CN112785581A - Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning - Google Patents

Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning Download PDF

Info

Publication number
CN112785581A
CN112785581A CN202110126100.9A CN202110126100A CN112785581A CN 112785581 A CN112785581 A CN 112785581A CN 202110126100 A CN202110126100 A CN 202110126100A CN 112785581 A CN112785581 A CN 112785581A
Authority
CN
China
Prior art keywords
deep learning
training
data
learning model
module
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.)
Pending
Application number
CN202110126100.9A
Other languages
Chinese (zh)
Inventor
汤敏
曾蒙苏
吴东
谢添武
金开璞
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.)
Zhongshan Hospital Fudan University
Original Assignee
Zhongshan Hospital Fudan University
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 Zhongshan Hospital Fudan University filed Critical Zhongshan Hospital Fudan University
Priority to CN202110126100.9A priority Critical patent/CN112785581A/en
Publication of CN112785581A publication Critical patent/CN112785581A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30101—Blood vessel; Artery; Vein; Vascular
    • G06T2207/30104—Vascular flow; Blood flow; Perfusion

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a training method and a device for extracting and training large blood vessel CTA imaging based on deep learning, belonging to the technical field of medical imaging. Compared with the traditional image processing method, the technical scheme has stronger robustness on conditions such as image extraction, reconstruction and analysis, and the like, and the detection is faster and more accurate. Compared with a measurement detection method based on manual reconstruction, the method uses a fast positioning network to extract the ring area before segmentation, reduces the segmented network sample space, reduces the input size of the segmented network, and ensures that the segmented network has less calculation amount, better real-time performance, lower requirement on hardware processor configuration and less energy consumption. The accuracy and speed of the existing deep learning-based method for segmenting the blood vessel can be improved.

