CN111340825A - Method and system for generating mediastinal lymph node segmentation model - Google Patents

Method and system for generating mediastinal lymph node segmentation model Download PDF

Info

Publication number
CN111340825A
CN111340825A CN202010148259.6A CN202010148259A CN111340825A CN 111340825 A CN111340825 A CN 111340825A CN 202010148259 A CN202010148259 A CN 202010148259A CN 111340825 A CN111340825 A CN 111340825A
Authority
CN
China
Prior art keywords
image
segmentation
lymph node
mediastinal lymph
dimensional
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
CN202010148259.6A
Other languages
Chinese (zh)
Other versions
CN111340825B (en
Inventor
刘馨月
陈昶
谢冬
佘云浪
邓家骏
王亭亭
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Pulmonary Hospital (shanghai Occupational Disease Prevention And Treatment Institute)
Original Assignee
Shanghai Pulmonary Hospital (shanghai Occupational Disease Prevention And Treatment Institute)
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 Shanghai Pulmonary Hospital (shanghai Occupational Disease Prevention And Treatment Institute) filed Critical Shanghai Pulmonary Hospital (shanghai Occupational Disease Prevention And Treatment Institute)
Priority to CN202010148259.6A priority Critical patent/CN111340825B/en
Publication of CN111340825A publication Critical patent/CN111340825A/en
Application granted granted Critical
Publication of CN111340825B publication Critical patent/CN111340825B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30061—Lung
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30096—Tumor; Lesion

Abstract

The invention provides a generation method and a system of a mediastinal lymph node segmentation model, which relate to the technical field of medical image processing and comprise the following steps: acquiring lung CT images of a plurality of thoracic surgery patients, and performing three-dimensional reconstruction on each lung CT image to obtain a three-dimensional image; respectively carrying out three-dimensional segmentation on each three-dimensional image to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area; grouping the three-dimensional labeling images to obtain a training set, a testing set and a correcting set; training a training set to obtain a mediastinal lymph node segmentation model; inputting the test set into a mediastinal lymph node segmentation model to obtain a corresponding segmentation image, and calculating the segmentation accuracy of the mediastinal lymph node segmentation model; if the segmentation accuracy is smaller than the accuracy threshold, the correction set corrects the mediastinal lymph node segmentation model; and if the segmentation accuracy is not less than the accuracy threshold, storing the mediastinal lymph node segmentation model. The method effectively improves the accuracy of the mediastinal lymph node segmentation, does not need manual intervention, and has strong practicability.

