A CT image cerebral hemorrhage auxiliary positioning system

Abstract

The present invention discloses a CT image cerebral hemorrhage auxiliary positioning system, which belongs to the field of image positioning. It includes: a data acquisition module, which is used to acquire multiple original CT images, mark the cranial and cerebral sections to which the CT images belong, and obtain a training set; a training module, which is used to train the CT section classification network using the training set; a first positioning module, which is used to acquire all CT images of the individual to be positioned, input them into the trained CT section classification network, obtain the predicted categories of cranial and cerebral sections, and sequentially represent them as the cranial and cerebral section classification sequence of the CT image; a third positioning module, which is used to divide the brain regions within each layer in the section classification sequence of the CT image using a brain region division mapping diagram, output the brain region where the cerebral hemorrhage occurs, and realize the positioning of the bleeding area. The present invention uses a residual neural network to build a CT section classification model, divides the brain regions within the layer through a mapping relationship, thereby roughly locating the cerebral hemorrhage area, plays an auxiliary diagnosis role, and shortens the waiting time for CT reports.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image positioning, and more specifically, relates to a CT image cerebral hemorrhage auxiliary positioning system. Background Art

[0002] Cerebral hemorrhage is a primary non-traumatic brain parenchymal hemorrhage. According to the bleeding part, cerebral hemorrhage can be divided into seven categories: basal ganglia hemorrhage, thalamic hemorrhage, lobar hemorrhage, brainstem hemorrhage, cerebellar hemorrhage and intraventricular hemorrhage. Different treatment methods are adopted for cerebral hemorrhage in different parts. Since the bleeding pressure of the part with high pressure flows to the part with low pressure, the cerebral hemorrhage will spread, which will affect the normal operation of the relevant areas of the brain to a certain extent. Effective measures should be taken as soon as possible, such as the use of dehydrators to prevent brain edema, and surgical removal of hematoma treatment when necessary. At present, CT is often used to screen cerebral hemorrhage in clinical practice. It has the advantages of no innovation, high density resolution, and contrast agent enhanced scanning, thereby improving the image with clear tissue relationship. However, CT has poor resolution for soft tissues, and patients have to wait for a long time for CT reports. It is not suitable for acute diseases related to cerebral hemorrhage. At the same time, except for special circumstances (such as emergency requirements), the conventional CT report does not include the calculation of the amount of bleeding in cerebral hemorrhage, and clinicians need to use the Tada formula for rough calculation.

[0003] With the development of machine learning technology in the field of medical image processing, different researchers have used this technology to establish image-assisted analysis tools for different diseases, using the powerful computing power of computers to conduct in-depth analysis of images, and assisting doctors to quickly acquire knowledge from images. Shanghai United Imaging Intelligent Medical Technology Co., Ltd. proposed a method for analyzing cerebral hemorrhage, the main idea of ​​which is: input the image to be analyzed into a convolutional neural network for feature extraction, input the feature map into the region of interest extraction network, obtain the detection frame containing the hemorrhage area, input the detection frame feature map into the classification network and the detection frame regression network, and obtain the classification result of the hemorrhage area and the location of the hemorrhage area respectively.

[0004] However, this method has the following disadvantages: it fails to take into account the difference in bleeding areas between different cranial and brain sections; it only obtains a detection box for the bleeding area, but does not obtain the pixel-level bleeding shape; the amount of bleeding has a crucial impact on the patient's condition, but this method fails to calculate the amount of bleeding. Summary of the invention

[0005] In view of the defects of the prior art and the need for improvement, the present invention provides a CT image cerebral hemorrhage auxiliary positioning system, which aims to use a deep residual neural network to build a model for cranial CT tomographic classification, and then divide the brain areas within the layer through mapping relationships, so as to roughly locate the area of ​​cerebral hemorrhage, which plays a role in auxiliary diagnosis and can "shorten" the waiting time for CT report to within 20 minutes. At the same time, it adds extra information not included in the CT report (such as the accurate calculation value of the amount of bleeding), which helps to improve the efficiency of CT image diagnosis and treatment of patients with cerebral hemorrhage.

