 an undivided plaque image corresponding to a first preset number of blood vessel walls and divided plaque images corresponding to the undivided plaque images are acquired.
  • the embodiment of the present invention is applicable to a model training device, and the model is used to perform plaque segmentation on a blood vessel wall image containing plaque (to mark a plaque portion in the blood vessel wall image).
  • the undivided plaque image is an originally obtained blood vessel wall image with unlabeled plaque parts
  • the segmented plaque image is an outline (marking) of the plaque part on the undivided plaque image in advance.
  • the subsequent image of the blood vessel wall can be outlined by ITK-SNAP.
  • an undivided plaque image of the blood vessel wall and a corresponding segmented plaque image are obtained to facilitate subsequent training of the model.
  • the undivided plaques The number of images and segmented plaque images are the same, and the undivided plaque image is a three-dimensional MRI image obtained by magnetic resonance imaging.
  • step S102 an unsegmented plaque image is input to a deep learning network model for segmentation training to obtain a first segmentation result of the unsegmented plaque image.
  • the deep learning network model is a U-Net network. Since the deep learning network can only convolve two-dimensional images, a certain number of two-dimensional cross-sectional views of a three-dimensional MRI image of a blood vessel wall are obtained first, and then Enter a certain number of 2D cross-sectional views into the deep learning network for segmentation training. You can also obtain the vascular wall expansion map generated by the 3D MRI image of the vascular wall based on the principle that the side of the cylinder is expanded to be rectangular, and then input the vascular wall expansion view. Go to deep learning network for segmentation training.
  • the deep learning network model is a V-Net network
  • the images received by the V-Net network are three-dimensional images, thereby simplifying the segmentation training process and increasing the accuracy of segmenting plaques.
  • the unsegmented plaque image is segmented according to a preset image feature of the unsegmented plaque image, thereby quickly obtaining a first segmentation result of the unsegmented plaque image
  • the preset image feature is a plaque portion
  • the characteristics of the image preferably, the preset image characteristics are pixel density characteristics of the plaque portion in the undivided plaque image, thereby speeding up the training speed of the model.
  • step S103 the deep learning network model after segmentation training is fine-tuned and iteratively trained through the first segmentation result and the segmented plaque image.
  • Reinforcement learning also known as re-encouraging learning and evaluation learning, is an agent that learns in a “trial and error” manner, and rewards and guides behaviors obtained by interacting with the environment. The goal is to make the agent receive the greatest rewards and strengthen Learning is different from supervised learning in connectionist learning, which is mainly manifested in teacher signals.
  • the reinforcement signals provided by the environment in reinforcement learning are an evaluation of the quality of the action (usually scalar signals), rather than telling reinforcement learning.
  • System RLS reinforcementment (reinforcement (learning system) how to generate correct actions. Since the external environment provides little information, RLS must learn from its own experience. In this way, RLS gains knowledge in an action-evaluation environment and refines action plans to suit the environment.
  • the first segmentation result is not a good segmentation result
  • here we introduce a reinforcement learning method use this first segmentation result to train a deep learning network model, and perform adjustments on the first segmentation result. Action, and evaluate the adjustment action based on the segmented plaque image, so as to implement fine-tune iterative training on the deep learning network model after segmentation training.
  • step S104 the fine-tuning deep learning network model for plaque segmentation is output.
  • a fine-tuned deep learning network model after iterative training is output, and the deep learning network model can accurately and accurately outline a partial image of a plaque on a blood vessel wall.
  • an undivided plaque image corresponding to a first preset number of blood vessel walls and a segmented plaque image corresponding to the undivided plaque image are obtained, and the undivided plaque image is input to a deep learning network model for segmentation.
  • FIG. 2 shows the implementation process of the model training method for plaque segmentation provided by the second embodiment of the present invention.
  • the details are as follows:
  • step S201 an undivided plaque image corresponding to a first preset number of blood vessel walls and divided plaque images corresponding to the undivided plaque images are acquired.
  • the embodiment of the present invention is applicable to a model training device, which is used to perform plaque segmentation on a blood vessel wall image containing plaque.
  • the segmented plaque image is a blood vessel wall image obtained by preliminarily delineating a plaque portion on an undivided plaque image. Specifically, it can be delineated by ITK-SNAP. Specifically, it can be delineated by ITK. -SNAP outlines, first obtain the undivided plaque image of the blood vessel wall and the corresponding segmented plaque image to facilitate the subsequent training of the model, where the number of undivided plaque images and the divided plaque images are the same
  • the unsegmented plaque image is a three-dimensional MRI image obtained by magnetic resonance imaging.
  • step S202 an image pre-processing operation is performed on the undivided plaque image and the divided plaque image.
  • the image size reduction is performed on the undivided plaque image and the divided plaque image.
  • an image binarization operation is performed on the undivided plaque image and the divided plaque image, so that the features of the image are more obvious, and the subsequent feature extraction of the undivided plaque image is facilitated.
  • step S203 the pre-processed unsegmented plaque image is input to a deep learning network model for segmentation training to obtain a first segmentation result of the unsegmented plaque image.
  • the deep learning network model is a U-Net network. Since the deep learning network can only convolve two-dimensional images, a certain number of two-dimensional cross-sectional views of a three-dimensional MRI image of a blood vessel wall are obtained first, and then Enter a certain number of 2D cross-sectional views into the deep learning network for segmentation training. You can also obtain the vascular wall expansion map generated by the 3D MRI image of the vascular wall based on the principle that the side of the cylinder is expanded to be rectangular, and then input the vascular wall expansion view. Go to deep learning network for segmentation training.
