WO2020007026A1 - 分割模型训练方法、装置及计算机可读存储介质 - Google Patents

分割模型训练方法、装置及计算机可读存储介质 Download PDF

Info

Publication number
WO2020007026A1
WO2020007026A1 PCT/CN2019/071501 CN2019071501W WO2020007026A1 WO 2020007026 A1 WO2020007026 A1 WO 2020007026A1 CN 2019071501 W CN2019071501 W CN 2019071501W WO 2020007026 A1 WO2020007026 A1 WO 2020007026A1
Authority
WO
WIPO (PCT)
Prior art keywords
segmentation result
image
cardiac
segmentation
similarity
Prior art date
Application number
PCT/CN2019/071501
Other languages
English (en)
French (fr)
Inventor
胡战利
吴垠
梁栋
杨永峰
刘新
郑海荣
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 WO2020007026A1 publication Critical patent/WO2020007026A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10072Tomographic images
    • G06T2207/10081Computed x-ray tomography [CT]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30048Heart; Cardiac

Definitions

  • the present invention relates to the field of image processing technology, and in particular, to a method, a device, and a computer-readable storage medium for training a segmentation model.
  • Cardiovascular disease is a serious threat to human life and health, and early quantitative diagnosis and risk assessment of cardiovascular disease plays a key role in extending human life and health.
  • CT computed tomography
  • the rapid development of Tomography (CT) technology has continuously affected the diagnosis of human diseases, and has gradually become an important diagnostic method of cardiac examination.
  • the ventricular area is the core area of the heart and has always been the focus of heart disease research.
  • the study of cardiac tissues, especially the left ventricle, with the help of cardiac CT images is very meaningful.
  • the main purpose of the present invention is to provide a segmentation model training method, which aims to solve the technical problem that the prior art lacks a tool capable of automatically performing ventricular myocardial segmentation.
  • a first aspect of the present invention provides a segmentation model training method, including:
  • a second aspect of the present invention provides a segmentation model training device, where the device includes:
  • An acquisition module for acquiring multiple CT computed tomography CT images and acquiring ventricular position information that has been outlined in each cardiac CT image
  • a training module configured to input the cardiac CT image into a deep learning network model for training, and obtain a first segmentation result of each of the cardiac CT images output by the deep learning network model;
  • a fine-tuning module configured to fine-tune the first segmentation result based on the reinforcement learning device according to the ventricular position information and the first segmentation result to obtain a second segmentation result of each of the cardiac CT images;
  • An iterative module configured to iteratively train the deep learning network model according to the second segmentation result and the ventricular position information, and use the trained deep learning network model as the segmentation model.
  • the third aspect of the present invention further provides a computer-readable storage medium having stored thereon a computer program that, when executed by a processor, implements each of the segmentation model training methods described in the first aspect. step.
  • the invention provides a segmentation model training method, a device and a computer-readable storage medium.
  • the method includes: acquiring multiple cardiac CT images, and acquiring ventricular position information outlined by each cardiac CT image, and inputting the cardiac CT image to depth Learning the network model for training to obtain the first segmentation result of each heart CT image output by the deep learning network model.
  • the first segmentation result is fine-tuned according to the ventricular position information and the first segmentation result to obtain each heart.
  • the second segmentation result of the CT image is iteratively trained according to the second segmentation result and the ventricular position information of the above-mentioned cardiac CT images, and the trained deep learning network model is used as a segmentation model for segmenting the heart muscle. .
  • the deep learning network model is trained by using cardiac CT images and ventricular position information of each cardiac CT image as training data, so that a segmentation model capable of automatically performing ventricular segmentation on cardiac CT images can be obtained, and further By combining fine-tuning and iterative training with reinforcement learning methods, a segmentation model with higher segmentation accuracy can be obtained.
  • FIG. 1 is a schematic flowchart of a segmentation model training method according to an embodiment of the present invention
  • FIG. 2 is another schematic flowchart of a segmentation model training method according to an embodiment of the present invention.
  • FIG. 3 is a schematic diagram of a first segmentation result and ventricular position information in an embodiment of the present invention
  • FIG. 4 is a schematic structural diagram of a segmentation model training device according to an embodiment of the present invention.
  • FIG. 5 is another schematic structural diagram of a segmentation model training apparatus according to an embodiment of the present invention.
  • FIG. 1 is a schematic flowchart of a segmentation model training method according to an embodiment of the present invention, including:
  • Step 101 Collect multiple cardiac CT images to obtain ventricular position information that has been outlined in each cardiac CT image
  • the above-mentioned segmented model training method may be implemented by a segmented model training device, which is composed of program modules and stored in a training device (such as a computer), and a processor in the training device may call the above Segmentation model training device to implement the above-mentioned segmentation model training method.
  • the training data needs to be prepared before training.
  • the training data includes multiple cardiac CT images and ventricular position information of each cardiac CT image.
  • the multiple cardiac CT images may be cardiac CT acquired through clinical methods. Images, and the ventricular position information of each cardiac CT image may be manually outlined, or may be highly accurate ventricular position information determined by other methods.
  • Step 102 input the cardiac CT image to a deep learning network model for training, and obtain a first segmentation result of each of the cardiac CT images output by the deep learning network model;
  • each cardiac CT image may also be pre-processed, and the pre-processing is specifically a normalization process for the size of each cardiac CT image. This is to better meet the training requirements and to obtain a more accurate segmentation model.
  • a plurality of cardiac CT images as training data are input to a deep learning network model for training, and a first segmentation result of each cardiac CT image output by the deep learning network model is obtained.
  • a deep learning network model is used, and the deep learning network model is trained to obtain a segmentation model capable of accurately segmenting ventricular myocardium.
  • the deep learning network model may specifically be a V-NET network model or a model of a convolutional neural network type. In actual applications, the model may be selected according to specific needs, which is not limited here.
  • Step 103 Based on the reinforcement learning method, fine-tune the first segmentation result according to the ventricular position information and the first segmentation result to obtain a second segmentation result of each of the cardiac CT images;
  • reinforcement learning also called re-energizing learning and evaluation learning
  • the so-called reinforcement learning is the learning of the intelligent system from environment to behavior mapping to maximize the value of the reinforcement signal function.
  • the first segmentation result is fine-tuned according to the ventricular position information and the first segmentation result to obtain the second segmentation result of each heart CT image.
  • Step 104 Iteratively train the deep learning network model according to the second segmentation result and the ventricular position information, and use the trained deep learning network model as the segmentation model.
  • the trained deep learning network model has high segmentation accuracy and can be used as a segmentation model. And using this segmentation model to segment cardiac CT images, it is possible to obtain ventricular position information with high segmentation accuracy.
  • a deep learning network model is trained by using cardiac CT images and ventricular position information of each cardiac CT image as training data, so that a segmentation model capable of automatically performing ventricular segmentation on cardiac CT images can be obtained, and further By combining fine-tuning and iterative training with reinforcement learning methods, a segmentation model with higher segmentation accuracy can be obtained.
  • FIG. 2 is another flowchart of the segmentation model training method in the embodiment of the present invention, including:
  • Step 201 Collect multiple CT computed tomography CT images to obtain ventricular position information that has been outlined in each cardiac CT image;
  • Step 202 input the cardiac CT image to a deep learning network model for training, and obtain a first segmentation result of each of the cardiac CT images output by the deep learning network model;
  • step 201 and step 202 the content described in step 201 and step 202 is similar to that in step 101 and step 102 in the embodiment shown in FIG. 1.