[0006] To achieve the above object, according to one aspect of the present invention, a CT image cerebral hemorrhage auxiliary positioning system is provided, the system comprising:

[0007] A data acquisition module is used to acquire multiple original CT images and mark the brain slice to which each CT image belongs to obtain a training set;

[0008] A training module, used for training a brain CT section classification network using the training set;

[0009] The first positioning module is used to obtain all CT images of the individual to be positioned, input them into the trained brain CT tomography classification network, obtain the brain tomography prediction category to which each CT image belongs, and express them in sequence as the brain tomography classification sequence of the CT image;

[0010] The third positioning module is used to use the brain area division mapping map within the cranial and cerebral tomography of the CT atlas to divide the brain areas within each layer in the cranial and cerebral tomography classification sequence of the CT image, and output the brain area where the cerebral hemorrhage occurs to achieve the positioning of the bleeding area.

[0011] Preferably, the system further comprises:

[0012] The preprocessing module of the training classification network is used to perform image enhancement on the multiple original brain CT images respectively. The image enhancement methods used include: random jittering with grayscale values ​​in the range of 0.8-0.85, contrast jittering in the range of 0.5-0.9, adding random Gaussian noise, Poisson noise, salt and pepper noise, and random bright spot superposition.

[0013] Beneficial effects: The present invention performs random jittering of the grayscale value of the brain CT image within the range of 0.8-0.85 and contrast jittering within the range of 0.5-0.9 on the brain CT image, so that the grayscale value range of the brain CT image is more consistent with the actual situation, thereby ensuring the prediction effect of the residual classification network model; the present invention simulates the random noise generated by various environmental factors during the image shooting process by adding random Gaussian noise, Poisson noise, and salt and pepper noise to the brain CT image to simulate the actual situation; the present invention simulates the highlight caused by the reflection of the CT film during CT image shooting by randomly superimposing bright spots on the brain CT image.

[0014] Preferably, the training module uses a residual network as the classification network structure, uses a cross entropy loss as a loss function, and uses Adam as an optimization algorithm.

[0015] Beneficial effects: The present invention utilizes a residual network to achieve a coarse classification of cranial and brain faults. On the one hand, due to the skip-layer connection method in the residual module, the gradient information can be effectively propagated between multiple neural network layers, alleviating the problem of gradient vanishing caused by increasing the depth in a deep neural network; on the other hand, since the skip-layer connection can solve the gradient vanishing problem caused by the depth of the network, the accuracy can be improved by increasing a considerable depth.

[0016] Preferably, the first positioning module refers to the CT atlas and sets the classification categories of cranial CT sections to 12 categories, and the classification output result of each CT is one of the 12 categories.

[0017] Beneficial effects: According to the classification of brain CT sections in the "CT and MRI Sectional Anatomy Atlas", brain CT is divided into 12 sections in sequence, which are represented as 12 output types for the classification network; the 12 types of sections are sufficient to clearly represent the differences between different sections, and at the same time will not reduce the classification accuracy due to the excessive number of categories.

[0018] Preferably, the system further comprises:

[0019] The second positioning module is used to perform the following operations on the brain section classification sequence in sequence according to the characteristics of the real brain section sequence:

[0020] 1) Sequence denoising, removing noise points whose classification confidence is lower than the threshold;

[0021] 2) Sequence increasing: the classification sequence is processed in a non-strict increasing manner, so that the monotonicity of the entire sequence shows a non-strict increasing trend;

[0022] 3) Sequence continuity: if the difference between the sequence numbers of the previous and next elements in the sequence is 2, the sequence is adjusted so that it gradually degenerates to a state where the difference between the sequence numbers of the previous and next elements in the sequence is less than 2.

[0023] Beneficial effect: After the above sequence denoising, sequence incrementing and sequence continuity operations, the classification sequence obtained by the present invention will meet the restriction conditions a and b, and the optimized classification sequence has more practical significance.