Description

Method and system for generating mediastinal lymph node segmentation model
Technical Field
The invention relates to the technical field of medical image processing, in particular to a generation method and a generation system of a mediastinal lymph node segmentation model.
Background
The incidence rate and the death rate of the lung cancer are high, the 5-year survival rate is low, and the lung cancer is the leading factor of cancer death worldwide. Spread and metastasis occur in advanced lung cancer, with mediastinal lymph node metastasis being a more common condition. Lung cancer does not have obvious symptoms in the early stage of lymphatic metastasis, and lymphadenectasis occurs in the late stage. As the disease condition increases, multiple lymph nodes may swell. After the occurrence of lymph node metastasis in lung cancer, most of the lung cancer is not treated well, because the spread and metastasis of cancer cells are severe at this time and spread all over the body through the lymphatic system, and a plurality of new cancer focuses are formed. Therefore, the accurate division of the lymph nodes is very important for removing the focus by the operation.
At present, medical Imaging becomes one of the most important means for lung cancer examination, treatment scheme selection, curative effect detection and the like, and according to medical Imaging examination methods such as X-ray Imaging, ultrasonic Imaging, Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET) and the like, internal focuses of a human body can be non-invasively 'snooped', tumors are examined, and disease changes of the tumors are monitored. However, in conventional lymph node segmentation, a specialist must manually segment the lymph node by observing the image. The method needs a professional doctor to perform complicated manual operation on a large amount of data, the doctor may have some errors of identification and edge segmentation due to long-term and large amount of repeated work, meanwhile, the accuracy and reliability of segmentation results of the method depend on the experience knowledge and professional quality of the doctor seriously, and the accuracy of the results is limited, so that the research on an automatic lymph node segmentation system is necessary.
In recent years, machine learning and deep learning have been developed, and excellent performance has been exerted in various fields. In the field of medical information analysis, machine learning and deep learning algorithms are widely applied in the fields of breast lump classification, breast molybdenum target tumor detection and the like. Successful segmentation of cell images by a U-network (U-Net) again demonstrates that deep learning can be well used for semantic segmentation of medical images. At present, the application of deep learning in clinical lung lesions is mostly limited to pulmonary nodules, however, the detection and segmentation of lung mediastinal lymph nodes have great significance for the formulation of doctor operation schemes and lymph node cleaning, but at present, no method for segmenting the lung cancer mediastinal lymph nodes from CT images of lung cancer lesions by utilizing deep learning exists.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a generation method of a mediastinal lymph node segmentation model, which specifically comprises the following steps:
step S1, acquiring lung CT images of a plurality of thoracic surgery patients, and performing three-dimensional reconstruction on the lung CT images to obtain a three-dimensional image corresponding to each thoracic surgery patient;
step S2, performing three-dimensional segmentation on each three-dimensional image to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area;
step S3, grouping the three-dimensional annotation images to obtain a training set, a test set and a correction set;
step S4, training according to each three-dimensional annotation image in the training set to obtain a mediastinal lymph node segmentation model;
step S5, inputting each three-dimensional labeling image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image, and calculating to obtain the segmentation accuracy of the mediastinal lymph node segmentation model according to each segmentation image and the corresponding real mediastinal lymph node focus area;
step S6, comparing the segmentation accuracy with a preset accuracy threshold:
if the segmentation accuracy is less than the accuracy threshold, then go to step S7;
if the segmentation accuracy is not less than the accuracy threshold, turning to step S8;
step S7, correcting the mediastinal lymph node segmentation model according to each three-dimensional annotation image in the correction set, and then returning to the step S6;
and step S8, storing the mediastinal lymph node segmentation model so as to segment the mediastinal lymph node.
Preferably, the step S4 specifically includes:
step S41, establishing a coordinate system on each three-dimensional annotation image in the training set, and respectively performing overlapping sampling on each three-dimensional annotation image according to the coordinate system to obtain a plurality of image blocks;
the image blocks obtained by sampling each three-dimensional labeling image have the same number and the same preset size, and each image block has a central coordinate related to the coordinate system;
step S42, inputting each image block into a pre-generated depth residual error U-net segmentation model for feature learning aiming at each image block obtained by sampling each three-dimensional labeling image to obtain a sub-segmentation probability map corresponding to each image block;
the sub-segmentation probability map comprises probability values of all voxel points of the corresponding image blocks as mediastinal lymph nodes;
step S43, restoring each sub-segmentation probability map in the coordinate system where the corresponding image block is located according to the central coordinate, and averaging the probability values of coincident pixel points between each sub-segmentation probability map to obtain a total segmentation probability map of the three-dimensional image;
step S44, comparing the probability value of each voxel point in the total segmentation probability map with a preset class probability threshold:
if the probability value is greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a first numerical value representing that the voxel point is a focal zone, and then turning to step S45;
if the probability value is not greater than the class probability threshold, setting the voxel value of the corresponding voxel point to a second numerical value representing that the voxel point is not a focal zone, and then turning to step S45;
step S45, generating a mediastinal lymph node segmentation result of the three-dimensional labeling image according to the first numerical value and the second numerical value, calculating an error between the mediastinal lymph node segmentation result and the real mediastinal lymph node lesion area, and adjusting parameters of the depth residual U-net segmentation model according to the error;
and S46, repeatedly executing the steps S42 to S45 until the training is finished, and obtaining the mediastinal lymph node segmentation model.
Preferably, in the step S41, the preset size is 24 pixels by 8 pixels.
Preferably, in step S44, the first value is 1.
Preferably, in step S44, the second value is 0.
Preferably, the step S5 specifically includes:
step S51, inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain corresponding segmentation images;
step S52, respectively calculating a coincidence rate between each segmented image and the corresponding real mediastinal lymph node lesion area according to each segmented image and the corresponding real mediastinal lymph node lesion area, and comparing the coincidence rate with a preset coincidence rate threshold:
if the coincidence rate is not less than the coincidence rate threshold, adding the corresponding three-dimensional annotation image into a first image library, and then turning to step S53;
if the coincidence rate is smaller than the coincidence rate threshold value, adding the corresponding three-dimensional annotation image into a second image library, and then turning to the step S53;
step S53, respectively counting a first number of the three-dimensional labeling images in the first image library and a second number of the three-dimensional labeling images in the second image library, and calculating a segmentation accuracy of the mediastinal lymph node segmentation model according to the first number and the second number.
Preferably, in step S53, the segmentation accuracy is calculated according to the following formula:
Ar＝N1/(N1+N2)
wherein the content of the first and second substances,
Arrepresenting the segmentation accuracy;
N1representing the first quantity;