  • step 101 and step in the embodiment shown in FIG. 1. 102 I will not repeat them here.
  • Step 203 Calculate the similarity between the ventricular position information of each cardiac CT image and the first segmentation result of each cardiac CT image;
  • Step 204 Fine-tune the first segmentation result of each heart CT image according to the similarity, to obtain a second segmentation result of each heart CT image;
  • the second segmentation result of each heart CT image is obtained based on the reinforcement learning method.
  • the reinforcement learning method involves reward and punishment, which can be judged by the Dice coefficient.
  • the Dice coefficient can be used to determine the similarity of the two sets. Degree, that is, the similarity of two sets can be calculated, the two sets are the set composed of the first segmentation results of each heart CT image, and the other set is the set composed of ventricular position information of each heart CT image,
  • a similarity calculation formula will be used to calculate the similarity between the ventricular position information of each cardiac CT image and the first segmentation result of each cardiac CT image, where the similarity calculation formula is as follows:
  • D represents the similarity between the set consisting of the first segmentation results of each cardiac CT image and the set consisting of the ventricular position information of each cardiac CT image
  • N represents the number of cardiac CT images
  • Ventricular position information representing the i-th heart CT image Represents the first segmentation result of the i-th heart CT image.
  • the calculated similarity is compared with a preset value.
  • the preset value is usually 1 and when the similarity is equal to 1, it indicates that the first segmentation result is completed correctly. At this time, it indicates that the deep learning has been completed. Training of network models.
  • the similarity is not equal to the above-mentioned preset value, it will be determined whether the first segmentation result of each cardiac CT image needs to be fine-tuned based on the similarity.
  • the fine-tuned second segmentation result is input to a deep learning network model for training, and the output first segmentation result is obtained, and then returns to step 203 to calculate the similarity, and then continues the similarity to determine whether to perform Fine-tuning. Therefore, a similarity will be calculated for each iteration process, and the calculated similarity can be saved. And after each similarity is calculated, the similarity is used as the current similarity, and the current similarity is compared with the similarity calculated last time. When the current similarity is less than the similarity calculated last time, , Indicating that the training of the deep learning network model has not improved the accuracy of the deep learning network model.
  • the reward value (reward) in the enhanced learning network is reduced.
  • Go to one, and at the same time use the first segmentation result of each cardiac CT image as the second segmentation result of each cardiac CT image, so as to subsequently use the second segmentation result for iteration.
  • the current similarity is greater than the similarity calculated last time, it indicates that the training of the deep learning network model has improved the accuracy of the deep learning network model, and the first segmentation results of each heart CT image will be fine-tuned.
  • the second segmentation result of each heart CT image is obtained.
  • the fine-tuning of the first segmentation result of the cardiac CT image may be at least one of moving up, moving down, moving left, moving right, reducing, enlarging, and maintaining the same. As for the specific fine-tuning method, it needs to be adjusted. It is determined based on the first segmentation result of the cardiac CT image and the ventricular position information of the cardiac CT image.
  • FIG. 3 is a schematic diagram of the first segmentation result and ventricular position information in the embodiment of the present invention.
  • the standard ventricular position information outlined in the method includes four coordinate points of up, down, left, and right, which are represented by Gt_top, Gt_bottom, Gt_left, and Gt_right.
  • the image on the right is the first segmentation result output by the deep learning network model.
  • the coordinate points are represented by V_top, V_bottom, V_left, and V_right, respectively, and judged as follows to determine the fine-tuning method:
  • the preset value of the movement can be any value in the range of 1 mm to 10 mm.
  • the preset multiple can be any multiple within the range of 0.01 to 0.2.
  • the second segmentation result may be obtained by fine-tuning the first segmentation result.
  • the deep learning network model will be iteratively trained based on the second segmentation results of each cardiac CT image and the ventricular position information, and the trained deep learning network model will be used as a segmentation model.
  • the segmentation model can accurately compare the cardiac CT images. For high segmentation, please refer to step 205 and step 206.
  • Step 205 Calculate an error between the second segmentation result of each cardiac CT image and the ventricular position information of each cardiac CT image;
  • Step 206 When the maximum error is greater than or equal to the error threshold, input the second segmentation result to the deep learning network model for iterative training until the calculated maximum error is less than the error threshold.
  • the error between the second segmentation result of each cardiac CT image and the ventricular position information of each cardiac CT image will be sequentially calculated.
  • the second segmentation result of each heart CT image is input into a deep learning network model for training.
  • the deep learning network model will output the first segmentation result and return to step 203 to perform iteration until the number of iterations reaches a preset value. Or the calculated maximum error is less than the error threshold.
  • the deep learning network model after the iteration is used as a segmentation model for segmenting cardiac CT images.
  • a deep learning network model is introduced in the embodiment of the present invention, so that the cardiac CT image can be segmented by using the deep learning network model, which can significantly improve the accuracy of segmentation and introduces reinforcement based on deep learning Learning method, using reinforcement learning to fine-tune the segmentation results and using the fine-tuned segmentation results to iteratively train the deep learning network model.
  • the segmentation accuracy of the deep learning network model can be further improved and used as a segmentation model. It can effectively realize the segmentation of the ventricle of the cardiac CT image, and the segmentation accuracy is high, no manual sketching is needed, and it can effectively save labor costs and reduce the probability of ventricular segmentation deviation.
  • FIG. 4 is a schematic structural diagram of a segmentation model training apparatus according to an embodiment of the present invention.
  • the apparatus includes:
  • the acquisition and acquisition module 401 is configured to acquire a plurality of cardiac computed tomography CT images, and obtain ventricular position information that has been outlined by each cardiac CT image;
  • the training data needs to be prepared before training.
  • the training data includes multiple cardiac CT images and ventricular position information of each cardiac CT image.
  • the multiple cardiac CT images may be cardiac CT acquired through clinical methods. Images, and the ventricular position information of each cardiac CT image may be manually outlined, or may be highly accurate ventricular position information determined by other methods.
  • a training module 402 configured to input the cardiac CT image into a deep learning network model for training, and obtain a first segmentation result of each of the cardiac CT images output by the deep learning network model;
  • each cardiac CT image may also be pre-processed, and the pre-processing is specifically a normalization process for the size of each cardiac CT image. This is to better meet the training requirements and to obtain a more accurate segmentation model.
  • the deep learning network model may specifically be a V-NET network model or a model of a convolutional neural network type. In actual applications, the model may be selected according to specific needs, which is not limited here.
  • a fine-tuning module 403, configured to fine-tune the first segmentation result based on the reinforcement learning device according to the ventricular position information and the first segmentation result to obtain a second segmentation result of each of the cardiac CT images;
  • reinforcement learning also called re-energizing learning and evaluation learning
  • the so-called reinforcement learning is the learning of the intelligent system from environment to behavior mapping to maximize the value of the reinforcement signal function.
  • An iteration module 404 is configured to perform iterative training on the deep learning network model according to the second segmentation result and the ventricular position information, and use the trained deep learning network model as the segmentation model.
  • a deep learning network model is trained by using cardiac CT images and ventricular position information of each cardiac CT image as training data, so that a segmentation model capable of automatically performing ventricular segmentation on cardiac CT images can be obtained, and further By combining fine-tuning and iterative training with reinforcement learning methods, a segmentation model with higher segmentation accuracy can be obtained.