  • the unsegmented plaque image is segmented according to a preset image feature of the unsegmented plaque image, thereby quickly obtaining a first segmentation result of the unsegmented plaque image
  • the preset image feature is a plaque portion
  • the characteristics of the image preferably, the preset image characteristics are pixel density characteristics of the plaque portion in the undivided plaque image, thereby speeding up the training speed of the model.
  • the deep learning network combines a residual function so that the input of each convolutional layer in the deep learning network is the sum of the input and output of the previous convolutional layer, thereby enriching the information input of the convolutional layer.
  • the Dense network is superimposed on the deep learning network, thereby enhancing the feature transfer and mitigating the disappearance of the gradient, thereby reducing the model parameters of the plaque segmentation model.
  • a feature network obtained by dilated convolution is added to the feature network obtained by the convolution, thereby increasing the receptive field of the image and enabling the feature output of the convolution layer. Contains more information.
  • step S204 fine-tune iterative training is performed on the deep learning network model after segmentation training through the first segmentation result and the segmented plaque image.
  • the first segmentation result is not a good segmentation result
  • here we introduce a reinforcement learning method use this first segmentation result to train a deep learning network model, and perform adjustments on the first segmentation result.
  • the first segmentation result is fine-tuned according to the segmented plaque image to obtain the second segmentation result of the undivided plaque image, and then the similarity between the second segmentation result and the segmented plaque image is calculated.
  • the similarity between the segmented plaque image and the second segmentation result can be calculated by the Dice coefficient.
  • a second preset number of coordinates of the same orientation on the first segmentation result and the segmentation boundary on the segmented plaque image are obtained (for convenience of subsequent description) , These coordinates are called azimuth coordinates), the azimuth coordinates of the first segmentation result are adjusted according to the azimuth coordinates of the segmented plaque image, thereby increasing the accuracy of fine-tuning, and thereby reducing the training time of the model.
  • the second preset number of the same azimuths be east, south, west, north, southeast, southwest, and northwest , Northeast, eight azimuth coordinates, so that the training duration and training accuracy are relatively balanced.
  • FIG. 3 shows the fine-tuning process of the first segmentation result.
  • the left image is the segmented plaque image and the right image is the first segmentation result.
  • the azimuth coordinates of the left and right images are set to east, south, and west.
  • a 1 , A 2 , A 3 , and A 4 are used in the segmented plaque image to indicate the 4 azimuth coordinates in the east, south, west, and north.
  • B 1 and B are used in the first segmentation result.
  • 2 , B 3 , and B 4 represent four azimuth coordinates of east, south, west, and north.
  • B 1 approaches A 1
  • B 2 approaches A 2
  • B 3 approaches A 3
  • B 4 approaches A 4.
  • the azimuth coordinates B 1 , B 2 , B 3 , and B 4 near the first segmentation result are adjusted according to the azimuth coordinates A 1 , A 2 , A 3 , and A 4, respectively.
  • step S205 the fine-tuning deep learning network model for plaque segmentation is output.
  • the similarity between the second segmentation result and the segmented plaque image is calculated. If the similarity is greater than a preset similarity threshold, it indicates that The segmentation accuracy of the currently trained model has reached the expected value. At this time, the fine-tuning iterative training of the deep learning network model is completed, and the deep learning network model after fine-tuning iterative training is output.
  • the deep learning network model can Partial image of the plaque is accurately outlined.
  • an undivided plaque image corresponding to a first preset number of blood vessel walls and a segmented plaque image corresponding to the undivided plaque image are obtained, and an image pre-processing operation is performed to unprocessed un-segmented
  • the plaque image is input to a deep learning network model for segmentation training to obtain the first segmentation result of the un-segmented plaque image.
  • the first training result and the segmented plaque image are used to fine-tune the iterative training of the segmented training deep learning network model.
  • the output fine-tunes the deep learning network model for plaque segmentation after iterative training, thereby realizing the automation of plaque segmentation on the vessel wall and improving the accuracy of plaque segmentation, thereby realizing the automation of plaque segmentation on the vessel wall And it improves the accuracy of plaque segmentation, thereby improving the efficiency of plaque segmentation, and further improving the efficiency of plaque segmentation.
  • FIG. 4 shows the structure of a model training device for plaque segmentation provided by Embodiment 3 of the present invention. For convenience of explanation, only parts related to the embodiment of the present invention are shown, including:
  • An image acquisition unit 41 configured to acquire an undivided plaque image and a segmented plaque image corresponding to the first preset number of blood vessel walls;
  • a segmentation training unit 42 configured to input an unsegmented plaque image to a deep learning network model for segmentation training to obtain a first segmentation result of the unsegmented plaque image
  • a fine-tuning training unit 43 for fine-tuning and iterative training of the deep learning network model after the segmentation training by using the first segmentation result and the segmented plaque image
  • a model output unit 44 is configured to output a fine-tuning deep learning network model for plaque segmentation after iterative training.
  • an undivided plaque image corresponding to a first preset number of blood vessel walls and a segmented plaque image corresponding to the undivided plaque image are obtained, and the undivided plaque image is input to a deep learning network model for segmentation.
  • each unit of the model training device for plaque segmentation may be implemented by corresponding hardware or software units.