N2representing said second number.
A generation system of a mediastinal lymph node segmentation model is applied to any one of the generation methods of the mediastinal lymph node segmentation model, and the generation system specifically comprises:
the data acquisition module is used for acquiring lung CT images of a plurality of thoracic surgery patients and performing three-dimensional reconstruction on the lung CT images to obtain a three-dimensional image corresponding to each thoracic surgery patient;
the data preprocessing module is connected with the data acquisition module and is used for respectively carrying out three-dimensional segmentation on each three-dimensional image to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area;
the data grouping module is connected with the data preprocessing module and used for grouping the three-dimensional annotation images to obtain a training set, a testing set and a correcting set;
the model training module is connected with the data grouping module and used for training according to each three-dimensional labeling image in the training set to obtain a mediastinal lymph node segmentation model;
the model evaluation module is respectively connected with the data grouping module and the model training module and is used for respectively inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image and calculating the segmentation accuracy of the mediastinal lymph node segmentation model according to each segmentation image and the corresponding real mediastinal lymph node focus area;
the data comparison module is connected with the model evaluation module and used for comparing the segmentation accuracy with a preset accuracy threshold, generating a first comparison result when the segmentation accuracy is not less than the accuracy threshold, and generating a second comparison result when the segmentation accuracy is less than the accuracy threshold;
the model correction module is respectively connected with the data grouping module and the data comparison module and is used for correcting the mediastinal lymph node segmentation model according to the first comparison result and each three-dimensional annotation image in the correction set;
and the model storage module is connected with the data comparison module and used for storing the mediastinal lymph node segmentation model according to the second comparison result so as to segment the mediastinal lymph node.
Preferably, the model training module specifically includes:
the image sampling unit is used for establishing a coordinate system on each three-dimensional annotation image in the training set and respectively sampling each three-dimensional annotation image according to the coordinate system to obtain a plurality of image blocks;
the image blocks obtained by sampling each three-dimensional labeling image have the same number and the same preset size, and each image block has a central coordinate related to the coordinate system;
the characteristic learning unit is connected with the image sampling unit and used for inputting each image block into a pre-generated depth residual error U-net segmentation model for characteristic learning aiming at each image block obtained by sampling each three-dimensional labeling image to obtain a sub-segmentation probability map corresponding to each image block;
the sub-segmentation probability map comprises probability values of all voxel points of the corresponding image blocks as mediastinal lymph nodes;
the image restoration unit is connected with the feature learning unit and used for restoring each sub-segmentation probability map into the coordinate system where each corresponding image block is located according to the central coordinate, and averaging the probability values of coincident pixel points among the sub-segmentation probability maps to obtain a total segmentation probability map of the three-dimensional image;
a probability comparison unit connected with the image restoration unit and used for comparing the probability value of each voxel point in the total segmentation probability map with a preset class probability threshold value and setting the corresponding voxel value of the voxel point as a first numerical value representing that the voxel point is a focus area when the probability value is greater than the class probability threshold value, and
if the probability value is not greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a second numerical value representing that the voxel point is not a focal zone;
the data adjusting unit is connected with the data comparing unit and used for generating a mediastinal lymph node segmentation result of the three-dimensional image according to the first numerical value and the second numerical value, calculating an error between the mediastinal lymph node segmentation result and the real mediastinal lymph node lesion area, and adjusting parameters of the depth residual U-net segmentation model according to the error;
and the model generation unit is connected with the data adjustment unit and is used for obtaining the mediastinal lymph node segmentation model after the training is finished.
Preferably, the model evaluation module specifically includes:
the image segmentation unit is used for respectively inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image;
an image processing unit connected with the image segmentation unit and used for respectively calculating the coincidence rate between each segmented image and the corresponding real mediastinal lymph node focus region according to each segmented image and the corresponding real mediastinal lymph node focus region, and adding the corresponding three-dimensional annotation image into a first image library when the coincidence rate is not less than a preset coincidence rate threshold value, and
adding the corresponding three-dimensional labeling image into a second image library when the coincidence rate is smaller than the coincidence rate threshold value;
and the data processing unit is connected with the image processing unit and is used for respectively counting to obtain a first number of the three-dimensional labeling images in the first image library and a second number of the three-dimensional labeling images in the second image library, and calculating to obtain the segmentation accuracy of the mediastinal lymph node segmentation model according to the first number and the second number.
The technical scheme has the following advantages or beneficial effects:
1) the training centralizes the overlapping sampling of the three-dimensional images, the accuracy and the specificity of the mediastinal lymph node segmentation model are effectively improved, manual intervention is not needed, and the practicability is high;
2) when the test result of the test set on the mediastinal lymph node segmentation model is not ideal, the correction set is adopted to correct the mediastinal lymph node segmentation model, so that the segmentation accuracy of the mediastinal lymph node segmentation model is further improved;
3) the mediastinal lymph node segmentation model can realize automatic segmentation of mediastinal lymph nodes, help doctors to make further diagnosis, timely diagnose illness, determine an operation scheme, further reduce the working strength of doctors and avoid delaying the best treatment opportunity of the illness.
Drawings
FIG. 1 is a flow chart illustrating a method for generating a mediastinal lymph node segmentation model according to a preferred embodiment of the present invention;
FIG. 2 is a flow chart illustrating a training process of a mediastinal lymph node segmentation model according to a preferred embodiment of the present invention;
FIG. 3 is a flow chart of a method for calculating segmentation accuracy according to a preferred embodiment of the present invention;
FIG. 4 is a schematic structural diagram of a generation system of a mediastinal lymph node segmentation model according to a preferred embodiment of the invention.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. The present invention is not limited to the embodiment, and other embodiments may be included in the scope of the present invention as long as the gist of the present invention is satisfied.
In a preferred embodiment of the present invention, based on the above problems in the prior art, a method for generating a mediastinal lymph node segmentation model is provided, as shown in fig. 1, which specifically includes the following steps:
step S1, acquiring lung CT images of a plurality of thoracic surgery patients, and performing three-dimensional reconstruction on each lung CT image to obtain a three-dimensional image corresponding to each thoracic surgery patient;
step S2, performing three-dimensional segmentation on each three-dimensional image respectively to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area;