Description

Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning
Technical Field
The invention relates to a training method and a device for extracting and training large blood vessel CTA imaging based on deep learning, belonging to the technical field of medical imaging.
Background
CT angiography imaging (CTA) is a first important tool for diagnosing acute aortic dissections. For patients suspected of aortic dissection, if clinical features of ascending aorta involvement are not suggested and hemodynamics are stable, CTA should be performed first, especially in emergency treatment, other examination conditions are limited, medical resources are under tension, and therefore, most patients suspected of acute aortic dissection are evaluated by thoracic/abdominal dynamic contrast enhancement thin-layer CTA. CT scanning, in contrast to other approaches, is minimally operator dependent, can provide useful anatomical information for surgery and intravascular treatment, and can collect information for follow-up analysis and measurement. Most importantly, three-dimensional CT reconstruction helps in treatment planning, and cross-sectional imaging most likely shows the structural relationship of the true lumen and potential aortic branch damage. However, manual segmentation of the vessels in the images is time and labor consuming, requires a radiologist with experience of more than 3-5 years to measure and reconstruct layer by layer, and requires more than 30 minutes for post-processing of a vessel reconstruction. However, emergency resources are limited, especially in a large-scale comprehensive medical center, patients are concentrated, the conditions of the patients are complicated, and accurate and rapid medical auxiliary blood vessel analysis is particularly needed.
The problem of fully automatic segmentation of blood vessels in medical images has been a major problem of concern in the medical and academic circles. The human large blood vessel system is mainly composed of the thoracic aorta and the abdominal aorta and a branch blood vessel network, and the whole set of blood vessel system provides vital nutrition and oxygen for the whole body. One of the most common techniques for imaging large vessel networks is vascular imaging (CTA), by which the display of blood vessels in medical images can be enhanced and visual analysis can be facilitated. In the prior art, the large vessel CT imaging data acquisition technology comprises the use of electrocardio-gating and no electrocardio-gating, and a deep learning method is utilized to automatically segment the large vessel artery network. However, the current blood vessel segmentation method for deep learning does not consider multi-scale information, so that the obtained final blood vessel segmentation effect is not ideal enough. Compared with manually defined features, the deep learning method is widely applied to actual production and life due to the super-strong feature expression capability of the deep learning method. The core idea of the deep learning method is to obtain the abstract expression of the original data by utilizing multiple nonlinear transformations, so as to further obtain the solution of the original task. In the field of image processing, a Convolutional Neural Network (CNN) in a deep learning method is far superior to a conventional method in performance of many problems. In recent years, convolutional neural networks are increasingly applied to the field of medical image processing, such as classification, segmentation, registration, denoising and other tasks of medical images. Similarly, deep learning has also advanced well in the field of vessel segmentation.
The existing method for segmenting blood vessels in medical images based on the deep learning technology mainly has the following problems:
(1) no multi-scale feature representation is obtained. Because the blood vessels in the medical image have multiple scales, the multi-scale information is not considered in many existing blood vessel segmentation methods based on deep learning, and therefore the obtained final blood vessel segmentation effect is not ideal enough;
(2) the extracted context information is not rich enough. Because the blood vessel is a slender structure, the context information in a small range cannot express a complex blood vessel structure, so that the network is easy to generate over-segmentation or under-segmentation results in the segmentation process;
(3) the network parameters are large, and overfitting is easy to occur in the training process. Since the amount of medical image data available for vessel segmentation is small, for larger network structures, overfitting is likely to occur during training.
Disclosure of Invention
The invention aims to solve the technical problem of the technical defects in the CTA imaging method for detecting the large blood vessel in the prior art.
In order to solve the problem, the technical scheme adopted by the invention is to provide a training method for extracting and training the imaging of the large blood vessel CTA based on deep learning, which comprises the following steps:
step 1: preprocessing the image to obtain preprocessed data;
step 2: taking the preprocessed data as the input of a deep learning model, and training the deep learning model to obtain a trained deep learning model, wherein the deep learning model comprises a U-shaped network;
and step 3: acquiring data to be segmented, and inputting the data to be segmented as a judgment model; carrying out segmentation prediction on data to be segmented by using a trained deep learning model;
and 4, step 4: and obtaining a large blood vessel segmentation result by a threshold segmentation method.
Preferably, the preprocessing in step 1 includes image normalization and normalization processing.
Preferably, in step 2, the step of using the preprocessed data as the input of the deep learning model includes segmenting the preprocessed data to obtain a plurality of data blocks; and taking the data block as an input of the deep learning model in a batch unit.
Preferably, the training of the deep learning model in step 2 to obtain the trained deep learning model includes training the deep learning network in a manner of iteratively back-propagating and optimizing network parameters, and obtaining the trained deep learning model when the training process converges.
Preferably, the U-type network in the step 2 includes an encoding part and a decoding part; the coding part comprises three reduction modules; the decoding part comprises three upsampling modules and four dense expansion modules, and each upsampling module is surrounded by a dense expansion module.
The invention also provides a great vessel segmentation device based on the deep learning technology, which applies the training method for extracting and training the CTA imaging of the great vessel based on the deep learning, and comprises a preprocessing module, a training module, a prediction module and a processing module; the data are processed sequentially through the preprocessing module, the training module, the prediction module and the processing module.
Preferably, the preprocessing module is configured to preprocess a human body large magnetic resonance angiography CTA image to obtain preprocessing data; the training module is used for taking the preprocessed data as the input of the deep learning model, and training the deep learning model to obtain a trained deep learning model; the deep learning model comprises a U-shaped network (U-Net); the prediction module is used for acquiring data to be segmented, taking the data to be segmented as the input of a trained deep learning model, and performing segmentation prediction on the data to be segmented by using the trained deep learning model to obtain large vessel binary volume data; the processing module is used for processing the large blood vessel binary volume data through a threshold segmentation method to obtain a large blood vessel segmentation result.