  • Each unit may be an independent software and hardware unit, or may be integrated into one software and hardware unit. This is not intended to limit the invention.
  • reference may be made to the description in Embodiment 1, and details are not described herein again.
  • Embodiment 4 is a diagrammatic representation of Embodiment 4:
  • FIG. 5 shows the structure of a model training device for plaque segmentation provided in Embodiment 4 of the present invention. For ease of description, only parts related to the embodiment of the present invention are shown, including:
  • An image obtaining unit 51 configured to obtain a first preset number of undivided plaque images and a divided plaque image corresponding to the undivided plaque image;
  • a segmentation training unit 52 configured to perform image preprocessing operations on an undivided plaque image and a segmented plaque image
  • a fine-tuning training unit 53 for inputting a pre-processed unsegmented plaque image to a deep learning network model for segmentation training to obtain a first segmentation result of the unsegmented plaque image;
  • the model output unit 54 is configured to output a fine-tuning deep learning network model for plaque segmentation after iterative training.
  • the fine-tuning training unit 53 includes:
  • An image fine-tuning unit 531 configured to fine-tune the first segmentation result according to the segmented plaque image to obtain a second segmentation result of the undivided plaque image
  • the calculation and judgment unit 532 is configured to calculate the similarity between the second segmentation result and the segmented plaque image, and determine whether to continue to perform fine-tuning and iterative training on the deep learning network model after segmentation training according to the similarity.
  • FIG. 6 shows a preferred structure of the image fine-tuning unit 531.
  • the image fine-tuning unit 531 includes:
  • a coordinate acquisition unit 61 configured to respectively acquire a first segmentation result and a second preset number of coordinates of the same orientation on the segmentation boundary on the segmented patch image;
  • the coordinate adjusting unit 62 is configured to adjust the coordinates of the first segmentation result according to the coordinates of the segmented plaque image.
  • an undivided plaque image corresponding to a first preset number of blood vessel walls and a segmented plaque image corresponding to the undivided plaque image are obtained, and an image pre-processing operation is performed to unprocessed un-segmented
  • the plaque image is input to a deep learning network model for segmentation training to obtain the first segmentation result of the un-segmented plaque image.
  • the first training result and the segmented plaque image are used to fine-tune the iterative training of the segmented training deep learning network model.
  • the output fine-tunes the deep learning network model for plaque segmentation after iterative training, thereby realizing the automation of plaque segmentation on the vessel wall and improving the accuracy of plaque segmentation, thereby realizing the automation of plaque segmentation on the vessel wall And it improves the accuracy of plaque segmentation, thereby improving the efficiency of plaque segmentation, and further improving the efficiency of plaque segmentation.
  • each unit of the model training device for plaque segmentation may be implemented by corresponding hardware or software units.
  • Each unit may be an independent software and hardware unit, or may be integrated into one software and hardware unit. This is not intended to limit the invention.
  • reference may be made to the description in Embodiment 2, and details are not described herein again.
  • Embodiment 5 is a diagrammatic representation of Embodiment 5:
  • FIG. 7 shows the structure of a model training device provided in Embodiment 5 of the present invention. For convenience of explanation, only parts related to the embodiment of the present invention are shown, including:
  • the model training device 7 includes a processor 71, a memory 72, and a computer program 73 stored in the memory 72 and executable on the processor 71.
  • the processor 71 executes the computer program 73
  • the steps in the foregoing embodiments of the model training method for plaque segmentation are implemented, such as steps S101 to S104 shown in FIG. 1 and steps S201 to S205 shown in FIG. 2.
  • the processor 71 executes the computer program 73
  • the functions of the units in the foregoing embodiments of the model training device for plaque segmentation are realized, for example, the functions of units 41 to 44 shown in FIG. 4 and the functions of units 51 to 54 shown in FIG. 5 .
  • an undivided plaque image corresponding to the first preset number of blood vessel walls and divided plaque images corresponding to the undivided plaque image are acquired, and the undivided plaque image is obtained.
  • a deep learning network model for plaque segmentation after iterative training thereby realizing the automation of plaque segmentation on the blood vessel wall and improving the accuracy of plaque segmentation, thereby improving the efficiency of plaque segmentation.
  • Embodiment 6 is a diagrammatic representation of Embodiment 6
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, each of the foregoing model training method embodiments for plaque segmentation is implemented. For example, steps S101 to S104 shown in FIG. 1 and steps S201 to S205 shown in FIG. 2.
  • the functions of the units in the foregoing embodiments of the model training device for plaque segmentation are realized, for example, the functions of units 41 to 44 shown in FIG. 4 and the functions of units 51 to 54 shown in FIG. 5 .
  • the first preset number of undivided plaque images and the divided plaque images corresponding to the undivided plaque images are obtained, and the undivided plaques are obtained.
  • the image is input to the deep learning network model for segmentation training to obtain the first segmentation result of the unsegmented plaque image.
  • the segmented trained deep learning network model is performed through the first segmentation result and the segmented plaque image. Fine-tune iterative training, output the fine-tuned deep learning network model for plaque segmentation after iterative training, thereby realizing the automation of plaque segmentation on the blood vessel wall and improving the accuracy of plaque segmentation, thereby improving the efficiency of plaque segmentation, and Improve plaque segmentation efficiency.
  • the computer-readable storage medium of the embodiment of the present invention may include any entity or device capable of carrying computer program code, and a storage medium, for example, a memory such as a ROM / RAM, a magnetic disk, an optical disk, a flash memory, or the like.
  • a storage medium for example, a memory such as a ROM / RAM, a magnetic disk, an optical disk, a flash memory, or the like.