[0024] Due to the restriction of the real sequence and the prior information of CT images, specifically, a. the sequence is not strictly increasing, that is, the classification sequence of cranial and brain sections can be allowed to be non-strictly increasing; b. the continuity of the sequence, that is, according to the actual situation, the actual CT image should have the same continuity as the section template, which is expressed in the sequence as the value of the next element minus the current element cannot be greater than 1; c. the confidence of the classification, the classification result should output the three confidences of the current layer and the adjacent layer of the classification, which will be used for the subsequent optimization of the classification sequence. Thus, the order and arrangement of the coarse classification sequence are optimized, and the influence of the classification network misclassification on the positioning result is reduced as much as possible, so that the optimized sequence meets the actual spatial sorting and is closer to the real cranial and brain CT section classification results.

[0025] Preferably, the third positioning module divides the brain regions within the layer, and the brain regions are divided into 8 types of brain regions with reference to the CT atlas, including: frontal lobe, parietal lobe, occipital lobe, temporal lobe, cerebellum, midbrain, pons and medulla oblongata.

[0026] Beneficial effects: After locating the cranial and brain sections, for each section, the classified image needs to be divided into brain regions. Use the matching method to find the center line of the skull area, correct the inclination angle of the center line, and put it in a vertical state. After the above alignment operation, each pixel position of the CT image is divided, and different brain region division domains can be obtained on the classified tomographic CT image. After the segmentation network segments the cerebral hemorrhage area, the pixel coordinate range of the segmented area can be obtained. By matching this pixel coordinate range to the brain region division domain, the brain region information where the cerebral hemorrhage occurs can be output.

[0027] Preferably, the system further comprises:

[0028] The bleeding volume estimation module is used to obtain the actual area S corresponding to each pixel in the CT image. pixel , the fixed spacing h between adjacent slices of the cranial CT image; for the segmented cranial CT image, first calculate the number of pixels n contained in the lesion area detected in the i-th image i , calculate the actual area S through the number of pixels i =S pixel ×n i ; The bleeding shapes in adjacent images are regarded as "pillars", and the bleeding volume between adjacent CT slices is calculated using the bleeding area of ​​the image Finally, all bleeding volumes are added together to obtain the patient's bleeding volume. Where k represents the number of all CT images of a patient.

[0029] Beneficial effects: Currently, clinicians use Tada's formula to roughly calculate the amount of cerebral hemorrhage in patients, and the results are often quite different from the actual situation. However, the efficiency of calculating cerebral hemorrhage in the radiology department is low, and it is not calculated unless necessary. Through the above steps, the present invention converts the amount of bleeding between adjacent layers into a "pillar" volume of a fixed height, and uses the "pillar" volume calculation formula to calculate the amount of bleeding. Since the process of traversing the number of pixels only needs to be performed once for each binary image, and the calculation accuracy is at the pixel level, the calculation of the amount of cerebral hemorrhage can be quickly realized, and the calculation accuracy is close to that of the radiology department.

[0030] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0031] In view of the long waiting time (2h) for CT reports in hospitals, the present invention realizes cranial CT section classification through residual classification network, optimizes the classification sequence on this basis, and finally realizes the functions such as intra-layer brain area positioning in each layer of the optimized sequence; due to the speed and accuracy advantages of convolutional neural network in image recognition, the processing time of a CT image (size of 512*512 pixels) by the classification network is about 1 second. Therefore, only in the reasoning process of the network model, the diagnosis time for all CT images of a patient is within 1 minute. Combined with the shooting and acquisition time of the patient's CT image, the input and output time of the program, etc., the patient's consultation time after getting the CT film can be shortened to less than 10 minutes, and the positioning accuracy is relatively consistent with the judgment accuracy of the radiology department. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic diagram of the use process of a CT image cerebral hemorrhage auxiliary diagnosis and positioning system is provided for the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0034] The present invention provides a CT image cerebral hemorrhage auxiliary diagnosis and positioning system, which uses a deep residual neural network to build a cranial CT tomography classification model, and then divides the brain area in the layer through a mapping relationship, so as to roughly locate the area of ​​cerebral hemorrhage, which plays a role in auxiliary diagnosis. In the process of diagnosing patients with cerebral hemorrhage, doctors can use mobile phones or desktop software to import CT images, and realize functions such as brain area positioning of hemorrhage lesions through trained neural network models and brain area division models, so as to help doctors quickly and deeply understand the details of patients' cerebral hemorrhage and improve the level of diagnosis. At the same time, the operability of mobile phones or desktop software at any time can effectively make up for the problems such as long waiting time for CT reports and delayed interpretation of CT information.