step S3, grouping the three-dimensional annotation images to obtain a training set, a test set and a correction set;
step S4, training according to each three-dimensional annotation image in the training set to obtain a mediastinal lymph node segmentation model;
step S5, inputting each three-dimensional labeling image in the test set into a mediastinal lymph node segmentation model to obtain a corresponding segmentation image, and calculating to obtain the segmentation accuracy of the mediastinal lymph node segmentation model according to each segmentation image and a corresponding real mediastinal lymph node focus area;
step S6, comparing the segmentation accuracy with a preset accuracy threshold:
if the segmentation accuracy is smaller than the accuracy threshold, turning to step S7;
if the segmentation accuracy is not less than the accuracy threshold, turning to step S8;
step S7, correcting the mediastinal lymph node segmentation model according to each three-dimensional annotation image in the correction set, and then returning to the step S6;
step S8, storing the mediastinal lymph node segmentation model to segment the mediastinal lymph node.
Specifically, in this embodiment, 3700 sets of lung CT images including all phases of imaging data set images of a thoracic surgical patient are preferably acquired, each set of lung CT images includes a plurality of scanning slice images, the scanning slice images are two-dimensional images, and a corresponding three-dimensional image is obtained by performing three-dimensional reconstruction on the scanning slice images in each set of lung CT images. And then, performing three-dimensional segmentation on the mediastinal lymph nodes by an imaging expert according to the three-dimensional images to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area. Preferably, 3700 three-dimensional annotation images are randomly divided into a training set, a testing set and a correction set, wherein the training set comprises 2000 three-dimensional annotation images which are used for training to obtain a mediastinal lymph node segmentation model; the test set comprises 1200 three-dimensional annotation images and is used for testing and evaluating the mediastinal lymph node segmentation model obtained through training, and the segmentation accuracy of the mediastinal lymph node segmentation model is used as an evaluation standard; the correction set comprises 500 three-dimensional labeling images and is used for correcting the mediastinal lymph node segmentation model when the segmentation accuracy rate does not reach the preset standard, and the finally obtained mediastinal lymph node segmentation model is used for segmenting mediastinal lymph nodes.
In a preferred embodiment of the present invention, as shown in fig. 2, step S4 specifically includes:
step S41, establishing a coordinate system on each three-dimensional annotation image in the training set, and respectively performing overlapping sampling on each three-dimensional annotation image according to the coordinate system to obtain a plurality of image blocks;
the image blocks sampled by each three-dimensional labeling image have the same number and the same preset size, and each image block has a central coordinate related to a coordinate system;
step S42, inputting each image block into a pre-generated depth residual error U-net segmentation model for feature learning aiming at each image block obtained by sampling each three-dimensional labeling image, and obtaining a sub-segmentation probability map corresponding to each image block;
each voxel point of the corresponding image block in the sub-segmentation probability map is the probability value of the mediastinal lymph node;
step S43, restoring each sub-segmentation probability map in a coordinate system where each corresponding image block is located according to the central coordinate, and averaging the probability values of coincident pixel points among the sub-segmentation probability maps to obtain a total segmentation probability map of the three-dimensional image;
step S44, comparing the probability value of each voxel point in the total segmentation probability map with a preset class probability threshold:
if the probability value is greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a first numerical value representing that the voxel point is a focal zone, and then turning to step S45;
if the probability value is not greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a second numerical value representing that the voxel point is not the focal zone, and then turning to step S45;
step S45, generating a mediastinal lymph node segmentation result of the three-dimensional labeling image according to the first numerical value and the second numerical value, calculating an error between the mediastinal lymph node segmentation result and a real mediastinal lymph node focus area, and adjusting parameters of a depth residual error U-net segmentation model according to the error;
and step S46, repeating the step S42 to the step S45 until the training is finished, and obtaining the mediastinal lymph node segmentation model.
Specifically, in this embodiment, the mediastinal lymph node segmentation model includes three processing procedures, wherein the first processing procedure is a process of sampling the three-dimensional labeled image:
in this embodiment, preferably, the three-dimensional labeling image is sampled into small blocks as input data of the next processing procedure in a sampling manner with the focal region of the real mediastinal lymph node as the center. Because the sizes of the real mediastinal lymph node lesion areas in each three-dimensional labeling image are different, in order to enable image blocks obtained by sampling of each three-dimensional labeling image to have the same number of preset sizes, different sampling intervals are preferably adopted for the three-dimensional labeling images with the real mediastinal lymph node lesion areas with different sizes in the sampling process, and larger sampling intervals are preferably adopted for the three-dimensional labeling images with the larger real mediastinal lymph node lesion areas; the three-dimensional labeling image with the smaller real mediastinal lymph node lesion area adopts smaller sampling intervals.
The second processing process is a rough segmentation process of the image block, and a depth residual U-net segmentation model is adopted to carry out rough segmentation of the image block:
the network structure of the depth residual U-net segmentation model comprises a compression process and an expansion process, wherein in the compression process, an input image block is processed by a convolutional layer, a batch normalization layer, an active layer, a convolutional layer and an additive layer to obtain a first characteristic diagram; the first feature map is subjected to down-sampling feature compression to obtain a second feature map; the second feature map is subjected to down-sampling feature compression to obtain a third feature map; the third feature map is subjected to down-sampling feature compression to obtain a fourth feature map; in the expansion process, the fourth feature map is subjected to up-sampling to obtain a fifth feature map, feature fusion is carried out on the fifth feature map and the third feature map, then, up-sampling is carried out to obtain a sixth feature map, feature fusion is carried out on the sixth feature map and the second feature map, then, up-sampling is carried out to obtain a seventh feature map, feature fusion is carried out on the seventh feature map and the first feature map, and then, after convolution processing, the seventh feature map and the first feature map are multiplied by Sigmoid activation to obtain a sub-segmentation probability map corresponding to the image block;
the third processing procedure is a fine segmentation procedure of the image block:
after the sub-segmentation probability maps are obtained, restoring the image blocks to the original positions according to the center coordinates of the image blocks, and averaging the probability values of the coincident pixel points among the sub-segmentation probability maps to obtain a total segmentation probability map of the three-dimensional image. The segmentation result of the image block is to take probability average to the coincident voxel points, and the similar probability is taken as the result of the voxel value of the focal region, which leads to the increase of false positive, thereby generating errors. In order to optimize the segmentation result, remove the false focus point, and further reduce the false positive, a class probability threshold is set, the voxel value greater than the class probability threshold is set as a first value, preferably 1, and the voxel value not greater than the class probability threshold is set as a second value, preferably 0. Some voxel values with low class probability can appear in the rough segmentation result, the region does not belong to a focus region, the region is set as a healthy region by setting a class probability threshold, an accurate segmentation result is obtained, and the establishment of a mediastinal lymph node segmentation model is realized.