Landscapes

  • Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Engineering & Computer Science (AREA)
  • Medical Informatics (AREA)
  • Primary Health Care (AREA)
  • General Health & Medical Sciences (AREA)
  • Epidemiology (AREA)
  • Data Mining & Analysis (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Biomedical Technology (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种用于斑块分割的模型训练方法、装置、设备及存储介质，适用医疗技术领域，该方法包括：获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像（S101），将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果（S102），通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练（S103），输出微调迭代训练后的用于斑块分割的深度学习网络模型（S104），从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率。

Description

用于斑块分割的模型训练方法、装置、设备及存储介质 技术领域
本发明属于医疗技术领域，尤其涉及一种用于斑块分割的模型训练方法、装置、设备及存储介质。
 背景技术
脑卒中是我国死亡率和致残率最高的疾病，脑卒中包括出血性脑卒中和缺血性脑卒中两种，我国脑卒中以缺血性脑卒中为主，占全部脑卒中病例的79%，而且有增加趋势，类似的  心血管疾病严重威胁着人类的生命健康，随着科技高速发展，对心血管疾病的早期定量诊断和风险评估对延长人类生命健康起着关键的作用。 研究表明，动脉粥样硬化斑块破裂引发血栓形成是缺血性脑卒中的主要发病机制，因此，及时发现脑卒中相关血管床，包括颅内动脉、颈动脉和胸主动脉的易损斑块或其他管壁病变是缺血性脑卒中早期预防和精准治疗的关键。
目前，在颈动脉和冠状动脉斑块的研究中，由于三维高分辨磁共振血管壁成像的数据量巨大，医生需要手动重建、配准、分割及标记的方式完成血管壁斑块的前期处理工作，每位检查者的图像可达到500幅，需要花费大量的时间才能完成一名检查者的诊断，而且，颅内动脉管径细小（1-2mm）、形态卷绕的特征，血管壁上斑块的图像分割标记过程较为繁琐。因此，采用磁共振对缺血性脑卒中相关血管床斑块进行全面、精确的影像评估，并利用人工智能进行快速准确诊断，对脑卒中高危人群筛查和病因探查防止再发具有十分重要的意义。
 技术问题
本发明的目的在于提供一用于斑块分割的模型训练方法、装置、设备以及存储介质，旨在解决由于现有技术无法提供一种有效的斑块分割模型，导致在分割血管壁上的斑块时人工操作繁琐以及斑块分割耗时长的问题。
 技术解决方案
一方面，本发明提供了一种用于斑块分割的模型训练方法，所述方法包括下述步骤：
获取第一预设数量张血管壁的未分割斑块图像和所述未分割斑块图像对应的已分割斑块图像；
将所述未分割斑块图像输入到深度学习网络模型进行分割训练，以得到所述所述未分割斑块图像的第一分割结果；
通过所述第一分割结果和所述已分割斑块图像对所述分割训练后的深度学习网络模型进行微调迭代训练；
输出微调迭代训练后的用于斑块分割的所述深度学习网络模型。
另一方面，本发明提供了一种用于斑块分割的模型训练装置，所述装置包括：
图像获取单元，用于获取第一预设数量张血管壁的未分割斑块图像和所述未分割斑块图像对应的已分割斑块图像；
分割训练单元，用于将所述未分割斑块图像输入到深度学习网络模型进行分割训练，以得到所述未分割斑块图像的第一分割结果；
微调训练单元，用于通过所述第一分割结果和所述已分割斑块图像对所述分割训练后的深度学习网络模型进行微调迭代训练；以及
模型输出单元，用于输出微调迭代训练后的用于斑块分割的所述深度学习网络模型。
另一方面，本发明还提供了一种模型训练设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上述用于斑块分割的模型训练方法的步骤。
另一方面，本发明还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现如上述用于斑块分割的模型训练方法的步骤。
 有益效果
本发明获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率，进而提高斑块分割效率。
 附图说明
图1是本发明实施例一提供的用于斑块分割的模型训练方法的实现流程图；
图2是本发明实施例二提供的用于斑块分割的模型训练方法的实现流程图；
图3是本发明实施例二提供的第一分割结果的微调过程示意图；
图4是本发明实施例三提供的用于斑块分割的模型训练装置的结构示意图；
图5是本发明实施例四提供的用于斑块分割的模型训练装置的结构示意图；
图6是是本发明实施例四提供的图像微调单元的优选结构示意图；以及
图7是本发明实施例五提供的一种模型训练设备的结构示意图。
 本发明的实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
以下结合具体实施例对本发明的具体实现进行详细描述：
实施例一：
图1示出了本发明实施例一提供的用于斑块分割的模型训练方法的实现流程，为了便于说明，仅示出了与本发明实施例相关的部分，详述如下：
在步骤S101中，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像。