[0035] The present invention provides a CT image cerebral hemorrhage auxiliary diagnosis and positioning system, which comprises:

[0036] The data acquisition module is used to acquire multiple original CT images and mark the brain slices to which the CT images belong.

[0037] Image preprocessing module: In order to improve the speed of the network model during the testing process, the secondary CT images need to be preprocessed during the use phase, including: image center cropping, image compression, image grayscale, etc.

[0038] Image enhancement module: In order to increase the size of the training set in the cranial CT section classification model and enhance the generalization ability of the model in practical applications, the original CT images of the training set need to be enhanced during the training phase. The data enhancement methods used include: random jittering of grayscale values ​​in the range of 0.8-0.85, contrast jittering in the range of 0.5-0.9 (to simulate the jitter blur that may occur during shooting), adding random Gaussian noise, Poisson noise, and salt and pepper noise (to simulate noise interference during shooting), and random bright spot superposition (to simulate reflections during shooting).

[0039] The model mainly consists of three parts: a rough classification model of brain CT sections (built using a residual network), optimization of brain CT section classification sequences, and an intra-section partitioning module.

[0040] The classification network structure used is a residual network, specifically ResNeXt101. Compared with ResNet, this network has significant effects on the segmentation problem of medical images and has a greater improvement in prediction effect.

[0041] The network consists of a series of residual blocks, and a single residual block can be expressed as: l+1 =x l +F(x l ,W l ). The residual block is divided into two parts: the direct mapping part x l and the residual part F(xl ,W l ), the residual part is generally composed of two or three convolution operations; loss back propagation: the loss function is cross entropy loss, the loss between the predicted category vector and the labeled category is calculated, and the loss back propagation is used to update the network weight parameters. Adam is used as the optimization algorithm, Batch_size is set to 2, and the learning rate is set to 0.001. Network output: For a single CT image, the output is the predicted category; for a patient's CT image sequence, the output is a classified sequence.

[0042] Model training module: Define the rough classification model of cranial CT tomography using shortcut connections, and perform end-to-end learning and stochastic gradient descent training through back propagation.

[0043] The first positioning module is used to obtain all CT images of the individual to be positioned, input them into the trained brain CT tomography classification network, obtain the brain tomography prediction category to which each CT image belongs, and express them in sequence as the brain tomography classification sequence of the CT image;

[0044] Furthermore, the brain CT tomographic rough classification utilizes a deep residual neural network to classify the patient CT image set into 12 types of tomographic images of sequential standard templates, thereby obtaining a rough classification result.

[0045] The second positioning module is used to perform the following operations on the brain section classification sequence in sequence according to the characteristics of the real brain section sequence:

[0046] 1) Sequence denoising, removing noise points whose classification confidence is lower than the threshold;

[0047] 2) Sequence increasing: the classification sequence is processed in a non-strict increasing manner, so that the monotonicity of the entire sequence shows a non-strict increasing trend;

[0048] 3) Sequence continuity: if the difference between the sequence numbers of the previous and next elements in the sequence is 2, the sequence is adjusted so that it gradually degenerates to a state where the difference between the sequence numbers of the previous and next elements in the sequence is less than 2.

[0049] The main purpose of this step of brain CT section classification sequence optimization is to use some prior information (confidence of network classification results) and constraints to optimize the classification sequence based on image-level coarse classification, thereby solving the problem of too small a gap between adjacent brain CT section images and too large a difference between brain CT images of different people.

[0050] The third positioning module is used to perform intra-layer brain region division processing on each layer in the cranial and cerebral tomography classification sequence of the CT image, divide the brain regions of each layer of the CT image using the brain region division mapping diagram within the cranial and cerebral tomography of the CT atlas, and output the brain region where the cerebral hemorrhage occurs to locate the bleeding area.

[0051] The intra-slice partitioning module divides the regions within the slices of the optimized CT slice classification sequence: the classified images are matched with the corresponding templates, and the segmented cerebral hemorrhage areas are located to one or several brain regions of the segmentation results, thus obtaining a diagnostic result with positioning information.

[0052] The reference CT atlas will be divided into 8 brain regions, including: frontal lobe, parietal lobe, occipital lobe, temporal lobe, cerebellum, midbrain, pons and medulla oblongata.

[0053] The CT imaging cerebral hemorrhage auxiliary diagnosis system is further configured as: a mobile phone terminal or a desktop terminal of a doctor in a clinic is provided with a packaged auxiliary diagnosis program. The mobile phone terminal system can be in the form of a WeChat applet.

[0054] As the front-end interactive part of the medical intelligent diagnosis service, the mobile or desktop terminal is responsible for forming an interface interaction with the user and collecting the input information required for the model service, such as: the patient's basic information, the patient's CT imaging data, and past medical history, and the patient's chief complaint text, which are submitted to the medical intelligent diagnosis service. After being processed by the back-end service, they are converted into input parameters that can ultimately be used by the medical intelligent diagnosis model, and then the model's inference action is started to generate the model output results, which are then returned to the mini program front-end interface for display to the user.

[0055] The system also includes:

[0056] The text information input module supports manual selection or filling of patient clinical information, chief complaint and other information.

[0057] The image input module supports doctors to take stable and high-precision photos of CT images under good lighting conditions or directly import the patient's CT image desktop file and transfer it to the target detection module.

[0058] The target detection module performs segmentation and positioning processing, and then imports the positioning results of multiple CT images detected into the diagnosis analysis module. The disease feature information contained in it comes from the CT image set of a single patient at different periods (if any).

[0059] The diagnostic analysis module outputs the patient's bleeding location information, whether or not the bleeding has spread, and other conclusions based on the location information (if any) included in the CT images of the same patient at different times, and returns the above auxiliary diagnostic information to the mobile phone or desktop, and the display module displays all the above conclusions on the mobile phone. Its image database is structured after knowledge input by multiple professional neurosurgeons.

[0060] The bleeding volume estimation module is used to obtain the actual area S corresponding to each pixel in the CT image. pixel, the fixed spacing h between adjacent slices of the brain CT image (these two are prior knowledge); for the segmented brain CT image, first calculate the number of pixels n contained in the lesion area detected in the i-th image i , calculate the actual area S through the number of pixels i =S pixel ×n i ; The bleeding shapes in adjacent images are regarded as "pillars", and the bleeding volume between adjacent CT slices is calculated using the bleeding area of ​​the image Finally, all bleeding volumes are added together to obtain the patient's bleeding volume.

[0061] Doctors launch the mini program on their mobile or desktop to detect CT image information of cerebral hemorrhage and import multiple CT images of patients. After CT image preprocessing is completed, redundant information in CT image data is eliminated, and image size is compressed and center cropped to the same size as the input image size when training the model, and then grayscale and image enhancement are performed. After launching the mobile or desktop version, the patient's basic information, clinical data, and chief complaint are interactively input, and the location of the patient's cerebral hemorrhage lesion, whether or not bleeding has spread, and other information are output based on the designed model.

[0062] The interface should include a section for inputting or selecting the patient's personal information, medical history and chief complaint; the interface should also include a section for directly taking pictures or retrieving the patient's CT images from local files.

[0063] Based on the interactive area selection on the mini-program page, you can manually zoom in on different bleeding lesions on the brain CT for further observation.

[0064] like Figure 1 As shown, the use process of the CT image cerebral hemorrhage auxiliary diagnosis system described in the present invention includes: CT image shooting and input, CT image processing, text information input, core module training and application, and diagnosis conclusion output process. Taking the diagnosis and treatment of patients with basal ganglia hemorrhage as an example, the process of CT image shooting and input, CT image processing, text information input, core module training and application, and diagnosis conclusion output is:

[0065] First, use the WeChat applet window to capture the original CT image and perform preliminary processing, mainly including: image center cropping, image compression, image grayscale, etc.

[0066] Next, the preprocessed images are placed into the trained deep residual network to achieve tomographic classification: the deep residual network is used to classify the patient's CT image set into 12 types of tomographic images arranged in order, and a rough classification sequence of cranial CT tomographic images is obtained.