  • FIG. 5 is another schematic structural diagram of a segmentation model training device according to an embodiment of the present invention, including a collection and acquisition module 401, a training module 402, a fine-tuning module 403, and an iteration module 404 in the embodiment shown in FIG. 4. Contents described in the embodiment shown in FIG. 4 are similar, and details are not described herein.
  • the fine-tuning module 403 includes:
  • a calculation module 501 configured to calculate a similarity between the ventricular position information of each heart CT image and the first segmentation result of each heart CT image;
  • the result fine-tuning module 502 is configured to fine-tune the first segmentation result of each heart CT image according to the similarity, to obtain the second segmentation result of each heart CT image.
  • the result fine-tuning module 502 includes:
  • a comparison module 503 configured to compare the similarity with a similarity calculated last time when the similarity is not equal to a preset value
  • a first processing module 504 configured to determine that the first segmentation result of each cardiac CT image is not fine-tuned when the similarity is less than the similarity calculated last time, and that the first segmentation result of each cardiac CT image is not adjusted. As the second segmentation result of each heart CT image;
  • a second processing module 505 is configured to fine-tune the first segmentation result of each cardiac CT image when the similarity is greater than the similarity calculated last time to obtain the second segmentation result.
  • the second processing module 505 is specifically configured to compare the coordinate values of the various positions in the ventricular position information of the CT image of each heart with the corresponding coordinate values of the various positions of the first segmentation result to obtain a comparison result. Adjusting the coordinate values of the respective bits of the corresponding first segmentation result according to the comparison result to obtain the second segmentation result, the adjustments include: move up, move down, move left, move right, reduce, enlarge , At least one of the same.
  • the iteration module 404 includes:
  • An error calculation module 506 configured to calculate an error between the second segmentation result of each heart CT image and the ventricular position information of each heart CT image;
  • An iterative training module 507 is configured to input the second segmentation result into the deep learning network model when the maximum error is greater than or equal to the error threshold to perform iterative training until the calculated maximum error is less than the error threshold.
  • the second segmentation result of each heart CT image is obtained based on the reinforcement learning method.
  • the reinforcement learning method involves reward and punishment, which can be judged by the Dice coefficient.
  • the Dice coefficient can be used to determine the similarity of the two sets. Degree, that is, the similarity of two sets can be calculated, the two sets are the set composed of the first segmentation results of each heart CT image, and the other set is the set composed of ventricular position information of each heart CT image,
  • a similarity calculation formula will be used to calculate the similarity between the ventricular position information of each cardiac CT image and the first segmentation result of each cardiac CT image, where the similarity calculation formula is as follows:
  • D represents the similarity between the set consisting of the first segmentation results of each cardiac CT image and the set consisting of the ventricular position information of each cardiac CT image
  • N represents the number of cardiac CT images
  • Ventricular position information representing the i-th heart CT image Represents the first segmentation result of the i-th heart CT image.
  • the calculated similarity is compared with a preset value.
  • the preset value is usually 1 and when the similarity is equal to 1, it indicates that the first segmentation result is completed correctly. At this time, it indicates that the deep learning has been completed. Training of network models.
  • the similarity is not equal to the above-mentioned preset value, it will be determined whether the first segmentation result of each cardiac CT image needs to be fine-tuned based on the similarity.
  • the fine-tuned second segmentation result is input to a deep learning network model for training, and the output first segmentation result is obtained, and then returns to step 203 to calculate the similarity, and then continues the similarity to determine whether to perform Fine-tuning. Therefore, a similarity will be calculated for each iteration process, and the calculated similarity can be saved. And after each similarity is calculated, the similarity is used as the current similarity, and the current similarity is compared with the similarity calculated last time. When the current similarity is less than the similarity calculated last time, , Indicating that the training of the deep learning network model did not improve the accuracy of the deep learning network model.
  • the first segmentation result of each heart CT image is used as the second segmentation result of each cardiac CT image, so as to iterate using the second segmentation result subsequently.
  • the current similarity is greater than the similarity calculated last time, it indicates that the training of the deep learning network model has improved the accuracy of the deep learning network model, and the first segmentation results of each heart CT image will be fine-tuned.
  • the second segmentation result of each heart CT image is obtained.
  • the fine-tuning of the first segmentation result of the cardiac CT image may be at least one of moving up, moving down, moving left, moving right, reducing, enlarging, and maintaining the same.
  • FIG. 3 is a schematic diagram of the first segmentation result and ventricular position information in the embodiment of the present invention.
  • a segmentation diagram of the same cardiac CT image the left image is manually
  • the standard ventricular position information outlined in the method includes four coordinate points of up, down, left, and right, which are represented by Gt_top, Gt_bottom, Gt_left, and Gt_right.
  • the image on the right is the first segmentation result output by the deep learning network model.
  • the coordinate points are represented by V_top, V_bottom, V_left, and V_right, respectively, and judged as follows to determine the fine-tuning method:
  • the second segmentation result may be obtained by fine-tuning the first segmentation result.
  • a deep learning network model is introduced in the embodiment of the present invention, so that the cardiac CT image can be segmented by using the deep learning network model, which can significantly improve the accuracy of segmentation, and introduces reinforcement based on deep learning.
  • Learning method using reinforcement learning to fine-tune the segmentation results and using the fine-tuned segmentation results to iteratively train the deep learning network model.
  • the segmentation accuracy of the deep learning network model can be further improved and used as a segmentation model. It can effectively realize the segmentation of the ventricle of the cardiac CT image, and the segmentation accuracy is high, no manual sketching is needed, and it can effectively save labor costs and reduce the probability of ventricular segmentation deviation.
  • a computer-readable storage medium is also provided in the embodiment of the present invention, and a computer program is stored thereon.
  • the computer program is executed by a processor, the method for implementing the segmentation model training method in the embodiment shown in FIG. 1 or FIG. 2 is implemented. Each step.
  • the disclosed apparatus and method may be implemented in other ways.
  • the device embodiments described above are only schematic.
  • the division of the modules is only a logical function division.
  • multiple modules or components may be combined or Can be integrated into another system, or some features can be ignored or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or modules, and may be electrical, mechanical or other forms.
  • the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, which may be located in one place, or may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the objective of the solution of this embodiment.
  • each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist separately physically, or two or more modules may be integrated into one module.
  • the above integrated modules can be implemented in the form of hardware or software functional modules.
  • the integrated module When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the technical solution of the present invention essentially or part that contributes to the existing technology or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium , Including a number of instructions to enable a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.
  • the aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), and a random access memory (RAM, Random Access) Memory), magnetic disks or optical disks, and other media that can store program code.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供一种分割模型训练方法,装置及计算机可读存储介质,方法包括:将采集的心脏CT图像输入到深度学习网络模型进行训练,得到各心脏CT图像的第一分割结果,基于强化学习方法,根据心室位置信息及第一分割结果对该第一分割结果进行微调,得到第二分割结果,根据该第二分割结果及上述心室位置信息对深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为对心脏心肌进行分割的分割模型。通过将心脏CT图像及各心脏CT图像的心室位置信息作为训练数据对深度学习网络模型进行训练,使得能够得到可自动对心脏CT图像进行心室分割的分割模型,且进一步的通过结合强化学习方法进行微调及迭代训练,能够得到分割精度更高的分割模型。