In a preferred embodiment of the present invention, in step S41, the predetermined size is 24 pixels by 8 pixels.
In a preferred embodiment of the present invention, in step S44, the first value is 1.
In the preferred embodiment of the present invention, in step S44, the second value is 0.
In a preferred embodiment of the present invention, as shown in fig. 3, step S5 specifically includes:
step S51, inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain corresponding segmentation images;
step S52, respectively calculating a coincidence rate between each segmented image and the corresponding real mediastinal lymph node lesion area according to each segmented image and the corresponding real mediastinal lymph node lesion area, and comparing the coincidence rate with a preset coincidence rate threshold:
if the coincidence rate is not less than the coincidence rate threshold, adding the corresponding three-dimensional annotation image into a first image library, and then turning to step S53;
if the coincidence rate is smaller than the coincidence rate threshold value, adding the corresponding three-dimensional annotation image into a second image library, and then turning to the step S53;
step S53, respectively counting to obtain a first number of the three-dimensional annotation images in the first image library and a second number of the three-dimensional annotation images in the second image library, and calculating to obtain the segmentation accuracy of the mediastinal lymph node segmentation model according to the first number and the second number.
In the preferred embodiment of the present invention, in step S53, the segmentation accuracy is calculated according to the following formula:
Ar＝N1/(N1+N2)
wherein the content of the first and second substances,
Arrepresenting the segmentation accuracy;
N1representing a first quantity;
N2a second number is indicated.
A generation system of a mediastinal lymph node segmentation model, which applies any one of the above generation methods of mediastinal lymph node segmentation models, as shown in fig. 4, the generation system specifically includes:
the data acquisition module 1 is used for acquiring lung CT images of a plurality of thoracic surgery patients and performing three-dimensional reconstruction on each lung CT image to obtain a three-dimensional image corresponding to each thoracic surgery patient;
the data preprocessing module 2 is connected with the data acquisition module 1 and is used for respectively carrying out three-dimensional segmentation on each three-dimensional image to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area;
the data grouping module 3 is connected with the data preprocessing module 2 and is used for grouping the three-dimensional labeling images to obtain a training set, a testing set and a correcting set;
the model training module 4 is connected with the data grouping module 4 and used for training according to each three-dimensional labeling image in the training set to obtain a mediastinal lymph node segmentation model;
the model evaluation module 5 is respectively connected with the data grouping module 3 and the model training module 4 and is used for respectively inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image, and calculating the segmentation accuracy of the mediastinal lymph node segmentation model according to each segmentation image and the corresponding real mediastinal lymph node lesion area;
the data comparison module 6 is connected with the model evaluation module 5 and used for comparing the segmentation accuracy with a preset accuracy threshold, generating a first comparison result when the segmentation accuracy is not less than the accuracy threshold, and generating a second comparison result when the segmentation accuracy is less than the accuracy threshold;
the model correction module 7 is respectively connected with the data grouping module 3 and the data comparison module 6 and is used for correcting the mediastinal lymph node segmentation model according to the first comparison result and each three-dimensional annotation image in the correction set;
and the model storage module 8 is connected with the data comparison module 7 and is used for storing the mediastinal lymph node segmentation model according to the second comparison result so as to segment the mediastinal lymph node.
In a preferred embodiment of the present invention, as shown in fig. 4, the model training module 4 specifically includes:
the image sampling unit 41 is configured to establish a coordinate system on each three-dimensional labeled image in the training set, and respectively sample each three-dimensional labeled image according to the coordinate system to obtain a plurality of image blocks;
the image blocks sampled by each three-dimensional labeling image have the same number and the same preset size, and each image block has a central coordinate related to a coordinate system;
the feature learning unit 42 is connected with the image sampling unit 41, and is used for inputting each image block into a pre-generated depth residual error U-net segmentation model for feature learning aiming at each image block obtained by sampling each three-dimensional labeled image, so as to obtain a sub-segmentation probability map corresponding to each image block;
each voxel point of the corresponding image block in the sub-segmentation probability map is the probability value of the mediastinal lymph node;
an image restoration unit 43 connected to the feature learning unit 42, configured to restore each sub-segmentation probability map in the coordinate system where the corresponding image block is located according to the center coordinate, and average the probability values of the coincident pixel points between each sub-segmentation probability map to obtain a total segmentation probability map of the three-dimensional image;
a probability comparing unit 44 connected to the image restoring unit 43 for comparing the probability value of each voxel point in the total segmentation probability map with a preset class probability threshold, and setting the voxel value of the corresponding voxel point as a first numerical value representing that the voxel point is a focus area when the probability value is greater than the class probability threshold, and
if the probability value is not greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a second numerical value representing that the voxel point is not a focus area;
the data adjusting unit 45 is connected with the data comparing unit 44 and is used for generating a mediastinal lymph node segmentation result of the three-dimensional image according to the first numerical value and the second numerical value, calculating an error between the mediastinal lymph node segmentation result and a real mediastinal lymph node lesion area, and adjusting parameters of the depth residual error U-net segmentation model according to the error;
and the model generating unit 46 is connected with the data adjusting unit 45 and is used for obtaining the mediastinal lymph node segmentation model when the training is finished.
In a preferred embodiment of the present invention, as shown in fig. 4, the model evaluation module 5 specifically includes:
an image segmentation unit 51, configured to input each three-dimensional annotation image in the test set into a mediastinal lymph node segmentation model to obtain a corresponding segmentation image;
an image processing unit 52 connected to the image segmentation unit 51, configured to calculate, according to each segmented image and the corresponding real mediastinal lymph node lesion area, a coincidence rate between each segmented image and the corresponding real mediastinal lymph node lesion area, and add the corresponding three-dimensional labeling image into a first image library when the coincidence rate is not less than a preset coincidence rate threshold, and
adding the corresponding three-dimensional annotation image into a second image library when the coincidence rate is smaller than the coincidence rate threshold value;
and the data processing unit 53 is connected to the image processing unit 52, and is configured to count a first number of the three-dimensional labeled images in the first image library and a second number of the three-dimensional labeled images in the second image library respectively, and calculate a segmentation accuracy of the mediastinal lymph node segmentation model according to the first number and the second number.
While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention.