[0067] Next, in order to achieve a more reasonable fault classification sequence, based on the rough classification sequence, with the help of prior information and constraints, the classification sequence needs to be optimized to solve the problem of too small gaps between adjacent brain CT tomographic images and too large differences between brain CT images of different people. The processing processes are: sequence denoising, sequence non-strict increment, and sequence continuity.

[0068] Finally, the optimized cranial and cerebral tomography classification sequence is processed with intra-layer brain region division, and the brain region of each layer of CT image is divided using the brain region division mapping diagram within the cranial and cerebral tomography of the CT atlas; and the brain region where the cerebral hemorrhage occurs is output to realize the positioning function of the bleeding area and obtain a diagnostic result with positioning information.

[0069] In the process of building the neural network model in the system, especially when annotating CT images, manual intervention can be carried out in real time, specifically inviting expert doctors to participate in the work together to improve the accuracy and guiding value of the model. In practical applications, the present invention can simultaneously conduct in-depth information mining on the disease of cerebral hemorrhage, form a more complete diagnosis conclusion, and enable doctors to obtain all medical information of CT images in a short time.

[0070] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A CT image cerebral hemorrhage auxiliary positioning system, characterized in that: The system includes: A data acquisition module is used to acquire multiple original CT images and mark the brain slice to which each CT image belongs to obtain a training set; A training module, used for training a brain CT section classification network using the training set; The first positioning module is used to obtain all CT images of the patient to be positioned, input them into the trained brain CT tomography classification network, obtain the brain tomography prediction category to which each CT image belongs, and express them in sequence as the brain tomography classification sequence of the CT image; The third positioning module is used to divide the brain regions in each layer of the brain section classification sequence of the CT image by using the brain section division map in the brain section of the CT atlas, and output the brain region where the cerebral hemorrhage occurs, so as to realize the positioning of the hemorrhage area; The third positioning module, intra-layer brain region division, refers to the CT atlas to divide the brain into 8 regions, including: frontal lobe, parietal lobe, occipital lobe, temporal lobe, cerebellum, midbrain, pons and medulla oblongata; The bleeding volume estimation module is used to obtain the actual area S corresponding to each pixel in the CT image. pixel , the fixed spacing h between adjacent slices of the cranial CT image; for the segmented cranial CT image, first calculate the number of pixels n contained in the lesion area detected in the i-th image i , calculate the actual area S through the number of pixels i =S pixel ×n i ; The bleeding shapes in adjacent images are regarded as "pillars", and the bleeding volume between adjacent CT slices is calculated using the bleeding area of ​​the image Finally, all bleeding volumes are added together to obtain the patient's bleeding volume. Where k represents the number of all CT images of a patient.

2. The system according to claim 1, characterized in that The system also includes: The preprocessing module of the training classification network is used to perform image enhancement on the multiple original CT images respectively. The image enhancement methods used include: random jittering with grayscale values ​​in the range of 0.8-0.85, contrast jittering in the range of 0.5-0.9, adding random Gaussian noise, Poisson noise, salt and pepper noise, and random bright spot superposition.

3. The system according to claim 1, characterized in that The training module uses a residual network as the classification network structure, uses a cross entropy loss as a loss function, and uses Adam as an optimization algorithm.

4. The system according to claim 1, characterized in that The first positioning module refers to the CT atlas and sets the classification categories of cranial CT tomography to 12 categories, and the classification output result of each CT is one of the 12 categories of tomography.

5. The system according to claim 1, characterized in that The system also includes: The second positioning module is used to perform the following operations on the brain section classification sequence in sequence according to the characteristics of the real brain section sequence: 1) Sequence denoising, removing noise points whose classification confidence is lower than the threshold; 2) Sequence increasing: the classification sequence is processed in a non-strict increasing manner, so that the monotonicity of the entire sequence shows a non-strict increasing trend; 3) Sequence continuity: if the difference between the sequence numbers of the previous and next elements in the sequence is 2, the sequence is adjusted so that it gradually degenerates to a state where the difference between the sequence numbers of the previous and next elements in the sequence is less than 2.

Images