本发明实施例适用于模型训练设备，该模型用于对包含斑块的血管壁图像进行斑块分割（标记血管壁图像中的斑块部分）。在本发明实施例中，未分割斑块图像为原始获取的未标记出斑块部分的血管壁图像，已分割斑块图像为预先对未分割斑块图像上的斑块部分进行勾勒（标记）后的血管壁图像，具体地，可通过ITK-SNAP进行勾勒，先获取血管壁的未分割斑块图像和对应的已分割斑块图像，以便于后续对模型进行训练，其中，未分割斑块图像和已分割斑块图像的数量是相同的，未分割斑块图像是通过磁共振成像获取的三维MRI图像。
在步骤S102中，将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果。
在本发明实施例中，深度学习网络模型为U-Net网络，由于该深度学习网络只能对二维图像进行卷积，因此，先获取血管壁的三维MRI图像的一定数量张二维切面图，再将这一定数量张二维切面图输入到深度学习网络进行分割训练，也可以根据圆柱侧面展开后为长方形的原理，获取血管壁的三维MRI图像生成的血管壁展开图，再将该血管壁展开图输入到深度学习网络进行分割训练。
优选地，该深度学习网络模型为V-Net网络，该V-Net网络接收的图像为三维图像，从而简化了分割训练过程以及增加了分割斑块的精确度。
在进行分割训练时，根据未分割斑块图像的预设图像特征对未分割斑块图像进行分割，从而快速地得到未分割斑块图像的第一分割结果，该预设图像特征为斑块部分图像的特征，优选地，该预设图像特征为未分割斑块图像中斑块部分的像素密度特征，从而加快了模型的训练速度。
在步骤S103中，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练。
强化学习，又称再励学习、评价学习，是智能体（Agent）以“试错”的方式进行学习，通过与环境进行交互获得的奖赏指导行为，目标是使智能体获得最大的奖赏，强化学习不同于连接主义学习中的监督学习，主要表现在教师信号上，强化学习中由环境提供的强化信号是对产生动作的好坏作一种评价(通常为标量信号)，而不是告诉强化学习系统RLS(reinforcement learning system)如何去产生正确的动作，由于外部环境提供的信息很少，RLS必须靠自身的经历进行学习。通过这种方式，RLS在行动-评价的环境中获得知识，改进行动方案以适应环境。
在本发明实施例中，由于第一分割结果并不是很好的分割结果，在此，我们引入强化学习方法，通过该第一分割结果进行训练深度学习网络模型，通过对第一分割结果执行调整动作，根据已分割斑块图像对该次调整动作进行评价，从而实现对分割训练后的深度学习网络模型进行微调迭代训练。
在步骤S104中，输出微调迭代训练后的用于斑块分割的深度学习网络模型。
在本发明实施例中，当训练完分割训练后的深度学习网络模型后，输出微调迭代训练后的深度学习网络模型，该深度学习网络模型可对血管壁上的斑块部分图像进行精准够勾勒。
在本发明实施例中，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率。
实施例二：
图2示出了本发明实施例二提供的用于斑块分割的模型训练方法的实现流程，为了便于说明，仅示出了与本发明实施例相关的部分，详述如下：
在步骤S201中，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像。
本发明实施例适用于模型训练设备，该模型用于对包含斑块的血管壁图像进行斑块分割。在本发明实施例中，已分割斑块图像为预先对未分割斑块图像上的斑块部分进行勾勒后的血管壁图像，具体地，可通过ITK-SNAP进行勾勒，具体地，可通过ITK-SNAP进行勾勒，先获取血管壁的未分割斑块图像和对应的已分割斑块图像，以便于后续对模型进行训练，其中，未分割斑块图像和已分割斑块图像的数量是相同的，未分割斑块图像是通过磁共振成像获取的三维MRI图像。
在步骤S202中，对未分割斑块图像和已分割斑块图像进行图像预处理操作。
在本发明实施例中，由于U/V-Net网络要求输入的图像尺寸必须相同，因此，在将图像输入深度学习网络之前，先对未分割斑块图像和已分割斑块图像进行图像尺寸归一化操作，优选地，对未分割斑块图像和已分割斑块图像进行图像二值化操作，从而使得图像的特征更明显，便于后续对未分割斑块图像的特征提取。
在步骤S203中，将预处理后的未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果。
在本发明实施例中，深度学习网络模型为U-Net网络，由于该深度学习网络只能对二维图像进行卷积，因此，先获取血管壁的三维MRI图像的一定数量张二维切面图，再将这一定数量张二维切面图输入到深度学习网络进行分割训练，也可以根据圆柱侧面展开后为长方形的原理，获取血管壁的三维MRI图像生成的血管壁展开图，再将该血管壁展开图输入到深度学习网络进行分割训练。
在进行分割训练时，根据未分割斑块图像的预设图像特征对未分割斑块图像进行分割，从而快速地得到未分割斑块图像的第一分割结果，该预设图像特征为斑块部分图像的特征，优选地，该预设图像特征为未分割斑块图像中斑块部分的像素密度特征，从而加快了模型的训练速度。
优选地，该深度学习网络结合残差函数，以使深度学习网络中每一个卷积层的输入为上一卷积层的输入以及输出之和，从而丰富了卷积层的信息输入。进一步优选地，在深度学习网络中叠加Dense网络，从而加强了特征传递以及减轻了梯度消失，进而减少了斑块分割模型的模型参数。
优选地，在预设深度学习网络中进行提取预设特征时，在卷积得到的特征网络中添加膨胀卷积得到的特征网络，从而增大了图像的感受野，使卷积层的特征输出包含更多的信息。
在步骤S204中，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练。