Claims (10)

1. A generation method of a mediastinal lymph node segmentation model is characterized by comprising the following steps:
step S1, acquiring lung CT images of a plurality of thoracic surgery patients, and performing three-dimensional reconstruction on the lung CT images to obtain a three-dimensional image corresponding to each thoracic surgery patient;
step S2, performing three-dimensional segmentation on each three-dimensional image to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area;
step S3, grouping the three-dimensional annotation images to obtain a training set, a test set and a correction set;
step S4, training according to each three-dimensional annotation image in the training set to obtain a mediastinal lymph node segmentation model;
step S5, inputting each three-dimensional labeling image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image, and calculating to obtain the segmentation accuracy of the mediastinal lymph node segmentation model according to each segmentation image and the corresponding real mediastinal lymph node focus area;
step S6, comparing the segmentation accuracy with a preset accuracy threshold:
if the segmentation accuracy is less than the accuracy threshold, then go to step S7;
if the segmentation accuracy is not less than the accuracy threshold, turning to step S8;
step S7, correcting the mediastinal lymph node segmentation model according to each three-dimensional annotation image in the correction set, and then returning to the step S6;
and step S8, storing the mediastinal lymph node segmentation model so as to segment the mediastinal lymph node.
2. The generation method of the mediastinal lymph node segmentation model according to claim 1, wherein the step S4 specifically includes:
step S41, establishing a coordinate system on each three-dimensional annotation image in the training set, and respectively performing overlapping sampling on each three-dimensional annotation image according to the coordinate system to obtain a plurality of image blocks;
the image blocks obtained by sampling each three-dimensional labeling image have the same number and the same preset size, and each image block has a central coordinate related to the coordinate system;
step S42, inputting each image block into a pre-generated depth residual error U-net segmentation model for feature learning aiming at each image block obtained by sampling each three-dimensional labeling image to obtain a sub-segmentation probability map corresponding to each image block;
the sub-segmentation probability map comprises probability values of all voxel points of the corresponding image blocks as mediastinal lymph nodes;
step S43, restoring each sub-segmentation probability map in the coordinate system where the corresponding image block is located according to the central coordinate, and averaging the probability values of coincident pixel points between each sub-segmentation probability map to obtain a total segmentation probability map of the three-dimensional image;
step S44, comparing the probability value of each voxel point in the total segmentation probability map with a preset class probability threshold:
if the probability value is greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a first numerical value representing that the voxel point is a focal zone, and then turning to step S45;
if the probability value is not greater than the class probability threshold, setting the voxel value of the corresponding voxel point to a second numerical value representing that the voxel point is not a focal zone, and then turning to step S45;
step S45, generating a mediastinal lymph node segmentation result of the three-dimensional labeling image according to the first numerical value and the second numerical value, calculating an error between the mediastinal lymph node segmentation result and the real mediastinal lymph node lesion area, and adjusting parameters of the depth residual U-net segmentation model according to the error;
and S46, repeatedly executing the steps S42 to S45 until the training is finished, and obtaining the mediastinal lymph node segmentation model.
3. The method for generating a mediastinal lymph node segmentation model according to claim 2, wherein the preset size is 24 pixels by 8 pixels in step S41.
4. The method for generating a mediastinal lymph node segmentation model according to claim 2, wherein in step S44, the first numerical value is 1.
5. The method for generating a mediastinal lymph node segmentation model according to claim 2, wherein in step S44, the second numerical value is 0.
6. The generation method of the mediastinal lymph node segmentation model according to claim 1, wherein the step S5 specifically includes:
step S51, inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain corresponding segmentation images;
step S52, respectively calculating a coincidence rate between each segmented image and the corresponding real mediastinal lymph node lesion area according to each segmented image and the corresponding real mediastinal lymph node lesion area, and comparing the coincidence rate with a preset coincidence rate threshold:
if the coincidence rate is not less than the coincidence rate threshold, adding the corresponding three-dimensional annotation image into a first image library, and then turning to step S53;
if the coincidence rate is smaller than the coincidence rate threshold value, adding the corresponding three-dimensional annotation image into a second image library, and then turning to the step S53;
step S53, respectively counting a first number of the three-dimensional labeling images in the first image library and a second number of the three-dimensional labeling images in the second image library, and calculating a segmentation accuracy of the mediastinal lymph node segmentation model according to the first number and the second number.
7. The generation method of a mediastinal lymph node segmentation model according to claim 6, wherein in step S53, the segmentation accuracy is calculated according to the following formula:
Ar＝N1/(N1+N2)
wherein the content of the first and second substances,
Arrepresenting the segmentation accuracy;
N1representing the first quantity;
N2representing said second number.
8. A generation system of a mediastinal lymph node segmentation model, which is characterized by applying the generation method of the mediastinal lymph node segmentation model according to any one of claims 1 to 7, and specifically comprises:
the data acquisition module is used for acquiring lung CT images of a plurality of thoracic surgery patients and performing three-dimensional reconstruction on the lung CT images to obtain a three-dimensional image corresponding to each thoracic surgery patient;
the data preprocessing module is connected with the data acquisition module and is used for respectively carrying out three-dimensional segmentation on each three-dimensional image to obtain a three-dimensional labeling image marked with a real mediastinal lymph node focus area;
the data grouping module is connected with the data preprocessing module and used for grouping the three-dimensional annotation images to obtain a training set, a testing set and a correcting set;
the model training module is connected with the data grouping module and used for training according to each three-dimensional labeling image in the training set to obtain a mediastinal lymph node segmentation model;
the model evaluation module is respectively connected with the data grouping module and the model training module and is used for respectively inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image and calculating the segmentation accuracy of the mediastinal lymph node segmentation model according to each segmentation image and the corresponding real mediastinal lymph node focus area;
the data comparison module is connected with the model evaluation module and used for comparing the segmentation accuracy with a preset accuracy threshold, generating a first comparison result when the segmentation accuracy is not less than the accuracy threshold, and generating a second comparison result when the segmentation accuracy is less than the accuracy threshold;
the model correction module is respectively connected with the data grouping module and the data comparison module and is used for correcting the mediastinal lymph node segmentation model according to the first comparison result and each three-dimensional annotation image in the correction set;
and the model storage module is connected with the data comparison module and used for storing the mediastinal lymph node segmentation model according to the second comparison result so as to segment the mediastinal lymph node.
9. The generation system of a mediastinal lymph node segmentation model according to claim 1, wherein the model training module specifically comprises:
the image sampling unit is used for establishing a coordinate system on each three-dimensional annotation image in the training set and respectively sampling each three-dimensional annotation image according to the coordinate system to obtain a plurality of image blocks;
the image blocks obtained by sampling each three-dimensional labeling image have the same number and the same preset size, and each image block has a central coordinate related to the coordinate system;
the characteristic learning unit is connected with the image sampling unit and used for inputting each image block into a pre-generated depth residual error U-net segmentation model for characteristic learning aiming at each image block obtained by sampling each three-dimensional labeling image to obtain a sub-segmentation probability map corresponding to each image block;
the sub-segmentation probability map comprises probability values of all voxel points of the corresponding image blocks as mediastinal lymph nodes;
the image restoration unit is connected with the feature learning unit and used for restoring each sub-segmentation probability map into the coordinate system where each corresponding image block is located according to the central coordinate, and averaging the probability values of coincident pixel points among the sub-segmentation probability maps to obtain a total segmentation probability map of the three-dimensional image;
a probability comparison unit connected with the image restoration unit and used for comparing the probability value of each voxel point in the total segmentation probability map with a preset class probability threshold value and setting the corresponding voxel value of the voxel point as a first numerical value representing that the voxel point is a focus area when the probability value is greater than the class probability threshold value, and
if the probability value is not greater than the class probability threshold, setting the voxel value of the corresponding voxel point as a second numerical value representing that the voxel point is not a focal zone;
the data adjusting unit is connected with the data comparing unit and used for generating a mediastinal lymph node segmentation result of the three-dimensional image according to the first numerical value and the second numerical value, calculating an error between the mediastinal lymph node segmentation result and the real mediastinal lymph node lesion area, and adjusting parameters of the depth residual U-net segmentation model according to the error;
and the model generation unit is connected with the data adjustment unit and is used for obtaining the mediastinal lymph node segmentation model after the training is finished.
10. The generation system of a mediastinal lymph node segmentation model according to claim 1, wherein the model evaluation module specifically comprises:
the image segmentation unit is used for respectively inputting each three-dimensional annotation image in the test set into the mediastinal lymph node segmentation model to obtain a corresponding segmentation image;
an image processing unit connected with the image segmentation unit and used for respectively calculating the coincidence rate between each segmented image and the corresponding real mediastinal lymph node focus region according to each segmented image and the corresponding real mediastinal lymph node focus region, and adding the corresponding three-dimensional annotation image into a first image library when the coincidence rate is not less than a preset coincidence rate threshold value, and
adding the corresponding three-dimensional labeling image into a second image library when the coincidence rate is smaller than the coincidence rate threshold value;
and the data processing unit is connected with the image processing unit and is used for respectively counting to obtain a first number of the three-dimensional labeling images in the first image library and a second number of the three-dimensional labeling images in the second image library, and calculating to obtain the segmentation accuracy of the mediastinal lymph node segmentation model according to the first number and the second number.
CN202010148259.6A 2020-03-05 2020-03-05 Method and system for generating mediastinum lymph node segmentation model Active CN111340825B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010148259.6A CN111340825B (en) 2020-03-05 2020-03-05 Method and system for generating mediastinum lymph node segmentation model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010148259.6A CN111340825B (en) 2020-03-05 2020-03-05 Method and system for generating mediastinum lymph node segmentation model