Description

分割模型训练方法、装置及计算机可读存储介质 技术领域
本发明涉及图像处理技术领域,尤其涉及一种分割模型训练方法、装置及计算机可读存储介质。
背景技术
心血管疾病严重威胁着人类的生命健康,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。随着科技高速发展,影像诊断设备的功能和成像质量有了很大提高。尤其是计算机断层扫面(Computed Tomography,CT)技术的高速发展不断影响人体疾病的诊断方式,逐步成为心脏检查的一种重要的诊断方式。心室区域是心脏核心区域,一直是心脏疾病研究的重点,借助于心脏CT图像对心脏组织尤其是左心室的研究,是非常有意义的。
左心室形态和运动的异常始终被视为心血管临床诊断的重要依据,为帮助脑血管疾病(Cerebrovascular Disease,CVD)的诊断,医生致力于确定左心室容积、心肌壁厚度,并测量在心动周期的心室血量(射血分数)和管壁增厚性质的变化,然而上述数据都依赖于左心室心肌的正确分割。目前,心室心肌的分割主要是专家手动分割,然而,手工分割对专家知识和经验要求很高,而且不可避免出现误差。因此,提供一种能够自动进行心室心肌分割的工具是目前亟待解决的问题。
技术问题
本发明的主要目的在于提供一种分割模型训练方法,旨在解决现有技术中缺少能够自动进行心室心肌分割的工具的技术问题。
技术解决方案
为实现上述目的,本发明第一方面提供一种分割模型训练方法,包括:
采集多张心脏计算机断层扫面CT图像,获取各心脏CT图像已勾勒的心室位置信息;
将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
基于强化学习方法,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果;
根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为所述分割模型。
为实现上述目的,本发明第二方面提供一种分割模型训练装置,所述装置包括:
采集获取模块,用于采集多张心脏计算机断层扫面CT图像,获取各心脏CT图像已勾勒的心室位置信息;
训练模块,用于将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
微调模块,用于基于强化学习装置,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果;
迭代模块,用于根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为所述分割模型。
为实现上述目的,本发明第三方面还提供一种计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时,实现如第一方面所述的分割模型训练方法的各个步骤。
有益效果
本发明提供一种分割模型训练方法,装置及计算机可读存储介质,在方法包括:采集多张心脏CT图像,并获取各心脏CT图像已勾勒的心室位置信息,将该心脏CT图像输入到深度学习网络模型进行训练,得到该深度学习网络模型输出的各心脏CT图像的第一分割结果,基于强化学习方法,根据心室位置信息及第一分割结果对该第一分割结果进行微调,得到各心脏CT图像的第二分割结果,根据该第二分割结果及上述各心脏CT图像的心室位置信息对深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为对心脏心肌进行分割的分割模型。相对于现有技术,通过将心脏CT图像及各心脏CT图像的心室位置信息作为训练数据对深度学习网络模型进行训练,使得能够得到可自动对心脏CT图像进行心室分割的分割模型,且进一步的通过结合强化学习方法进行微调及迭代训练,能够得到分割精度更高的分割模型。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案,下面将对实施例或现有技术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,对于本领域技术人员来讲,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为本发明实施例中分割模型训练方法的流程示意图;
图2为本发明实施例中分割模型训练方法的另一流程示意图;
图3为本发明实施例中第一分割结果和心室位置信息的示意图;
图4为本发明实施例中分割模型训练装置的结构示意图;
图5为本发明实施例中分割模型训练装置的另一结构示意图。
本发明的实施方式
为使得本发明的发明目的、特征、优点能够更加的明显和易懂,下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而非全部实施例。基于本发明中的实施例,本领域技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。
请参阅图1,为本发明实施例中分割模型训练方法的流程示意图,包括:
步骤101、采集多张心脏CT图像,获取各心脏CT图像已勾勒的心室位置信息;
在发明实施例中,上述的分割模型训练方法可以由分割模型训练装置实现,该分割模型训练装置由程序模块构成,并存储在训练设备(如计算机)内,训练设备内的处理器可以调用上述的分割模型训练装置,以实现上述的分割模型训练方法。
其中,训练之前需要准备进行训练的训练数据,该训练数据包含多张心脏CT图像,及各心脏CT图像的心室位置信息,其中,该多张心脏CT图像可以是通过临床方式采集到的心脏CT图像,且各心脏CT图像的心室位置信息可以是通过手动方式勾勒出来的,或者可以是通过其他方式确定的准确度高的心室位置信息。
步骤102、将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
在本发明实施例中,在将心脏CT图像输入到深度学习网络模型进行训练之前,还可以对各心脏CT图像进行预处理,预处理具体为对各心脏CT图像的尺寸进行归一化处理,这是为了能够更好的满足训练要求及得到精确度更高的分割模型。且在预处理之后,将作为训练数据的多张心脏CT图像输入到深度学习网络模型进行训练,得到该深度学习网络模型输出的各心脏CT图像的第一分割结果。
在本发明实施例中,使用的是深度学习网络模型,通过对该深度学习网络模型进行训练,以得到能够对心室心肌进行准确分割的分割模型。其中,该深度学习网络模型具体可以是V-NET网络模型,或者卷积神经网络类的模型,在实际应用中可以根据具体的需要选择模型,此处不做限定。
步骤103、基于强化学习方法,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果;
在本发明实施例中,强化学习又称再励学习、评价学习,是一种重要的机器学习方法。所谓强化学习就是智能系统从环境到行为映射的学习,以使强化信号函数值最大。将基于强化学习方法,根据心室位置信息及第一分割结果对该第一分割结果进行微调,得到各心脏CT图像的第二分割结果。
步骤104、根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为所述分割模型。
通过迭代的方式,对每次深度学习网络模型输出的第一分割结果进行微调,通过多次迭代及微调,使得训练得到的深度学习网络模型具有较高的分割精度,并可作为分割模型使用,且使用该分割模型对心脏CT图像进行分割,能够得到分割准确度高的心室位置信息。
在本发明实施例中,通过将心脏CT图像及各心脏CT图像的心室位置信息作为训练数据对深度学习网络模型进行训练,使得能够得到可自动对心脏CT图像进行心室分割的分割模型,且进一步的通过结合强化学习方法进行微调及迭代训练,能够得到分割精度更高的分割模型。
为了更好的理解本发明实施例中的技术方案,请参阅图2,为本发明实施例中分割模型训练方法的另一流程图,包括:
步骤201、采集多张心脏计算机断层扫面CT图像,获取各心脏CT图像已勾勒的心室位置信息;
步骤202、将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
在本发明实施例中,上述步骤201及步骤202描述的内容与图1所示实施例中的步骤101及步骤102中的内容相似,具体可以参阅图1所示实施例中的步骤101及步骤102,此处不做赘述。
步骤203、计算所述各心脏CT图像的心室位置信息与所述各心脏CT图像的第一分割结果之间的相似度;
步骤204、根据所述相似度对所述各心脏CT图像的第一分割结果进行微调,得到所述各心脏CT图像的第二分割结果;
在本发明实施例中,是基于强化学习方法,得到各心脏CT图像的第二分割结果的,强化学习方法中涉及到奖惩,可以通过Dice系数判断,Dice系数可以用来判断两个集合的相似程度,即可以计算两个集合的相似度,该两个集合为各心脏CT图像的第一分割结果构成的集合,另一个集合为各心脏CT图像的心室位置信息构成的集合,
[援引加入(细则20.6) 29.04.2019] 
具体的,将使用相似度计算公式计算各心脏CT图像的心室位置信息与各心脏CT图像的第一分割结果之间的相似度,其中,相似度计算公式如下:
Figure WO-DOC-FIGURE-1
Figure 209906dest_path_image001
[援引加入(细则20.6) 29.04.2019] 
其中,D表示各心脏CT图像的第一分割结果构成的集合与各心脏CT图像的心室位置信息构成的集合之间的相似度,N表示心脏CT图像的张数,
Figure WO-DOC-FIGURE-2
表示第i张心脏CT图像的心室位置信息,
Figure WO-DOC-FIGURE-3
表示第i张心脏CT图像的第一分割结果。
其中,将计算得到的相似度与预设值进行比较,该预设值通常为1,且相似度等于1时,表明该第一分割结果是完成正确的,此时表明已经完成了对深度学习网路模型的训练。当相似度不等于上述预设值时,将基于相似度确定是否需要对各心脏CT图像的第一分割结果进行微调。
其中,由于需要进行多次迭代计算,即将微调后的第二分割结果输入至深度学习网络模型进行训练,得到输出的第一分割结果,然后返回步骤203计算相似度,再继续相似度确定是否进行微调。因此,每一次迭代过程都将会计算得到一个相似度,可以保存计算得到的相似度。且在每次计算得到相似度之后,将该相似度作为当前相似度,并将该当前相似度与上一次计算得到的相似度进行比较,当该当前相似度小于上一次计算得到的相似度时,表明此次对深度学习网络模型的训练并没有使得深度学习网络模型的精度提高,确定不对该各心脏CT图像的第一分割结果进行微调,且将加强学习网络中的奖励值(reward)减去一,同时将各心脏CT图像的第一分割结果作为所述各心脏CT图像的第二分割结果,以便后续利用该第二分割结果进行迭代。当该当前相似度大于上一次计算得到的相似度时,表明此次对深度学习网络模型的训练使得深度学习网络模型的精度有了提高,将对各心脏CT图像的第一分割结果进行微调,得到各心脏CT图像的第二分割结果。
其中,对心脏CT图像的第一分割结果的微调可以是上移、下移、左移、右移、缩小、放大和保持不变中的至少一种,至于具体采取何种微调方式,则需要基于心脏CT图像的第一分割结果与心脏CT图像的心室位置信息进行确定。
为了更好的理解微调,请参阅图3,为本发明实施例中第一分割结果和心室位置信息的示意图,在图3中,为同一张心脏CT图像的分割示意图,左侧图像为通过手动方式勾勒的作为标准的心室位置信息,包括上下左右四个坐标点,分别使用Gt_top、Gt_bottom、Gt_left及Gt_right表示,右侧图像为深度学习网络模型输出的第一分割结果,也包括上下左右四个坐标点,分别使用V_top、V_bottom、V_left及V_right表示,按照如下方式进行判别,以确定微调方式:
1、当Gt_top > V_top & Gt_bottom =V_bottom & Gt_left =V_left & Gt_right = V_right时,确定微调方式为上移,即将心脏CT图像的第一分割结果整体上移预设值。
2、当Gt_top < V_top & Gt_bottom < V_bottom & Gt_left =V_left & Gt_right =V_right,确定微调方式为下移,即将心脏CT图像的第一分割结果整体上移预设值。
3、当Gt_top =V-top & Gt_bottom =V_bottom &Gt_left < V_left & Gt_right < V_right ,确定微调方式为左移,即将心脏CT图像的第一分割结果整体上移预设值。
4、当Gt_top =V-top & Gt_bottom = V_bottom &Gt_left > V_left & Gt_right > V_right,确定微调方式为右移,即将心脏CT图像的第一分割结果整体右移预设值。
5、当Gt_top < V-top & Gt_bottom > V_bottom &Gt_left > V_left & Gt_right < V_right,确定微调方式为缩小,即将心脏CT图像的第一分割结果整体缩小预设倍数。
6、当Gt_top > V-top & Gt_bottom < V_bottom &Gt_left < V_left & Gt_right > V_right,确定微调方式为放大,即将心脏CT图像的第一分割结果整体放大预设倍数。
7、当Gt_top = V-top & Gt_bottom = V_bottom &Gt_left =V_left & Gt_right = V_right,确定微调方式为保持不变。
可以理解的是,在上移、下移、左移、或者右移时,移动的预设值可以是1毫米至10毫米范围内的任意一个值,且在缩小或放大时,缩小或放大的预设倍数可以是0.01至0.2倍范围内的任意一个倍数。
在本发明实施例中,可以通过对第一分割结果进行微调的方式得到第二分割结果。
进一步的,将根据各心脏CT图像的第二分割结果及心室位置信息对深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为分割模型,该分割模型能够对心脏CT图像进行精度较高的分割,具体请参阅步骤205及步骤206。
步骤205、计算所述各心脏CT图像的第二分割结果与所述各心脏CT图像的心室位置信息之间的误差;
步骤206、当最大误差大于或等于误差阈值时,将所述第二分割结果输入所述深度学习网络模型,以进行迭代训练,直至计算得到的最大误差小于所述误差阈值。
在本发明实施例中,将依次计算各心脏CT图像的微调后的第二分割结果与各心脏CT图像的心室位置信息之间的误差,当误差中的最大误差大于或等于误差阈值,则将各心脏CT图像的第二分割结果输入到深度学习网络模型中,进行训练,该深度学习网络模型又将输出第一分割结果,返回执行步骤203,以进行迭代,直至迭代次数达到预设数值,或者计算得到的最大误差小于误差阈值。且将迭代结束后的深度学习网络模型作为分割模型,用于对心脏CT图像进行分割。
相对于传统的手工勾勒方法,本发明实施例中引入深度学习网路模型,使得利用深度学习网络模型对心脏CT图像进行分割,能够显著提高分割的精度,且在深度学习的基础上引入了强化学习的方法,利用强化学习的方法对分割结果进行微调及利用微调后的分割结果对深度学习网络模型进行迭代训练,通过多次迭代,可以进一步提高深度学习网络模型的分割精度,作为分割模型使用,能够有效实现对心脏CT图像的心室的分割,且分割精度高,不需要手动勾勒,能够有效的节约人工成本及降低心室分割偏差出现的概率。
请参阅图4,为本发明实施例中分割模型训练装置的结构示意图,该装置包括:
采集获取模块401,用于采集多张心脏计算机断层扫面CT图像,获取各心脏CT图像已勾勒的心室位置信息;
其中,训练之前需要准备进行训练的训练数据,该训练数据包含多张心脏CT图像,及各心脏CT图像的心室位置信息,其中,该多张心脏CT图像可以是通过临床方式采集到的心脏CT图像,且各心脏CT图像的心室位置信息可以是通过手动方式勾勒出来的,或者可以是通过其他方式确定的准确度高的心室位置信息。
训练模块402,用于将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
在本发明实施例中,在将心脏CT图像输入到深度学习网络模型进行训练之前,还可以对各心脏CT图像进行预处理,预处理具体为对各心脏CT图像的尺寸进行归一化处理,这是为了能够更好的满足训练要求及得到精确度更高的分割模型。
其中,深度学习网络模型具体可以是V-NET网络模型,或者卷积神经网络类的模型,在实际应用中可以根据具体的需要选择模型,此处不做限定。
微调模块403,用于基于强化学习装置,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果;
在本发明实施例中,强化学习又称再励学习、评价学习,是一种重要的机器学习方法。所谓强化学习就是智能系统从环境到行为映射的学习,以使强化信号函数值最大。
迭代模块404,用于根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为所述分割模型。
在本发明实施例中,通过将心脏CT图像及各心脏CT图像的心室位置信息作为训练数据对深度学习网络模型进行训练,使得能够得到可自动对心脏CT图像进行心室分割的分割模型,且进一步的通过结合强化学习方法进行微调及迭代训练,能够得到分割精度更高的分割模型。
请参阅图5,为本发明实施例中分割模型训练装置的另一结构示意图,包括如图4所示实施例中的采集获取模块401、训练模块402、微调模块403及迭代模块404,且与图4所示实施例中描述的内容相似,此处不做赘述。
在本发明实施例中,微调模块403包括:
计算模块501,用于计算所述各心脏CT图像的心室位置信息与所述各心脏CT图像的第一分割结果之间的相似度;
结果微调模块502,用于根据所述相似度对所述各心脏CT图像的第一分割结果进行微调,得到所述各心脏CT图像的第二分割结果。
其中,结果微调模块502包括:
比较模块503,用于在所述相似度不等于预设值时,将所述相似度与上一次计算得到的相似度进行比较;
第一处理模块504,用于当所述相似度小于上一次计算得到的相似度时,确定不对所述各心脏CT图像的第一分割结果进行微调,且将各心脏CT图像的第一分割结果作为所述各心脏CT图像的第二分割结果;