在本发明实施例中，由于第一分割结果并不是很好的分割结果，在此，我们引入强化学习方法，通过该第一分割结果进行训练深度学习网络模型，通过对第一分割结果执行调整动作，得到第二分割结果，根据已分割斑块图像对该次调整动作（或者第二分割结果）进行评价，从而实现对分割训练后的深度学习网络模型进行微调迭代训练。具体地，先根据已分割斑块图像对第一分割结果进行微调，以得到未分割斑块图像的第二分割结果，再计算第二分割结果与已分割斑块图像的相似度，根据相似度判断是否继续对分割训练后的深度学习网络模型进行微调迭代训练。若继续对深度学习网络模型进行微调迭代训练，则表明当前训练得到的模型的分割精确度不够，具体地，可通过Dice系数来计算已分割斑块图像与第二分割结果的相似度。
优选地，在根据已分割斑块图像对第一分割结果进行微调时，分别获取第一分割结果和已分割斑块图像上分割边界的第二预设数量个相同方位的坐标（为便于后续描述，将这些坐标称为方位坐标），根据已分割斑块图像的方位坐标调整第一分割结果的方位坐标，从而增加了微调的准确度，进而减少了模型的训练时长。由于方位坐标个数与训练精度成正比，但是方位坐标个数又与训练时长成反比，因此，优选地，第二预设数量个相同方位为东、南、西、北、东南、西南、西北、东北，八个方位坐标，从而使训练时长和训练精度得到相对平衡。
作为示例地，图3示出了第一分割结果的微调过程，左图为已分割斑块图像，右图为第一分割结果，其中，左右两图的方位坐标都设置为东、南、西、北共4个方位坐标，已分割斑块图像中使用A 1 、A 2、A 3、A 4表示其东、南、西、北4个方位坐标，第一分割结果中使用B 1 、B 2、B 3、B 4表示其东、南、西、北4个方位坐标，此时，B 1向A 1靠近，B 2向A 2靠近，B 3向A 3靠近，B 4向A 4靠近第一分割结果的方位坐标B 1 、B 2、B 3、B 4分别根据方位坐标A 1 、A 2、A 3、A 4进行调整。
在步骤S205中，输出微调迭代训练后的用于斑块分割的深度学习网络模型。
在本发明实施例中，对分割训练后的深度学习网络模型进行微调迭代训练时，计算了第二分割结果与已分割斑块图像的相似度，若相似度大于预设相似度阈值时，表明当前训练得到的模型的分割精确度已达到预期值，此时，完成对深度学习网络模型的微调迭代训练，并输出微调迭代训练后的深度学习网络模型，该深度学习网络模型可对血管壁上的斑块部分图像进行精准够勾勒。
在本发明实施例中，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，并进行图像预处理操作，将预处理后的未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率，进而提高斑块分割效率。
实施例三：
图4示出了本发明实施例三提供的用于斑块分割的模型训练装置的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
图像获取单元41，用于获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像；
分割训练单元42，用于将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果；
微调训练单元43，用于通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练；以及
模型输出单元44，用于输出微调迭代训练后的用于斑块分割的深度学习网络模型。
在本发明实施例中，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率。
在本发明实施例中，用于斑块分割的模型训练装置的各单元可由相应的硬件或软件单元实现，各单元可以为独立的软、硬件单元，也可以集成为一个软、硬件单元，在此不用以限制本发明。各单元的具体实施方式可参考实施例一的描述，在此不再赘述。
实施例四：
图5示出了本发明实施例四提供的用于斑块分割的模型训练装置的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
图像获取单元51，用于获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像；
分割训练单元52，用于对未分割斑块图像和已分割斑块图像进行图像预处理操作；
微调训练单元53，用于将预处理后的未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果；以及
模型输出单元54，用于输出微调迭代训练后的用于斑块分割的深度学习网络模型。
其中，所述微调训练单元53包括：
图像微调单元531，用于根据已分割斑块图像对第一分割结果进行微调，以得到未分割斑块图像的第二分割结果；以及
计算判断单元532，用于计算第二分割结果与已分割斑块图像的相似度，根据相似度判断是否继续对分割训练后的深度学习网络模型进行微调迭代训练。
图6示出了图像微调单元531的优选结构，优选地，图像微调单元531包括：
坐标获取单元61，用于分别获取第一分割结果和已分割斑块图像上分割边界的第二预设数量个相同方位的坐标；以及
坐标调整单元62，用于根据已分割斑块图像的坐标调整第一分割结果的坐标。
在本发明实施例中，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，并进行图像预处理操作，将预处理后的未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率，进而提高斑块分割效率。
在本发明实施例中，用于斑块分割的模型训练装置的各单元可由相应的硬件或软件单元实现，各单元可以为独立的软、硬件单元，也可以集成为一个软、硬件单元，在此不用以限制本发明。各单元的具体实施方式可参考实施例二的描述，在此不再赘述。
实施例五：
图7示出了本发明实施例五提供的模型训练设备的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