Publications (2)

Publication Number Publication Date
CN111340825A true CN111340825A (en) 2020-06-26
CN111340825B CN111340825B (en) 2023-05-09

Family

ID=71185914

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010148259.6A Active CN111340825B (en) 2020-03-05 2020-03-05 Method and system for generating mediastinum lymph node segmentation model

Country Status (1)

Country Link
CN (1) CN111340825B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112258500A (en) * 2020-11-10 2021-01-22 北京深睿博联科技有限责任公司 Lymph node partition method, apparatus, device and computer readable storage medium
CN113408633A (en) * 2021-06-29 2021-09-17 北京百度网讯科技有限公司 Method, apparatus, device and storage medium for outputting information
CN113450381A (en) * 2021-06-16 2021-09-28 上海深至信息科技有限公司 System and method for evaluating accuracy of image segmentation model
CN114820535A (en) * 2022-05-05 2022-07-29 深圳市铱硙医疗科技有限公司 Image detection method and device for aneurysm, computer device and storage medium
CN115272206A (en) * 2022-07-18 2022-11-01 深圳市医未医疗科技有限公司 Medical image processing method, medical image processing device, computer equipment and storage medium
WO2024045395A1 (en) * 2022-08-29 2024-03-07 杭州堃博生物科技有限公司 Ct image-based lesion detection method and apparatus, and computer readable storage medium