第二处理模块505,用于当所述相似度大于上一次计算得到的相似度时,对所述各心脏CT图像的第一分割结果进行微调,得到所述第二分割结果。
所述第二处理模块505具体用于将所述各心脏的CT图像的心室位置信息中的各方位的坐标值与相应的所述第一分割结果的各方位的坐标值进行比较,得到比较结果;按照所述比较结果对相应的所述第一分割结果的各方位的坐标值进行调整,得到所述第二分割结果,调整包括:上移、下移、左移、右移、缩小、放大、保持不变中的至少一种。
在本发明实施例中,迭代模块404包括:
误差计算模块506,用于计算所述各心脏CT图像的第二分割结果与所述各心脏CT图像的心室位置信息之间的误差;
迭代训练模块507,用于当最大误差大于或等于误差阈值时,将所述第二分割结果输入所述深度学习网络模型,以进行迭代训练,直至计算得到的最大误差小于所述误差阈值。
在本发明实施例中,是基于强化学习方法,得到各心脏CT图像的第二分割结果的,强化学习方法中涉及到奖惩,可以通过Dice系数判断,Dice系数可以用来判断两个集合的相似程度,即可以计算两个集合的相似度,该两个集合为各心脏CT图像的第一分割结果构成的集合,另一个集合为各心脏CT图像的心室位置信息构成的集合,
[援引加入(细则20.6) 29.04.2019] 
具体的,将使用相似度计算公式计算各心脏CT图像的心室位置信息与各心脏CT图像的第一分割结果之间的相似度,其中,相似度计算公式如下:
Figure WO-DOC-FIGURE-1
Figure 881736dest_path_image001
[援引加入(细则20.6) 29.04.2019] 
其中,D表示各心脏CT图像的第一分割结果构成的集合与各心脏CT图像的心室位置信息构成的集合之间的相似度,N表示心脏CT图像的张数,
Figure WO-DOC-FIGURE-2
表示第i张心脏CT图像的心室位置信息,
Figure WO-DOC-FIGURE-3
表示第i张心脏CT图像的第一分割结果。
其中,将计算得到的相似度与预设值进行比较,该预设值通常为1,且相似度等于1时,表明该第一分割结果是完成正确的,此时表明已经完成了对深度学习网路模型的训练。当相似度不等于上述预设值时,将基于相似度确定是否需要对各心脏CT图像的第一分割结果进行微调。
其中,由于需要进行多次迭代计算,即将微调后的第二分割结果输入至深度学习网络模型进行训练,得到输出的第一分割结果,然后返回步骤203计算相似度,再继续相似度确定是否进行微调。因此,每一次迭代过程都将会计算得到一个相似度,可以保存计算得到的相似度。且在每次计算得到相似度之后,将该相似度作为当前相似度,并将该当前相似度与上一次计算得到的相似度进行比较,当该当前相似度小于上一次计算得到的相似度时,表明此次对深度学习网络模型的训练并没有使得深度学习网络模型的精度提高,确定不对该各心脏CT图像的第一分割结果进行微调,且将加强学习网络中的奖励值减去一,且将各心脏CT图像的第一分割结果作为各心脏CT图像的第二分割结果,以便后续利用第二分割结果进行迭代。当该当前相似度大于上一次计算得到的相似度时,表明此次对深度学习网络模型的训练使得深度学习网络模型的精度有了提高,将对各心脏CT图像的第一分割结果进行微调,得到各心脏CT图像的第二分割结果。
其中,对心脏CT图像的第一分割结果的微调可以是上移、下移、左移、右移、缩小、放大和保持不变中的至少一种,至于具体采取何种微调方式,则需要基于心脏CT图像的第一分割结果与心脏CT图像的心室位置信息进行确定。
为了更好的理解微调,请参阅图3,为本发明实施例中第一分割结果和心室位置信息的示意图,在图3中,为同一张心脏CT图像的分割示意图,左侧图像为通过手动方式勾勒的作为标准的心室位置信息,包括上下左右四个坐标点,分别使用Gt_top、Gt_bottom、Gt_left及Gt_right表示,右侧图像为深度学习网络模型输出的第一分割结果,也包括上下左右四个坐标点,分别使用V_top、V_bottom、V_left及V_right表示,按照如下方式进行判别,以确定微调方式:
1、当Gt_top > V_top & Gt_bottom =V_bottom & Gt_left =V_left & Gt_right = V_right时,确定微调方式为上移,即将心脏CT图像的第一分割结果整体上移预设值。
2、当Gt_top < V_top & Gt_bottom < V_bottom & Gt_left =V_left & Gt_right =V_right,确定微调方式为下移,即将心脏CT图像的第一分割结果整体上移预设值。
3、当Gt_top =V-top & Gt_bottom =V_bottom &Gt_left < V_left & Gt_right < V_right ,确定微调方式为左移,即将心脏CT图像的第一分割结果整体上移预设值。
4、当Gt_top =V-top & Gt_bottom = V_bottom &Gt_left > V_left & Gt_right > V_right,确定微调方式为右移,即将心脏CT图像的第一分割结果整体右移预设值。
5、当Gt_top < V-top & Gt_bottom > V_bottom &Gt_left > V_left & Gt_right < V_right,确定微调方式为缩小,即将心脏CT图像的第一分割结果整体缩小预设倍数。
6、当Gt_top > V-top & Gt_bottom < V_bottom &Gt_left < V_left & Gt_right > V_right,确定微调方式为放大,即将心脏CT图像的第一分割结果整体放大预设倍数。
7、当Gt_top = V-top & Gt_bottom = V_bottom &Gt_left =V_left & Gt_right = V_right,确定微调方式为保持不变。
在本发明实施例中,可以通过对第一分割结果进行微调的方式得到第二分割结果。
相对于传统的手工勾勒方法,本发明实施例中引入深度学习网路模型,使得利用深度学习网络模型对心脏CT图像进行分割,能够显著提高分割的精度,且在深度学习的基础上引入了强化学习的方法,利用强化学习的方法对分割结果进行微调及利用微调后的分割结果对深度学习网络模型进行迭代训练,通过多次迭代,可以进一步提高深度学习网络模型的分割精度,作为分割模型使用,能够有效实现对心脏CT图像的心室的分割,且分割精度高,不需要手动勾勒,能够有效的节约人工成本及降低心室分割偏差出现的概率。
在本发明实施例中还提供一种计算机可读存储介质,其上存储有计算机程序,该计算机程序被处理器执行时,实现如图1或图2所示实施例中的分割模型训练方法的各个步骤。
在本申请所提供的几个实施例中,应该理解到,所揭露的装置和方法,可以通过其它的方式实现。例如,以上所描述的装置实施例仅仅是示意性的,例如,所述模块的划分,仅仅为一种逻辑功能划分,实际实现时可以有另外的划分方式,例如多个模块或组件可以结合或者可以集成到另一个系统,或一些特征可以忽略,或不执行。另一点,所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口,装置或模块的间接耦合或通信连接,可以是电性,机械或其它的形式。
所述作为分离部件说明的模块可以是或者也可以不是物理上分开的,作为模块显示的部件可以是或者也可以不是物理模块,即可以位于一个地方,或者也可以分布到多个网络模块上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。
另外,在本发明各个实施例中的各功能模块可以集成在一个处理模块中,也可以是各个模块单独物理存在,也可以两个或两个以上模块集成在一个模块中。上述集成的模块既可以采用硬件的形式实现,也可以采用软件功能模块的形式实现。
所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。基于这样的理解,本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来,该计算机软件产品存储在一个存储介质中,包括若干指令用以使得一台计算机设备(可以是个人计算机,服务器,或者网络设备等)执行本发明各个实施例所述方法的全部或部分步骤。而前述的存储介质包括:U盘、移动硬盘、只读存储器(ROM,Read-Only Memory)、随机存取存储器(RAM,Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
需要说明的是,对于前述的各方法实施例,为了简便描述,故将其都表述为一系列的动作组合,但是本领域技术人员应该知悉,本发明并不受所描述的动作顺序的限制,因为依据本发明,某些步骤可以采用其它顺序或者同时进行。其次,本领域技术人员也应该知悉,说明书中所描述的实施例均属于优选实施例,所涉及的动作和模块并不一定都是本发明所必须的。
在上述实施例中,对各个实施例的描述都各有侧重,某个实施例中没有详述的部分,可以参见其它实施例的相关描述。
以上为对本发明所提供的一种分割模型训练方法、装置及计算机可读存储介质的描述,对于本领域的技术人员,依据本发明实施例的思想,在具体实施方式及应用范围上均会有改变之处,综上,本说明书内容不应理解为对本发明的限制。

Claims (10)

  1. 一种分割模型训练方法,其特征在于,所述方法包括:
    采集多张心脏计算机断层扫面CT图像,获取各心脏CT图像已勾勒的心室位置信息;
    将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
    基于强化学习方法,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果;
    根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为所述分割模型。
  2. 根据权利1所述的方法,其特征在于,所述基于强化学习方法,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果,包括:
    计算所述各心脏CT图像的心室位置信息与所述各心脏CT图像的第一分割结果之间的相似度;
    根据所述相似度对所述各心脏CT图像的第一分割结果进行微调,得到所述各心脏CT图像的第二分割结果。
  3. 根据权利要求2所述的方法,其特征在于,所述根据所述相似度对所述各心脏CT图像的分割结果进行微调,得到所述各心脏CT图像的第二分割结果,包括:
    在所述相似度不等于预设值时,将所述相似度与上一次计算得到的相似度进行比较;
    当所述相似度小于上一次计算得到的相似度时,确定不对所述各心脏CT图像的第一分割结果进行微调,并将所述各心脏CT图像的第一分割结果作为所述各心脏CT图像的第二分割结果;
    当所述相似度大于上一次计算得到的相似度时,对所述各心脏CT图像的第一分割结果进行微调,得到所述第二分割结果。
  4. 根据权利要求3所述的方法,其特征在于,所述对所述各心脏CT图像的第一分割结果进行微调,得到所述第二分割结果,包括:
    将所述各心脏的CT图像的心室位置信息中的各方位的坐标值与相应的所述第一分割结果的各方位的坐标值进行比较,得到比较结果;
    按照所述比较结果对相应的所述第一分割结果的各方位的坐标值进行调整,得到所述第二分割结果,调整包括:上移、下移、左移、右移、缩小、放大、保持不变中的至少一种。
  5. 根据权利要求1至4任意一项所述的方法,其特征在于,所述根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,包括:
    计算所述各心脏CT图像的第二分割结果与所述各心脏CT图像的心室位置信息之间的误差;
    当最大误差大于或等于误差阈值时,将所述第二分割结果输入所述深度学习网络模型,以进行迭代训练,直至计算得到的最大误差小于所述误差阈值。
  6. 一种分割模型训练装置,其特征在于,所述装置包括:
    采集获取模块,用于采集多张心脏计算机断层扫面CT图像,获取各心脏CT图像已勾勒的心室位置信息;
    训练模块,用于将所述心脏CT图像输入到深度学习网络模型进行训练,得到所述深度学习网络模型输出的各所述心脏CT图像的第一分割结果;
    微调模块,用于基于强化学习装置,根据所述心室位置信息及所述第一分割结果对所述第一分割结果进行微调,得到各所述心脏CT图像的第二分割结果;
    迭代模块,用于根据所述第二分割结果及所述心室位置信息对所述深度学习网络模型进行迭代训练,将训练后的深度学习网络模型作为所述分割模型。
  7. 根据权利6所述的装置,其特征在于,所述微调模块包括:
    计算模块,用于计算所述各心脏CT图像的心室位置信息与所述各心脏CT图像的第一分割结果之间的相似度;
    结果微调模块,用于根据所述相似度对所述各心脏CT图像的第一分割结果进行微调,得到所述各心脏CT图像的第二分割结果。
  8. 根据权利要求6所述的装置,其特征在于,所述结果微调模块包括:
    比较模块,用于在所述相似度不等于预设值时,将所述相似度与上一次计算得到的相似度进行比较;
    第一处理模块,用于当所述相似度小于上一次计算得到的相似度时,确定不对所述各心脏CT图像的第一分割结果进行微调,且将各心脏CT图像的第一分割结果作为所述各心脏CT图像的第二分割结果;
    第二处理模块,用于当所述相似度大于上一次计算得到的相似度时,对所述各心脏CT图像的第一分割结果进行微调,得到所述第二分割结果。
    所述第二处理模块具体用于将所述各心脏的CT图像的心室位置信息中的各方位的坐标值与相应的所述第一分割结果的各方位的坐标值进行比较,得到比较结果;按照所述比较结果对相应的所述第一分割结果的各方位的坐标值进行调整,得到所述第二分割结果,调整包括:上移、下移、左移、右移、缩小、放大、保持不变中的至少一种。
  9. 根据权利要求5至8任意一项所述的装置,其特征在于,所述迭代模块包括:
    误差计算模块,用于计算所述各心脏CT图像的第二分割结果与所述各心脏CT图像的心室位置信息之间的误差;
    迭代训练模块,用于当最大误差大于或等于误差阈值时,将所述第二分割结果输入所述深度学习网络模型,以进行迭代训练,直至计算得到的最大误差小于所述误差阈值。
  10. 一种计算机可读存储介质,其上存储有计算机程序,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任意一项所述的分割模型训练方法的各个步骤。
PCT/CN2019/071501 2018-07-03 2019-01-12 分割模型训练方法、装置及计算机可读存储介质 WO2020007026A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810716876.4A CN109102509B (zh) 2018-07-03 2018-07-03 分割模型训练方法、装置及计算机可读存储介质
CN201810716876.4 2018-07-03

Publications (1)

Publication Number Publication Date
WO2020007026A1 true WO2020007026A1 (zh) 2020-01-09

Family

ID=64845516

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/071501 WO2020007026A1 (zh) 2018-07-03 2019-01-12 分割模型训练方法、装置及计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN109102509B (zh)
WO (1) WO2020007026A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115272374A (zh) * 2022-08-01 2022-11-01 北京安德医智科技有限公司 心脏磁共振影像的心室分割方法和装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109102509B (zh) * 2018-07-03 2021-06-25 深圳先进技术研究院 分割模型训练方法、装置及计算机可读存储介质
CN112365504A (zh) * 2019-10-29 2021-02-12 杭州脉流科技有限公司 Ct左心室分割方法、装置、设备和存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107578416A (zh) * 2017-09-11 2018-01-12 武汉大学 一种由粗到精级联深度网络的心脏左心室全自动分割方法
CN108109151A (zh) * 2017-12-19 2018-06-01 哈尔滨工业大学 一种基于深度学习和形变模型的超声心动图心室分割方法和装置
US20180174001A1 (en) * 2016-12-15 2018-06-21 Samsung Electronics Co., Ltd. Method of training neural network, and recognition method and apparatus using neural network
CN109102509A (zh) * 2018-07-03 2018-12-28 深圳先进技术研究院 分割模型训练方法、装置及计算机可读存储介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101226155A (zh) * 2007-12-21 2008-07-23 中国人民解放军第八一医院 智能化肺癌早期细胞病理图像识别处理方法
US8391603B2 (en) * 2009-06-18 2013-03-05 Omisa Inc. System and method for image segmentation
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
CN107358600A (zh) * 2017-06-14 2017-11-17 北京全域医疗技术有限公司 放疗计划中的自动勾靶方法、装置和电子设备
CN107274408B (zh) * 2017-06-16 2019-11-19 厦门大学 一种基于新型粒子滤波算法的图像分割方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180174001A1 (en) * 2016-12-15 2018-06-21 Samsung Electronics Co., Ltd. Method of training neural network, and recognition method and apparatus using neural network
CN107578416A (zh) * 2017-09-11 2018-01-12 武汉大学 一种由粗到精级联深度网络的心脏左心室全自动分割方法
CN108109151A (zh) * 2017-12-19 2018-06-01 哈尔滨工业大学 一种基于深度学习和形变模型的超声心动图心室分割方法和装置
CN109102509A (zh) * 2018-07-03 2018-12-28 深圳先进技术研究院 分割模型训练方法、装置及计算机可读存储介质

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115272374A (zh) * 2022-08-01 2022-11-01 北京安德医智科技有限公司 心脏磁共振影像的心室分割方法和装置

Also Published As

Publication number Publication date
CN109102509A (zh) 2018-12-28
CN109102509B (zh) 2021-06-25

Similar Documents

Publication Publication Date Title
CN110232383B (zh) 一种基于深度学习模型的病灶图像识别方法及病灶图像识别系统
WO2020119679A1 (zh) 三维左心房分割方法、装置、终端设备及存储介质
CN107578416B (zh) 一种由粗到精级联深度网络的心脏左心室全自动分割方法
CN109035284B (zh) 基于深度学习的心脏ct图像分割方法、装置、设备及介质
WO2017193251A1 (zh) 识别超声图像中感兴趣区域轮廓的方法及系统
Liu et al. Automatic whole heart segmentation using a two-stage u-net framework and an adaptive threshold window
Nurmaini et al. Accurate detection of septal defects with fetal ultrasonography images using deep learning-based multiclass instance segmentation
CN110164550B (zh) 一种基于多视角协同关系的先天性心脏病辅助诊断方法
CN111127527B (zh) 基于ct影像骨骼配准实现肺结节自适应匹配的方法与装置
WO2020007026A1 (zh) 分割模型训练方法、装置及计算机可读存储介质
CN109255354B (zh) 面向医用ct的计算机图像处理方法和装置
WO2022213654A1 (zh) 一种超声图像的分割方法、装置、终端设备和存储介质
CN109215040B (zh) 一种基于多尺度加权学习的乳腺肿瘤分割方法
US20220335600A1 (en) Method, device, and storage medium for lesion segmentation and recist diameter prediction via click-driven attention and dual-path connection
CN110619635B (zh) 基于深度学习的肝细胞癌磁共振图像分割系统和方法
CN110751187B (zh) 异常区域图像生成网络的训练方法和相关产品
CN111080592B (zh) 一种基于深度学习的肋骨提取方法及装置
CN111340825A (zh) 一种纵膈淋巴结分割模型的生成方法及系统
CN111754534A (zh) 基于深度神经网络的ct左心室短轴图像分割方法、装置、计算机设备和存储介质
CN111815597A (zh) 基于ct影像的左心室长短轴切面提取方法、装置、计算机设备和存储介质
CN116130090A (zh) 射血分数测量方法和装置、电子设备和存储介质
EP2498222B1 (en) Method and system for regression-based 4D mitral valve segmentation from 2D+T magnetic resonance imaging slices
CN110739050B (zh) 一种左心室全参数及置信度的量化方法
CN112164447B (zh) 图像处理方法、装置、设备及存储介质
CN113592802B (zh) 一种基于超声图像的二尖瓣环位移自动检测系统

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: 19831177

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: 19831177

Country of ref document: EP

Kind code of ref document: A1