本发明实施例的模型训练设备7包括处理器71、存储器72以及存储在存储器72中并可在处理器71上运行的计算机程序73。该处理器71执行计算机程序73时实现上述各个用于斑块分割的模型训练方法实施例中的步骤，例如图1所示的步骤S101至S104以及图2所示的步骤S201至S205。或者，处理器71执行计算机程序73时实现上述各个用于斑块分割的模型训练装置实施例中各单元的功能，例如图4所示单元41至44以及图5所示单元51至54的功能。
在本发明实施例中，该处理器执行计算机程序时，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率。
该处理器执行计算机程序时实现上述用于斑块分割的模型训练方法实施例中的步骤可参考实施例一和实施例二的描述，在此不再赘述。
实施例六：
在本发明实施例中，提供了一种计算机可读存储介质，该计算机可读存储介质存储有计算机程序，该计算机程序被处理器执行时实现上述各个用于斑块分割的模型训练方法实施例中的步骤，例如，图1所示的步骤S101至S104以及图2所示的步骤S201至S205。或者，该计算机程序被处理器执行时实现上述各个用于斑块分割的模型训练装置实施例中各单元的功能，例如图4所示单元41至44以及图5所示单元51至54的功能。
在本发明实施例中，在计算机程序被处理器执行后，获取第一预设数量张血管壁的未分割斑块图像和未分割斑块图像对应的已分割斑块图像，将未分割斑块图像输入到深度学习网络模型进行分割训练，以得到未分割斑块图像的第一分割结果，基于强化学习方法，通过第一分割结果和已分割斑块图像对分割训练后的深度学习网络模型进行微调迭代训练，输出微调迭代训练后的用于斑块分割的深度学习网络模型，从而实现了血管壁上斑块分割的自动化以及提高了斑块分割的精确度，进而提高斑块分割效率，进而提高斑块分割效率。
本发明实施例的计算机可读存储介质可以包括能够携带计算机程序代码的任何实体或装置、存储介质，例如，ROM/RAM、磁盘、光盘、闪存等存储器。
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种用于斑块分割的模型训练方法，其特征在于，所述方法包括下述步骤：
    获取第一预设数量张血管壁的未分割斑块图像和所述未分割斑块图像对应的已分割斑块图像；
    将所述未分割斑块图像输入到深度学习网络模型进行分割训练，以得到所述未分割斑块图像的第一分割结果；
    通过所述第一分割结果和所述已分割斑块图像对所述分割训练后的深度学习网络模型进行微调迭代训练；
    输出微调迭代训练后的用于斑块分割的所述深度学习网络模型。
  2. 如权利要求1所述的方法，其特征在于，通过所述第一分割结果和所述已分割斑块图像对所述分割训练后的深度学习网络模型进行微调迭代训练的步骤，包括：
    根据所述已分割斑块图像对所述第一分割结果进行微调，以得到所述未分割斑块图像的第二分割结果；
    计算所述第二分割结果与所述已分割斑块图像的相似度，根据所述相似度判断是否继续对分割训练后的所述深度学习网络模型进行微调迭代训练。
  3. 如权利要求2所述的方法，其特征在于，根据所述已分割斑块图像对所述第一分割结果进行微调的步骤，包括：
    分别获取所述第一分割结果和所述已分割斑块图像上分割边界的第二预设数量个相同方位的方位坐标；
    根据所述已分割斑块图像的方位坐标调整所述第一分割结果的方位坐标。
  4. 如权利要求1所述的方法，其特征在于，将所述未分割斑块图像输入到深度学习网络模型进行分割训练的步骤，包括：
    根据所述未分割斑块图像的预设图像特征对所述未分割斑块图像进行分割。
  5. 如权利要求1所述的方法，其特征在于，获取第一预设数量张血管壁的未分割斑块图像和所述未分割斑块图像对应的已分割斑块图像的步骤之后，将所述未分割斑块图像输入到深度学习网络模型进行分割训练的步骤之前，所述方法还包括：
    对所述未分割斑块图像和所述已分割斑块图像进行图像预处理操作，所述图像预处理操作包括尺寸的归一化操作。
  6. 一种用于斑块分割的模型训练装置，其特征在于，所述装置包括：
    图像获取单元，用于获取第一预设数量张血管壁的未分割斑块图像和所述未分割斑块图像对应的已分割斑块图像；
    分割训练单元，用于将所述未分割斑块图像输入到深度学习网络模型进行分割训练，以得到所述未分割斑块图像的第一分割结果；
    微调训练单元，用于通过所述第一分割结果和所述已分割斑块图像对所述分割训练后的深度学习网络模型进行微调迭代训练；以及
    模型输出单元，用于输出微调迭代训练后的用于斑块分割的所述深度学习网络模型。
  7. 如权利要求6所述的装置，其特征在于，所述微调训练调单元包括：
    图像微调单元，用于根据所述已分割斑块图像对所述第一分割结果进行微调，以得到所述未分割斑块图像的第二分割结果；以及
    计算判断单元，用于计算所述第二分割结果与所述已分割斑块图像的相似度，根据所述相似度判断是否继续对分割训练后的所述深度学习网络模型进行微调迭代训练。
  8. 如权利要求7所述的装置，其特征在于，所述图像微调单元包括：
    坐标获取单元，用于分别获取所述第一分割结果和所述已分割斑块图像上分割边界的第二预设数量个相同方位的坐标；以及
    坐标调整单元，用于根据所述已分割斑块图像的坐标调整所述第一分割结果的坐标。
  9. 一种模型训练设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1至5项所述方法的步骤。
  10. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至5项所述方法的步骤。
PCT/CN2019/078469 2018-09-20 2019-03-18 用于斑块分割的模型训练方法、装置、设备及存储介质 WO2020057074A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811100334.0A CN109087708B (zh) 2018-09-20 2018-09-20 用于斑块分割的模型训练方法、装置、设备及存储介质
CN201811100334.0 2018-09-20

Publications (1)

Publication Number Publication Date
WO2020057074A1 true WO2020057074A1 (zh) 2020-03-26

Family

ID=64841977

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/078469 WO2020057074A1 (zh) 2018-09-20 2019-03-18 用于斑块分割的模型训练方法、装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN109087708B (zh)
WO (1) WO2020057074A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109087708B (zh) * 2018-09-20 2021-08-31 深圳先进技术研究院 用于斑块分割的模型训练方法、装置、设备及存储介质
CN109741346B (zh) * 2018-12-30 2020-12-08 上海联影智能医疗科技有限公司 感兴趣区域提取方法、装置、设备及存储介质
CN109829892A (zh) * 2019-01-03 2019-05-31 众安信息技术服务有限公司 一种预测模型的训练方法、使用该模型的预测方法及装置
CN111598891B (zh) * 2019-02-20 2023-08-08 深圳先进技术研究院 斑块稳定性的识别方法、装置、设备及存储介质
CN113192062A (zh) * 2021-05-25 2021-07-30 湖北工业大学 一种基于图像复原的动脉斑块超声图像自监督分割方法
CN113538471B (zh) * 2021-06-30 2023-09-22 上海联影医疗科技股份有限公司 斑块的分割方法、装置、计算机设备和存储介质
CN113940640B (zh) * 2021-11-12 2023-04-18 清华大学 心血管疾病风险控制方法、系统及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104794708A (zh) * 2015-04-10 2015-07-22 浙江工业大学 一种基于多特征学习的动脉粥样硬化斑块成分分割方法
CN106408562A (zh) * 2016-09-22 2017-02-15 华南理工大学 基于深度学习的眼底图像视网膜血管分割方法及系统
CN106558058A (zh) * 2016-11-29 2017-04-05 北京图森未来科技有限公司 分割模型训练方法、道路分割方法、车辆控制方法及装置
CN109087708A (zh) * 2018-09-20 2018-12-25 深圳先进技术研究院 用于斑块分割的模型训练方法、装置、设备及存储介质

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108229479B (zh) * 2017-08-01 2019-12-31 北京市商汤科技开发有限公司 语义分割模型的训练方法和装置、电子设备、存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104794708A (zh) * 2015-04-10 2015-07-22 浙江工业大学 一种基于多特征学习的动脉粥样硬化斑块成分分割方法
CN106408562A (zh) * 2016-09-22 2017-02-15 华南理工大学 基于深度学习的眼底图像视网膜血管分割方法及系统
CN106558058A (zh) * 2016-11-29 2017-04-05 北京图森未来科技有限公司 分割模型训练方法、道路分割方法、车辆控制方法及装置
CN109087708A (zh) * 2018-09-20 2018-12-25 深圳先进技术研究院 用于斑块分割的模型训练方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN109087708A (zh) 2018-12-25
CN109087708B (zh) 2021-08-31

Similar Documents

Publication Publication Date Title
WO2020057074A1 (zh) 用于斑块分割的模型训练方法、装置、设备及存储介质
CN111091589B (zh) 基于多尺度监督学习的超声和核磁图像配准方法及装置
US10769791B2 (en) Systems and methods for cross-modality image segmentation
CN111047594B (zh) 肿瘤mri弱监督学习分析建模方法及其模型
CN109544510B (zh) 一种基于卷积神经网络的三维肺结节识别方法
CN109410188B (zh) 用于对医学图像进行分割的系统和方法
De Craene et al. 3D strain assessment in ultrasound (straus): A synthetic comparison of five tracking methodologies
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
CN107481252A (zh) 医学图像的分割方法、装置、介质和电子设备
CN107680110B (zh) 基于统计形状模型的内耳三维水平集分割方法
CN111080681B (zh) 一种基于LoG算子的3D/2D医学图像配准方法
CN103325143A (zh) 基于模型匹配的标记点自动注册方法
WO2017092973A1 (en) Removal of image artifacts in sense-mri
CN110838140A (zh) 基于混合监督学习的超声和核磁图像配准融合方法及装置
CN113674330A (zh) 一种基于生成对抗网络的伪ct影像生成系统
WO2023124830A1 (zh) 一种基于中心线提取的血管壁影像自动曲面重建方法
CN111815593B (zh) 基于对抗学习的肺结节域适应分割方法、装置及存储介质
CN108898578B (zh) 一种医疗图像的处理方法、装置及计算机存储介质
CN114266939A (zh) 一种基于ResTLU-Net模型的脑提取方法
CN110458842B (zh) 基于双通道三维密集连接网络的脑肿瘤分割方法
CN109559296B (zh) 基于全卷积神经网络和互信息的医学图像配准方法及系统
CN100411587C (zh) 基于机器学习的立体核磁共振脑图像弹性配准方法
CN109741439A (zh) 一种二维mri胎儿图像的三维重建方法
CN112581385B (zh) 基于多先验约束扩散峰度成像张量估计方法、介质和设备
CN117911432A (zh) 图像分割方法、装置及存储介质

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: 19863223

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: 19863223

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 09/09/2021)