Preferably, the pre-processing comprises normalization and normalization processing; the training module is used for segmenting the preprocessed data to obtain a plurality of data blocks; taking the data block as the input of a deep learning model by taking the batch as a unit; the training module trains the deep learning network in a mode of iteratively propagating the optimized network parameters in the reverse direction, and a trained deep learning model is obtained when the training process is converged.
Preferably, the U-type network comprises an encoding part and a decoding part; the encoding portion includes three reduction modules; the decoding part comprises three upsampling modules and four dense expansion modules, and each upsampling module is surrounded by a dense expansion module.
Compared with the prior art, the invention has the following beneficial effects:
compared with the existing depth model for automatically segmenting the great vessel network from TOF-CTA, the method for extracting and training the great vessel CTA imaging based on the deep learning and the great vessel segmentation device based on the deep learning provided by the invention have the following advantages that:
(1) the segmentation precision is higher. Experiments prove that the evaluation index of DSC (clock Similarity coefficient) is superior to that of the prior deep learning model;
(2) the number of model parameters is less. Compared with the existing model for segmentation, the model training parameters provided by the invention are reduced by 39-44%;
(3) the feature extraction capability is stronger. Through mechanisms such as convolution, dense connection, jumping convolution and the like, the network can extract multi-scale features and fuse high-level semantic information, so that the model has better feature extraction capability on an original image.
The accuracy and speed of the existing deep learning-based method for segmenting the blood vessel can be improved.
Compared with the traditional image processing method, the training method and the training device for extracting and training the large blood CTA imaging based on deep learning have the advantages that robustness is stronger in conditions of image extraction, reconstruction, analysis and the like, and detection is quicker and more accurate. Compared with a measurement detection method based on manual reconstruction, the method uses a fast positioning network to extract the ring area before segmentation, reduces the segmented network sample space, reduces the input size of the segmented network, and ensures that the segmented network has less calculation amount, better real-time performance, lower requirement on hardware processor configuration and less energy consumption.
Drawings
Fig. 1 is a general flowchart of a thoracoabdominal aorta segmentation method based on a deep learning technique according to an embodiment of the present invention;
FIG. 2 is a flowchart of a deep learning model training process according to an embodiment of the present invention;
fig. 3 is a flowchart of another thoracoabdominal aorta segmentation method based on a deep learning technique according to an embodiment of the present invention;
fig. 4 is a flowchart of a thoracoabdominal aorta segmentation method based on a deep learning technique according to another embodiment of the present invention;
fig. 5 is a schematic structural diagram of a thoracoabdominal aorta segmentation apparatus based on a deep learning technique according to an embodiment of the present invention.
Detailed Description
In order to make the invention more comprehensible, preferred embodiments are described in detail below with reference to the accompanying drawings:
as shown in fig. 1-5, the present invention provides a training method for extracting and training a large blood vessel CTA image based on deep learning, comprising the following steps:
step 1: preprocessing the image to obtain preprocessed data;
step 2: taking the preprocessed data as the input of a deep learning model, and training the deep learning model to obtain a trained deep learning model, wherein the deep learning model comprises a U-shaped network;
and step 3: acquiring data to be segmented, and inputting the data to be segmented as a judgment model; carrying out segmentation prediction on data to be segmented by using a trained deep learning model;
and 4, step 4: and obtaining a large blood vessel segmentation result by a threshold segmentation method.
The preprocessing in the step 1 includes image normalization and normalization processing.
In the step 2, the step of taking the preprocessed data as the input of the deep learning model comprises the step of segmenting the preprocessed data to obtain a plurality of data blocks; and taking the data block as an input of the deep learning model in a batch unit.
Training the deep learning model in the step 2 to obtain the trained deep learning model comprises training the deep learning network in a mode of iteratively propagating back and optimizing network parameters, and obtaining the trained deep learning model when the training process is converged.
The U-type network in the above step 2 includes an encoding part and a decoding part; the coding part comprises three reduction modules; the decoding part comprises three upsampling modules and four dense expansion modules, and each upsampling module is surrounded by a dense expansion module.
The invention also provides a great vessel segmentation device based on the deep learning technology, which applies the training method for extracting and training the CTA imaging of the great vessel based on the deep learning, and comprises a preprocessing module, a training module, a prediction module and a processing module; the data are processed sequentially through the preprocessing module, the training module, the prediction module and the processing module.
The preprocessing module is used for preprocessing a human body large magnetic resonance angiography CTA image to obtain preprocessing data; the training module is used for taking the preprocessed data as the input of the deep learning model, and training the deep learning model to obtain a trained deep learning model; the deep learning model comprises a U-shaped network; the prediction module is used for acquiring data to be segmented, taking the data to be segmented as the input of a trained deep learning model, and performing segmentation prediction on the data to be segmented by using the trained deep learning model to obtain large vessel binary volume data; the processing module is used for processing the large blood vessel binary volume data through a threshold segmentation method to obtain a large blood vessel segmentation result.
The pretreatment comprises normalization and normalization; the training module is used for segmenting the preprocessed data to obtain a plurality of data blocks; taking the data block as the input of a deep learning model by taking the batch as a unit; the training module trains the deep learning network in a mode of iteratively propagating the optimized network parameters in the reverse direction, and a trained deep learning model is obtained when the training process is converged.
The U-type network includes an encoding part and a decoding part; the coding part comprises three reduction modules; the decoding part comprises three upsampling modules and four dense expansion modules, and each upsampling module is surrounded by a dense expansion module.
Examples
Fig. 1 shows a general flowchart of a cerebral artery segmentation method based on a deep learning technique according to an embodiment of the present invention, and referring to fig. 1, the cerebral artery segmentation method based on a deep learning technique according to an embodiment of the present invention includes:
step 101, data preprocessing:
wherein, the data is human thoraco-abdominal major blood vessel magnetic resonance angiography CTA data.
In the embodiment of the invention, the input data is a human body thoracoabdominal major blood vessel CTA image.
In the embodiment of the invention, the preprocessing operation comprises two steps of normalization preprocessing and normalization processing.
The CTA data and the labeled data are made to be isotropic in voxel spacing, a mm x a mm, by a trilinear interpolation method. Because the labeling data is a binary image, a nearest neighbor interpolation mode is adopted in the process of trilinear interpolation, and the gray value of the interpolated voxel is determined by the gray value of the nearest neighbor voxel, so that the labeling image is still a binary image after interpolation. And then, the data is normalized, namely the gray value of each voxel is divided by the maximum gray value in the current image, so that the convergence performance of the network in the training process can be improved.
Step 102, deep learning model training:
by designing and realizing a deep neural network suitable for cerebral artery segmentation, a model capable of automatically predicting cerebral arteries in data is obtained after big data training.
Specifically, firstly, a deep neural network suitable for human thoracoabdominal major vascular artery segmentation is designed and realized; the input of the deep neural network is a data preprocessing result, the deep learning network designed by the embodiment of the invention is trained in a mode of iteratively propagating back and optimizing network parameters, and a model for cerebral artery segmentation can be obtained when the training process is converged.
The embodiment of the invention provides a new end-to-end convolution neural network-U-type network U-Net for fully automatically segmenting the major thoracoabdominal aorta artery, the network can obtain multi-scale representation of the major thoracoabdominal aorta artery network, and the final major thoracoabdominal aorta artery network segmentation result can be obtained with less calculation amount.
The U-shaped network structure provided by the embodiment of the invention consists of an encoding part and a decoding part. The reduction module and the dense expansion module provided by the embodiment of the invention are respectively applied to the encoding part and the decoding part. Wherein: the encoding part is composed of three reduction modules, and the resolution of input data or a characteristic graph is reduced by one time after passing through each reduction module; each up-sampling module of the decoding part is surrounded by the dense expansion module, wherein the number of the up-sampling modules is the same as that of the reduction modules in the encoding stage, and the network of the embodiment of the invention is composed of four dense expansion modules which are specifically arranged as a dense expansion module 1, an up-sampling module 1, a dense expansion module 2, an up-sampling module 2, a dense expansion module 3, an up-sampling module 3 and a dense expansion module 4. The model can extract the characteristics of multi-scale receptive fields, can fuse high-level semantic characteristics together, and obtains a good cerebral artery segmentation effect under the condition of less parameter quantity.
Therefore, the network is one of full convolution coding and decoding networks, a good image segmentation effect can be obtained only by a small amount of training data for the coding and decoding networks, and the segmentation accuracy and speed are superior to those of a convolution neural network based on a sliding window. The codec network is a full convolutional neural network, i.e. the network architecture does not include a full connection layer. The encoding stage of the encoding and decoding network mainly learns the feature representation of the original data in a mode of rolling and reducing the image scale, and captures the context information of the input image. After the encoding stage, a concentrated representation of the original data can be obtained, the feature map is gradually restored to the size of the original image in the decoding stage through steps of up-sampling and the like, and the number of final output channels of the network is the same as the number of segmentation classes. The main function of the decoder is accurate positioning.
The neuron network adopts a dense connection mechanism, an inclusion mechanism, a hole convolution mechanism and the like in specific modules of encoding and decoding, so that the network can obtain a better segmentation result by needing less parameter quantity. In addition, in order to control the parameter size of the network, the neuron network of the invention also adopts a one-dimensional convolution operation, namely the convolution kernel size is 1 × 1, and is used for reducing the number of the characteristic diagrams. The split network consists of two parts, an encoder and a decoder. The encoder part continuously reduces the input size and increases the receptive field, the decoder part performs up-sampling on the input, and finally the output size is the original input size.
Inputting the preprocessed data into a network for training, and updating parameters in the network through a back propagation algorithm until the model converges.
Step 103, applying the trained model to carry out cerebral artery prediction segmentation:
after the trained model is obtained in step 102, the model can be directly used for cerebral artery segmentation prediction in new data.
104, obtaining a cerebral artery segmentation result by using a threshold segmentation method:
the probability that each voxel of the original image belongs to the cerebral artery is obtained in step 103, and the final cerebral artery binary volume data can be obtained by segmentation through the threshold segmentation method in the step.
As an optional implementation manner of the embodiment of the present invention, as a refinement and extension of step 102 in the embodiment shown in fig. 1, the present invention further provides a flowchart of a deep learning model training process. As shown in fig. 2, the deep learning model training process in this embodiment includes:
step 201, CTA data acquisition:
this step acquires a sufficient amount of human head CTA data for model training.
Step 202, data preprocessing:
this step is the same as step 101.
Step 203, data segmentation:
since the original image is large and hardware resources such as video memory and memory are limited, the input image is divided into blocks (patch) of size m × n × k.
Step 204, inputting the data block into the deep neural network to train the model by taking the batch as a unit:
and inputting the data blocks obtained in the step 203 into a neural network in batches for training, and updating parameters in the network through a back propagation algorithm until the model converges.
Step 205, obtaining a trained deep learning model for cerebral artery segmentation:
and saving the model architecture and the model weight of the model trained in the step 204 for the prediction stage.
As an optional implementation manner of the embodiment of the present invention, as an implementation manner of the method shown in fig. 1, fig. 3 further illustrates a flowchart of another cerebral artery segmentation prediction method based on deep learning according to the embodiment of the present invention, where the cerebral artery segmentation prediction method based on deep learning includes the following steps:
step 301, CTA data acquisition: this step 301 performs the same operation as step 201.
Step 302, data preprocessing: this step 302 performs the same operation as step 202.
Step 303, data segmentation: this step 303 performs the same operation as step 203.
Step 304, inputting the data into the trained deep neural network by taking batches as units to obtain a segmentation prediction result: in the model prediction stage, after data is input into a trained model, the final output is a prediction result, and the process does not change network parameters.
Step 305, threshold segmentation:
the U-shaped network realized by the invention finally adopts a Sigmoid activation function to obtain a final output result, so that the network outputs the probability that each voxel point belongs to the blood vessel. That is, the voxels in the probability map output by the network whose median is greater than the threshold are foreground (blood vessels), otherwise, the voxels are background.
Step 306, outputting a cerebral artery binary segmentation result:
after the steps are carried out, the segmentation result can be stored in a file form, or the segmentation result is drawn by adopting an isosurface generation or volume drawing method and displayed to a user.
As an optional implementation manner of the embodiment of the present invention, fig. 4 shows a flowchart of a further method for segmenting a cerebral artery based on a deep learning technique according to the embodiment of the present invention, and referring to fig. 4, the further method for segmenting a cerebral artery based on a deep learning technique according to the embodiment of the present invention includes:
step 401, preprocessing a human body thoracoabdominal major blood vessel magnetic resonance angiography CTA image to obtain preprocessed data. As an optional implementation of the embodiment of the present invention, the preprocessing includes: normalization preprocessing and normalization processing.
Specifically, the CTA data and the labeled data are first rendered with isotropic voxel spacing, a mm x a mm, using a trilinear interpolation method. Because the labeling data is a binary image, a nearest neighbor interpolation mode is adopted in the process of trilinear interpolation, and the gray value of the interpolated voxel is determined by the gray value of the nearest neighbor voxel, so that the labeling image is still a binary image after interpolation. And then, normalizing the data, namely dividing the gray value of each voxel by the maximum gray value in the current image, so that the convergence performance of the network in the training process can be improved.
And step 402, taking the preprocessed data as input of a deep learning model, and training the deep learning model to obtain a trained deep learning model, wherein the deep learning model comprises a U-shaped network.
As an optional implementation of the embodiment of the present invention, the using the preprocessed data as input to the deep learning model includes: segmenting the preprocessed data to obtain a plurality of data blocks; and taking the data block as an input of the deep learning model in a batch unit.
As an optional implementation manner of the embodiment of the present invention, training the deep learning model to obtain the trained deep learning model includes: and training the deep learning network in a mode of iteratively propagating the optimized network parameters in the reverse direction, and obtaining a well-trained deep learning model when the training process is converged.
As an optional implementation manner of the embodiment of the present invention, the U-type network includes: an encoding section and a decoding section; the encoding section includes: three reduction modules; the decoding section includes: three upsampling modules and four dense expansion modules, each upsampling module being surrounded by a dense expansion module.
Specifically, the embodiment of the invention provides a new end-to-end fully-automatic thoracoabdominal main great vessel artery segmentation network-U-type network by utilizing a convolutional neural network, wherein the network can obtain multi-scale representation of the thoracoabdominal main great vessel artery network, and a final thoracoabdominal main great vessel artery network segmentation result can be obtained by less calculation.
The U-shaped network structure provided by the embodiment of the invention consists of an encoding part and a decoding part. The reduction module and the dense expansion module provided by the embodiment of the invention are respectively applied to the encoding part and the decoding part. Wherein: the encoding part is composed of three reduction modules, and the resolution of input data or a characteristic graph is reduced by one time after passing through each reduction module; each up-sampling module of the decoding part is surrounded by a dense expansion module, wherein the number of the up-sampling modules is the same as that of the reduction modules in the encoding stage, and the U-shaped network of the embodiment of the invention is composed of four dense expansion modules which are specifically arranged as a dense expansion module 1, an up-sampling module 1, a dense expansion module 2, an up-sampling module 2, a dense expansion module 3, an up-sampling module 3 and a dense expansion module 4. The model can extract the characteristics of multi-scale receptive fields, can fuse high-level semantic characteristics together, and obtains a good cerebral artery segmentation effect under the condition of less parameter quantity.
Therefore, the network is one of full convolution coding and decoding networks, for the coding and decoding networks such as U-Net, a network structure specially designed for a segmentation task is adopted, good image segmentation effect can be obtained only by a small amount of training data, and the segmentation accuracy and speed are superior to those of a convolution neural network based on a sliding window. The codec network is a full convolutional neural network, i.e. the network architecture does not include a full connection layer. The encoding stage of the encoding and decoding network mainly learns the feature representation of the original data in a mode of rolling and reducing the image scale, and captures the context information of the input image. After the encoding stage, a concentrated representation of the original data can be obtained, the feature map is gradually restored to the size of the original image in the decoding stage through steps of up-sampling and the like, and the number of final output channels of the network is the same as the number of segmentation classes. The main function of the decoder is accurate positioning.
The neuron network adopts a dense connection mechanism, an inclusion mechanism, a hole convolution mechanism and the like in specific modules of encoding and decoding, so that the network can obtain a better segmentation result by needing less parameter quantity. In addition, in order to control the parameter size of the network, the neuron network of the invention also adopts a one-dimensional convolution operation, namely the convolution kernel size is 1 × 1, and is used for reducing the number of the characteristic diagrams. The neuron network is composed of an encoder and a decoder. The encoder part continuously reduces the input size and increases the receptive field, the decoder part performs up-sampling on the input, and finally the output size is the original input size.
And inputting the data blocks obtained in the step 203 into a neural network in batches for training, and updating parameters in the network through a back propagation algorithm until the model converges.
Step 403, acquiring data to be segmented, using the data to be segmented as input of a trained deep learning model, and performing segmentation prediction on the data to be segmented by using the trained deep learning model to obtain thoracic and abdominal major vascular artery binary volume data;
and step 404, processing the thoracic and abdominal major vessel artery binary volume data through a threshold segmentation method to obtain a cerebral artery segmentation result.
Specifically, the neural network realized by the invention finally adopts a Sigmoid activation function to obtain a final output result, so that the network output is the probability that each voxel point belongs to a blood vessel. That is, the voxels in the probability map output by the network whose median is greater than the threshold are foreground (blood vessels), otherwise, the voxels are background.
As an optional implementation manner of the embodiment of the present invention, after step 404, the method for segmenting a cerebral artery based on a deep learning technique according to the embodiment of the present invention may further include: and outputting a cerebral artery segmentation result. Specifically, after the steps are carried out, the segmentation result can be stored in a file form, or the segmentation result is drawn by adopting an isosurface generation or volume drawing method and displayed to a user.
Therefore, compared with the existing depth model for automatically segmenting the cerebral artery network from TOF-CTA, the method for segmenting the cerebral artery based on the deep learning technology provided by the invention has the advantages that:
(1) the segmentation precision is higher. Experiments prove that the evaluation index of DSC (clock Similarity coefficient) is superior to that of the prior deep learning model;
(2) the number of model parameters is less. Compared with the existing model for segmentation, the model training parameters provided by the invention are reduced by 39-44%;
(3) the feature extraction capability is stronger. Through mechanisms such as hole convolution, dense connection, jumping convolution and the like, the network can extract multi-scale features and fuse high-level semantic information, so that the model has better feature extraction capability on an original image.
The accuracy and speed of the existing deep learning-based method for segmenting the blood vessel can be improved.
Fig. 5 is a schematic structural diagram of a cerebral artery segmentation apparatus based on a deep learning technique according to an embodiment of the present invention, where the cerebral artery segmentation apparatus based on a deep learning technique according to an embodiment of the present invention uses the cerebral artery segmentation method based on a deep learning technique, and only the cerebral artery segmentation apparatus based on a deep learning technique according to an embodiment of the present invention is briefly described herein, but other matters are not considered to be the best, please refer to the related description of the cerebral artery segmentation method based on a deep learning technique, which is not described herein again, referring to fig. 5, and the cerebral artery segmentation apparatus based on a deep learning technique according to an embodiment of the present invention includes:
the preprocessing module is used for preprocessing a human body thoracoabdominal major blood vessel magnetic resonance angiography CTA image to obtain preprocessing data;
the training module is used for taking the preprocessed data as the input of the deep learning model, training the deep learning model and obtaining the trained deep learning model, wherein the deep learning model comprises a U-shaped network;
the prediction module is used for acquiring data to be segmented, taking the data to be segmented as the input of a trained deep learning model, and performing segmentation prediction on the data to be segmented by using the trained deep learning model to obtain thoracic and abdominal major vascular artery binary volume data;
and the processing module is used for processing the thoracic and abdominal main great vessel artery binary volume data through a threshold segmentation method to obtain a cerebral artery segmentation result.
As an optional implementation of the embodiment of the present invention, the preprocessing includes: normalization preprocessing and normalization processing.
As an optional implementation of the embodiment of the present invention, the training module takes the preprocessed data as input to the deep learning model by: the training module is specifically used for segmenting the preprocessed data to obtain a plurality of data blocks; and taking the data block as an input of the deep learning model in a batch unit.
As an optional implementation manner of the embodiment of the present invention, the training module trains the deep learning model to obtain a trained deep learning model as follows: and the training module is specifically used for training the deep learning network in a mode of iteratively back-propagating and optimizing network parameters, and obtaining a well-trained deep learning model when the training process is converged.
As an optional implementation manner of the embodiment of the present invention, the U-type network includes: an encoding section and a decoding section; the encoding section includes: three reduction modules; the decoding section includes: three upsampling modules and four dense expansion modules, each upsampling module being surrounded by a dense expansion module.
As an optional implementation manner of the embodiment of the present invention, the cerebral artery segmentation apparatus based on the deep learning technique provided in the embodiment of the present invention may further include: and the output module is used for outputting the cerebral artery segmentation result.
Therefore, compared with the existing depth model for automatically segmenting the cerebral artery network from TOF-CTA, the cerebral artery segmentation device based on the deep learning technology provided by the invention has the following advantages that:
(1) the segmentation precision is higher. Experiments prove that the evaluation index of DSC (clock Similarity coefficient) is superior to that of the prior deep learning model;
(2) the number of model parameters is less. Compared with the existing model for segmentation, the model training parameters provided by the invention are reduced by 39-44%;
(3) the feature extraction capability is stronger. Through mechanisms such as hole convolution, dense connection, jumping convolution and the like, the network can extract multi-scale features and fuse high-level semantic information, so that the model has better feature extraction capability on an original image.
The accuracy and speed of the existing deep learning-based method for segmenting the blood vessel can be improved.
While the invention has been described with respect to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention. Those skilled in the art can make various changes, modifications and equivalent arrangements, which are equivalent to the embodiments of the present invention, without departing from the spirit and scope of the present invention, and which may be made by utilizing the techniques disclosed above; meanwhile, any changes, modifications and variations of the above-described embodiments, which are equivalent to those of the technical spirit of the present invention, are within the scope of the technical solution of the present invention.

Claims (9)

1. A training method for extracting and training large blood vessel CTA imaging based on deep learning is characterized in that: the method comprises the following steps:
step 1: preprocessing the image to obtain preprocessed data;
step 2: taking the preprocessed data as the input of a deep learning model, and training the deep learning model to obtain a trained deep learning model, wherein the deep learning model comprises a U-shaped network;
and step 3: acquiring data to be segmented, and inputting the data to be segmented as a judgment model; carrying out segmentation prediction on data to be segmented by using a trained deep learning model;
and 4, step 4: and obtaining a large blood vessel segmentation result by a threshold segmentation method.
2. The deep learning based method for training CTA imaging of large blood vessels by extraction as claimed in claim 1, wherein: the preprocessing in the step 1 comprises image standardization and normalization processing.
3. The deep learning based method for training CTA imaging of large blood vessels by extraction as claimed in claim 1, wherein: in the step 2, the step of taking the preprocessed data as the input of the deep learning model comprises the step of segmenting the preprocessed data to obtain a plurality of data blocks; and taking the data block as an input of the deep learning model in a batch unit.
4. The deep learning based method for training CTA imaging of large blood vessels by extraction as claimed in claim 1, wherein: and 2, training the deep learning model in the step 2 to obtain the trained deep learning model, wherein the training of the deep learning network is realized by iteratively propagating the optimized network parameters in the reverse direction, and the trained deep learning model is obtained when the training process is converged.
5. The deep learning based method for training CTA imaging of large blood vessels by extraction as claimed in claim 1, wherein: the U-type network in the step 2 comprises an encoding part and a decoding part; the coding part comprises three reduction modules; the decoding part comprises three upsampling modules and four dense expansion modules, and each upsampling module is surrounded by a dense expansion module.
6. A great vessel segmentation device based on deep learning technology applying a training method for training CTA imaging of the great vessel based on deep learning as claimed in any one of claims 1 to 5, characterized in that: the device comprises a preprocessing module, a training module, a prediction module and a processing module; the data are processed sequentially through the preprocessing module, the training module, the prediction module and the processing module.
7. The great vessel segmentation device based on the deep learning technique as claimed in claim 6, wherein: the preprocessing module is used for preprocessing a human body large magnetic resonance angiography CTA image to obtain preprocessing data; the training module is used for taking the preprocessed data as the input of the deep learning model, and training the deep learning model to obtain a trained deep learning model; the deep learning model comprises a U-shaped network; the prediction module is used for acquiring data to be segmented, taking the data to be segmented as the input of a trained deep learning model, and performing segmentation prediction on the data to be segmented by using the trained deep learning model to obtain large vessel binary volume data; the processing module is used for processing the large blood vessel binary volume data through a threshold segmentation method to obtain a large blood vessel segmentation result.
8. The large vessel segmentation device based on the deep learning technique as claimed in claim 7, wherein: the preprocessing comprises normalization and normalization processing; the training module is used for segmenting the preprocessed data to obtain a plurality of data blocks; taking the data block as the input of a deep learning model by taking the batch as a unit; the training module trains the deep learning network in a mode of iteratively propagating the optimized network parameters in the reverse direction, and a trained deep learning model is obtained when the training process is converged.
9. The large vessel segmentation device based on the deep learning technique as claimed in claim 8, wherein: the U-type network includes an encoding part and a decoding part; the encoding portion includes three reduction modules; the decoding part comprises three upsampling modules and four dense expansion modules, and each upsampling module is surrounded by a dense expansion module.
CN202110126100.9A 2021-01-29 2021-01-29 Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning Pending CN112785581A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110126100.9A CN112785581A (en) 2021-01-29 2021-01-29 Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110126100.9A CN112785581A (en) 2021-01-29 2021-01-29 Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning

Publications (1)

Publication Number Publication Date
CN112785581A true CN112785581A (en) 2021-05-11

Family

ID=75759788

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110126100.9A Pending CN112785581A (en) 2021-01-29 2021-01-29 Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning

Country Status (1)

Country Link
CN (1) CN112785581A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113592939A (en) * 2021-07-20 2021-11-02 燕山大学 Deep learning method for judging size of narrow blood vessel based on coronary angiography image
CN113674291A (en) * 2021-08-16 2021-11-19 北京理工大学 Full-type aortic dissection real-false lumen image segmentation method and system
CN114494482A (en) * 2021-12-24 2022-05-13 中国人民解放军总医院第一医学中心 Method for generating CT blood vessel imaging based on flat scanning CT
CN114943682A (en) * 2022-02-25 2022-08-26 清华大学 Method and device for detecting anatomical key points in three-dimensional angiography image

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106408562A (en) * 2016-09-22 2017-02-15 华南理工大学 Fundus image retinal vessel segmentation method and system based on deep learning
CN108629784A (en) * 2018-05-08 2018-10-09 上海嘉奥信息科技发展有限公司 A kind of CT image intracranial vessel dividing methods and system based on deep learning
CN110942466A (en) * 2019-11-22 2020-03-31 北京灵医灵科技有限公司 Cerebral artery segmentation method and device based on deep learning technology
CN112164069A (en) * 2020-07-29 2021-01-01 南通大学 CT abdominal blood vessel segmentation method based on deep learning

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106408562A (en) * 2016-09-22 2017-02-15 华南理工大学 Fundus image retinal vessel segmentation method and system based on deep learning
CN108629784A (en) * 2018-05-08 2018-10-09 上海嘉奥信息科技发展有限公司 A kind of CT image intracranial vessel dividing methods and system based on deep learning
CN110942466A (en) * 2019-11-22 2020-03-31 北京灵医灵科技有限公司 Cerebral artery segmentation method and device based on deep learning technology
CN112164069A (en) * 2020-07-29 2021-01-01 南通大学 CT abdominal blood vessel segmentation method based on deep learning

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113592939A (en) * 2021-07-20 2021-11-02 燕山大学 Deep learning method for judging size of narrow blood vessel based on coronary angiography image
CN113592939B (en) * 2021-07-20 2024-02-27 燕山大学 Deep learning method for judging size of narrow blood vessel based on coronary angiography image
CN113674291A (en) * 2021-08-16 2021-11-19 北京理工大学 Full-type aortic dissection real-false lumen image segmentation method and system
CN114494482A (en) * 2021-12-24 2022-05-13 中国人民解放军总医院第一医学中心 Method for generating CT blood vessel imaging based on flat scanning CT
CN114943682A (en) * 2022-02-25 2022-08-26 清华大学 Method and device for detecting anatomical key points in three-dimensional angiography image

Similar Documents

Publication Publication Date Title
CN110942466B (en) Cerebral artery segmentation method and device based on deep learning technology
CN110992338B (en) Primary stove transfer auxiliary diagnosis system
CN112785581A (en) Training method and device for extracting and training large blood vessel CTA (computed tomography angiography) imaging based on deep learning
CN109949276B (en) Lymph node detection method for improving SegNet segmentation network
CN111369565B (en) Digital pathological image segmentation and classification method based on graph convolution network
CN112712528B (en) Intestinal tract focus segmentation method combining multi-scale U-shaped residual error encoder and integral reverse attention mechanism
CN112446892A (en) Cell nucleus segmentation method based on attention learning
CN112991346B (en) Training method and training system for learning network for medical image analysis
CN113539402B (en) Multi-mode image automatic sketching model migration method
CN111862136A (en) Multi-modal nuclear magnetic image ischemic stroke lesion segmentation method based on convolutional neural network
CN111951288A (en) Skin cancer lesion segmentation method based on deep learning
CN113256592B (en) Training method, system and device of image feature extraction model
CN111462146A (en) Medical image multi-mode registration method based on space-time intelligent agent
CN116862877A (en) Scanning image analysis system and method based on convolutional neural network
CN114240955B (en) Semi-supervised cross-domain self-adaptive image segmentation method
CN116309571B (en) Three-dimensional cerebrovascular segmentation method and device based on semi-supervised learning
CN113744215B (en) Extraction method and device for central line of tree-shaped lumen structure in three-dimensional tomographic image
CN117934824A (en) Target region segmentation method and system for ultrasonic image and electronic equipment
CN111383222A (en) Intervertebral disc MRI image intelligent diagnosis system based on deep learning
Li et al. wUnet: A new network used for ultrasonic tongue contour extraction
CN114581668A (en) Segmentation model construction and contour recognition method and device and computer equipment
CN116385467B (en) Cerebrovascular segmentation method based on self-supervision learning and related equipment
Lee et al. Cardiac CT Image Segmentation for Deep Learning-Based Coronary Calcium Detection Using K-Means Clustering and Grabcut Algorithm.
CN116824146B (en) Small sample CT image segmentation method, system, terminal and storage medium
CN117633558A (en) Multi-excitation fusion zero-sample lesion detection method based on visual language model

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