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016665A (en) * 2017-02-16 2017-08-04 浙江大学 A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks
US20190057488A1 (en) * 2017-08-17 2019-02-21 Boe Technology Group Co., Ltd. Image processing method and device
CN109685811A (en) * 2018-12-24 2019-04-26 北京大学第三医院 PET/CT hypermetabolism lymph node dividing method based on dual path U-net convolutional neural networks
CN109949276A (en) * 2019-02-28 2019-06-28 华中科技大学 A kind of lymph node detection method in improvement SegNet segmentation network
CN110310289A (en) * 2019-06-17 2019-10-08 北京交通大学 Lung tissue's image partition method based on deep learning
US20190333222A1 (en) * 2018-04-26 2019-10-31 NeuralSeg Ltd. Systems and methods for segmenting an image
CN110415253A (en) * 2019-05-06 2019-11-05 南京大学 A kind of point Interactive medical image dividing method based on deep neural network
CN110689525A (en) * 2019-09-09 2020-01-14 上海中医药大学附属龙华医院 Method and device for recognizing lymph nodes based on neural network

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016665A (en) * 2017-02-16 2017-08-04 浙江大学 A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks
US20190057488A1 (en) * 2017-08-17 2019-02-21 Boe Technology Group Co., Ltd. Image processing method and device
US20190333222A1 (en) * 2018-04-26 2019-10-31 NeuralSeg Ltd. Systems and methods for segmenting an image
CN109685811A (en) * 2018-12-24 2019-04-26 北京大学第三医院 PET/CT hypermetabolism lymph node dividing method based on dual path U-net convolutional neural networks
CN109949276A (en) * 2019-02-28 2019-06-28 华中科技大学 A kind of lymph node detection method in improvement SegNet segmentation network
CN110415253A (en) * 2019-05-06 2019-11-05 南京大学 A kind of point Interactive medical image dividing method based on deep neural network
CN110310289A (en) * 2019-06-17 2019-10-08 北京交通大学 Lung tissue's image partition method based on deep learning
CN110689525A (en) * 2019-09-09 2020-01-14 上海中医药大学附属龙华医院 Method and device for recognizing lymph nodes based on neural network

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112258500A (en) * 2020-11-10 2021-01-22 北京深睿博联科技有限责任公司 Lymph node partition method, apparatus, device and computer readable storage medium
CN112258500B (en) * 2020-11-10 2023-08-04 北京深睿博联科技有限责任公司 Lymph node partition method, apparatus, device, and computer-readable storage medium
CN113450381A (en) * 2021-06-16 2021-09-28 上海深至信息科技有限公司 System and method for evaluating accuracy of image segmentation model
CN113408633A (en) * 2021-06-29 2021-09-17 北京百度网讯科技有限公司 Method, apparatus, device and storage medium for outputting information
CN114820535A (en) * 2022-05-05 2022-07-29 深圳市铱硙医疗科技有限公司 Image detection method and device for aneurysm, computer device and storage medium
CN114820535B (en) * 2022-05-05 2023-09-12 深圳市铱硙医疗科技有限公司 Image detection method and device for aneurysm, computer equipment and storage medium
CN115272206A (en) * 2022-07-18 2022-11-01 深圳市医未医疗科技有限公司 Medical image processing method, medical image processing device, computer equipment and storage medium
WO2024045395A1 (en) * 2022-08-29 2024-03-07 杭州堃博生物科技有限公司 Ct image-based lesion detection method and apparatus, and computer readable storage medium

Also Published As

Publication number Publication date
CN111340825B (en) 2023-05-09

Similar Documents

Publication Publication Date Title
CN111340825B (en) Method and system for generating mediastinum lymph node segmentation model
JP5081390B2 (en) Method and system for monitoring tumor burden
CN111080584B (en) Quality control method for medical image, computer device and readable storage medium
TWI750583B (en) Medical image dividing method, device, and system, and image dividing method
US8649843B2 (en) Automated calcium scoring of the aorta
CN112885453A (en) Method and system for identifying pathological changes in subsequent medical images
CN108062749B (en) Identification method and device for levator ani fissure hole and electronic equipment
CN111553892A (en) Lung nodule segmentation calculation method, device and system based on deep learning
CN110619635B (en) Hepatocellular carcinoma magnetic resonance image segmentation system and method based on deep learning
US8520923B2 (en) Reporting organ volume for a medical digital image
KR102328229B1 (en) AI based tumor detection and diagnostic method using 3D medical image
CN111311626A (en) Skull fracture automatic detection method based on CT image and electronic medium
CN114270397A (en) System and method for determining fluid and tissue volume estimates using electrical property tomography
CN115100067A (en) Image reconstruction method and device
CN114332132A (en) Image segmentation method and device and computer equipment
CN116091466A (en) Image analysis method, computer device, and storage medium
WO2022164374A1 (en) Automated measurement of morphometric and geometric parameters of large vessels in computed tomography pulmonary angiography
CN116580037B (en) Nasopharyngeal carcinoma image segmentation method and system based on deep learning
CN116309647B (en) Method for constructing craniocerebral lesion image segmentation model, image segmentation method and device
Armya et al. Medical images segmentation based on unsupervised algorithms: a review
CN110992312B (en) Medical image processing method, medical image processing device, storage medium and computer equipment
Dandıl et al. A Mask R-CNN based Approach for Automatic Lung Segmentation in Computed Tomography Scans
Benrabha et al. Automatic ROI detection and classification of the achilles tendon ultrasound images
Malinda et al. Lumbar vertebrae synthetic segmentation in computed tomography images using hybrid deep generative adversarial networks
CN114820483A (en) Image detection method and device and computer equipment

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant