阅读说明：本技术 一种ct影像脑出血辅助定位系统 (CT image cerebral hemorrhage auxiliary positioning system ) 是由 骆汉宾 张佳乐 谭毅华 闫培 郑自鹏 于 2021-09-26 设计创作，主要内容包括：本发明公开一种CT影像脑出血辅助定位系统,属于图像定位领域。包括：数据获取模块,用于获取多张原始CT图像,标记出CT图像所属颅脑断层,得到训练集；训练模块,用于利用训练集对CT断层分类网络进行训练；第一定位模块,用于获取待定位个体的全部CT图像,输入至训练好的CT断层分类网络,得到颅脑断层预测类别,按序表示为CT图像的颅脑断层分类序列；第三定位模块,用于利用脑区划分映射图对CT图像的断层分类序列中的每一层进行层内脑区划分,将发生脑出血所在的脑区部位输出,实现出血区域定位。本发明利用残差神经网络搭建CT断层分类模型,通过映射关系对层内脑区划分,从而对脑出血区域进行大致定位,起到辅助诊断作用,将CT报告等待时间缩短。(The invention discloses a CT image cerebral hemorrhage auxiliary positioning system, and belongs to the field of image positioning. The method comprises the following steps: the data acquisition module is used for acquiring a plurality of original CT images and marking the craniocerebral fault to which the CT images belong to obtain a training set; the training module is used for training the CT fault classification network by utilizing a training set; the first positioning module is used for acquiring all CT images of an individual to be positioned, inputting the CT images into a trained CT fault classification network to obtain a prediction type of the craniocerebral fault, and sequentially representing the prediction type of the craniocerebral fault as a craniocerebral fault classification sequence of the CT images; and the third positioning module is used for carrying out in-layer brain area division on each layer in the fault classification sequence of the CT image by utilizing the brain area division mapping image, outputting the brain area part where the cerebral hemorrhage occurs, and realizing the positioning of the hemorrhage area. According to the method, the CT fault classification model is built by utilizing the residual error neural network, and the in-layer brain area is divided through the mapping relation, so that the cerebral hemorrhage area is roughly positioned, the auxiliary diagnosis effect is achieved, and the waiting time of the CT report is shortened.)

1. An auxiliary positioning system for cerebral hemorrhage based on CT images, which is characterized in that the system comprises:

the data acquisition module is used for acquiring a plurality of original CT images and marking the craniocerebral fault to which each CT image belongs to obtain a training set;

the training module is used for training the craniocerebral CT fault classification network by utilizing the training set;

the first positioning module is used for acquiring all CT images of an individual to be positioned, inputting the CT images into a trained craniocerebral CT fault classification network to obtain the craniocerebral fault prediction category to which each CT image belongs, and sequentially representing the craniocerebral fault prediction categories as the craniocerebral fault classification sequences of the CT images;

and the third positioning module is used for carrying out in-layer brain area division on each layer in the brain fault classification sequence of the CT image by utilizing a brain area division mapping image in the brain fault layer of the CT map, and outputting the brain area part where cerebral hemorrhage occurs, so that the positioning of a hemorrhage area is realized.

2. The system of claim 1, further comprising:

a preprocessing module of the training classification network, configured to perform image enhancement on the multiple original brain CT images, respectively, where the image enhancement method includes: random dithering with gray value in the range of 0.8-0.85, contrast dithering in the range of 0.5-0.9, addition of random Gaussian noise, Poisson noise, salt and pepper noise, and superposition of random bright spots.

3. The system of claim 1, wherein the training module employs a residual network as a classification network structure, cross entropy loss as a loss function, and Adam as an optimization algorithm.

4. The system of claim 1, wherein the first positioning module refers to a CT atlas to set the classification categories of the CT skull slices as 12 classes, and the classification output result of each CT is one of the 12 classes of slices.

5. The system of claim 1, further comprising:

the second positioning module is used for sequentially carrying out the following steps on the brain fault classification sequence according to the characteristics of the real brain fault sequence:

1) denoising the sequence, namely removing noise points with classification confidence degrees lower than a threshold value;

2) increasing the sequence, namely performing non-strict increasing treatment on the classified sequence to ensure that the monotonicity of the whole sequence presents a non-strict increasing trend;

3) and (4) sequences are serialized, if the difference between the sequence numbers of the elements before and after the sequences is 2, the sequences are gradually degenerated to be less than 2 by sequence adjustment.

6. The system of claim 1, wherein the third localization module, intralayer brain region partitioning, reference CT atlas will be partitioned into 8 brain regions, comprising: frontal lobe, parietal lobe, occipital lobe, temporal lobe, cerebellum, midbrain, pons and medulla oblongata.

7. The system of any one of claims 1 to 6, further comprising:

a bleeding amount estimation module for obtaining the actual area S corresponding to each pixel in the CT image_pixelThe fixed distance h between adjacent slices of the craniocerebral CT image; for the segmented brain CT image, the number n of pixels included in the detected lesion region in the ith image is calculated_iThe size S of the actual area is obtained by the number of pixels_i＝S_pixel×n_i(ii) a The bleeding shape in the adjacent images is regarded as a pillar, and the bleeding volume between the adjacent CT slices is calculated by using the bleeding volume area of the imagesFinally, all the bleeding volumes are added to obtain the bleeding volume of the patientWhere k represents the number of total CT images for a patient.

Technical Field

The invention belongs to the technical field of medical image positioning, and particularly relates to a CT image cerebral hemorrhage auxiliary positioning system.

Background

Cerebral hemorrhage is primary non-traumatic brain parenchymal hemorrhage, and can be divided into 7 types, such as basal ganglia hemorrhage, thalamus hemorrhage, cerebral lobe hemorrhage, brain stem hemorrhage, cerebellar hemorrhage, and ventricular hemorrhage, according to hemorrhage parts. The treatment means adopted for the cerebral hemorrhage of different parts are different. Cerebral hemorrhage may spread due to the flow of hemorrhage from a site with high hemorrhage pressure to a site with low hemorrhage pressure, and may affect the normal operation of the relevant area of the brain to some extent. Effective measures such as the application of a dehydrator and the like to prevent cerebral edema and the surgical removal of hematoma treatment when necessary need to be taken as soon as possible. At present, CT is adopted to screen cerebral hemorrhage clinically, and the CT has the advantages of no innovation, high density resolution, capability of enhancing scanning by using a contrast agent and the like, so that clear images of tissue relations are improved. However, the resolution of CT to soft tissue is poor, and the patient has a long waiting time for CT report, which is not suitable for acute diseases related to cerebral hemorrhage. Meanwhile, except for special conditions (such as emergency treatment), the conventional CT report does not include the calculation of the amount of cerebral hemorrhage and requires a clinician to roughly calculate by using a multidata formula.

With the development of the machine learning technology in the field of medical image processing, different researchers use the technology to establish image auxiliary analysis tools for different diseases, and use the powerful computing power of a computer to carry out depth analysis on images so as to assist doctors to quickly acquire knowledge from the images. The Shanghai Lianying Intelligent medical science and technology Limited company provides a cerebral hemorrhage analysis method, which has the main ideas as follows: inputting the image to be analyzed into a convolutional neural network for feature extraction, inputting the feature map into an interested area extraction network to obtain a detection frame containing a bleeding area, and inputting the feature map of the detection frame into a classification network and a detection frame regression network to respectively obtain a classification result of the bleeding area and the position of the bleeding area.

However, this method has the following disadvantages: this approach fails to take into account the difference in bleeding areas from different cranial faults; the bleeding area obtained by the method only has a detection frame, and does not obtain a bleeding shape of a pixel level; the amount of bleeding has a crucial influence on the patient's condition, but the method fails to calculate the outcome of the bleeding.

Disclosure of Invention

Aiming at the defects and improvement requirements of the prior art, the invention provides a CT image cerebral hemorrhage auxiliary positioning system, which aims to build a craniocerebral CT fault classification model by using a depth residual error neural network and divide brain areas in the layer by a mapping relation, so that the cerebral hemorrhage area is roughly positioned, the auxiliary diagnosis effect is realized, the waiting time of a CT report can be shortened to 20 minutes, meanwhile, extra information (such as an accurate calculated value of the amount of hemorrhage) which is not related to the CT report is increased, and the CT image diagnosis and treatment efficiency of a cerebral hemorrhage patient is improved.

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

the data acquisition module is used for acquiring a plurality of original CT images and marking the craniocerebral fault to which each CT image belongs to obtain a training set;

the training module is used for training the craniocerebral CT fault classification network by utilizing the training set;

the first positioning module is used for acquiring all CT images of an individual to be positioned, inputting the CT images into a trained craniocerebral CT fault classification network to obtain the craniocerebral fault prediction category to which each CT image belongs, and sequentially representing the craniocerebral fault prediction categories as the craniocerebral fault classification sequences of the CT images;

and the third positioning module is used for carrying out in-layer brain area division on each layer in the brain fault classification sequence of the CT image by utilizing a brain area division mapping image in the brain fault layer of the CT map, and outputting the brain area part where cerebral hemorrhage occurs, so that the positioning of a hemorrhage area is realized.

Preferably, the system further comprises:

a preprocessing module of the training classification network, configured to perform image enhancement on the multiple original brain CT images, respectively, where the image enhancement method includes: random dithering with gray value in the range of 0.8-0.85, contrast dithering in the range of 0.5-0.9, addition of random Gaussian noise, Poisson noise, salt and pepper noise, and superposition of random bright spots.

Has the advantages that: the method ensures that the gray value range of the brain CT image is more consistent with the real situation by performing random dithering of the gray value of the brain CT image within the range of 0.8-0.85 and contrast dithering of the brain CT image within the range of 0.5-0.9, thereby ensuring the prediction effect of a residual classification network model; random Gaussian noise, Poisson noise and salt and pepper noise added in the brain CT image under the actual condition are simulated, so that random noise generated by various environmental factors in the shooting process of the image is simulated; according to the invention, the highlight condition caused by the reflection of the CT film in the CT image shooting is simulated by carrying out random bright spot superposition on the brain CT image.

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

Has the advantages that: according to the invention, coarse classification of the cranial fault is realized by using the residual error network, on one hand, gradient information can be effectively transmitted among a plurality of neural network layers due to a layer jump connection mode in the residual error module, and the problem of gradient disappearance caused by depth increase in the deep neural network is relieved; on the other hand, the layer jump connection can solve the problem of gradient disappearance caused by the network depth, so that the accuracy can be improved by increasing the equivalent depth.

Preferably, the first positioning module refers to a CT atlas, sets classification categories of the craniocerebral CT faults as 12 categories, and the classification output result of each CT is one of the 12 categories of faults.

Has the advantages that: according to the classification of CT faults in CT and MRI (computed tomography and magnetic resonance imaging) tomography, the CT faults of the brain are divided into 12 faults in sequence, and the 12 output types are expressed for a classification network; the 12 types of faults are enough to clearly represent the difference between different faults, and the classification accuracy rate is not reduced due to the excessive number of classes.

Preferably, the system further comprises:

the second positioning module is used for sequentially carrying out the following steps on the brain fault classification sequence according to the characteristics of the real brain fault sequence:

1) denoising the sequence, namely removing noise points with classification confidence degrees lower than a threshold value;

2) increasing the sequence, namely performing non-strict increasing treatment on the classified sequence to ensure that the monotonicity of the whole sequence presents a non-strict increasing trend;

3) and (4) sequences are serialized, if the difference between the sequence numbers of the elements before and after the sequences is 2, the sequences are gradually degenerated to be less than 2 by sequence adjustment.

Has the advantages that: after the sequence denoising, sequence increasing and sequence continuous operation are carried out, the obtained classification sequence meets the limiting conditions a and b, and the optimized classification sequence has practical significance.

Due to the fact that the limiting conditions of the real sequence and the prior information of the CT image are utilized, specifically, a, the sequence is not strictly increased, namely, the classification sequence of the craniocerebral fault is allowed to be not strictly increased; b. sequence continuity, that is, according to the actual situation, the actual CT image should have continuity similar to the fault template, and the value represented by subtracting the current element from the next element in the sequence cannot be greater than 1; c. and the classified confidence degrees and the classification result should output three confidence degrees of the classified current layer and the adjacent layer together for optimizing the classification sequence subsequently. Therefore, the sequence and the arrangement mode of the coarse classification sequence are optimized, the influence of classification network misclassification on the positioning result is reduced as much as possible, the optimized sequence meets the actual spatial sequencing, and the optimal sequence is closer to the real brain CT fault classification result.

Preferably, the third localization module, the layer inner brain region is divided, and the reference CT atlas is divided into 8 kinds of brain regions, including: frontal lobe, parietal lobe, occipital lobe, temporal lobe, cerebellum, midbrain, pons and medulla oblongata.

Has the advantages that: after the cranial slice is located, for each slice, brain region division needs to be performed on the classified image. And finding the center line of the skull region by using a matching method, correcting the inclination angle of the center line, and keeping the center line in a vertical state. After the alignment operation, different brain regions can be obtained on the classified tomographic CT images by dividing each pixel position of the CT images. After the cerebral hemorrhage area is divided by the dividing network, the pixel coordinate range of the divided area can be obtained, and the pixel coordinate range corresponds to the cerebral hemorrhage area, so that the cerebral hemorrhage information can be output.

Preferably, the system further comprises:

a bleeding amount estimation module for obtaining the actual area S corresponding to each pixel in the CT image_pixelThe fixed distance h between adjacent slices of the craniocerebral CT image; for the segmented brain CT image, the number n of pixels included in the detected lesion region in the ith image is calculated_iThe size S of the actual area is obtained by the number of pixels_i＝S_pixel×n_i(ii) a The bleeding shape in the adjacent images is regarded as a pillar, and the bleeding volume between the adjacent CT slices is calculated by using the bleeding volume area of the imagesFinally, all the bleeding volumes are added to obtain the bleeding volume of the patientWhere k represents the number of total CT images for a patient.

Has the advantages that: at present, clinicians roughly calculate the cerebral hemorrhage volume of patients by using a multidata formula, and the result is often greatly different from the actual situation. However, the radiology department is inefficient in calculating cerebral hemorrhage and does not necessarily calculate it. According to the invention, through the steps, the bleeding amount between adjacent layers is converted into the volume of the column with fixed height, and the bleeding amount is obtained by utilizing a column volume calculation formula. The calculation of the cerebral hemorrhage amount can be quickly realized because the process of traversing and calculating the pixel number is only needed once for each binary image and the calculation precision is on the pixel level, and the calculation precision is close to the precision of a radiology department.

Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:

aiming at the problem that the current hospital CT report has long waiting time (2h), the invention realizes the classification of the craniocerebral CT fault through a residual error classification network, optimizes a classification sequence on the basis, and finally realizes the functions of positioning the inner brain region and the like in each layer of the optimization sequence; because of the speed and precision advantages of the convolutional neural network in image recognition, the processing time of the classification network on one CT image (with the size of 512 x 512 pixels) is about 1 second, so that the diagnosis time of all the CT images of one patient is within 1 minute in the inference process of the network model, the shooting and acquisition time of the CT images of the patient, the input and output time of a program and the like are added, the diagnosis time of the patient after the patient is taken to the CT image can be shortened to be within 10 minutes, and the positioning precision is more consistent with the radiology judgment precision.

Drawings

Fig. 1 is a schematic view of a usage process of a positioning system for auxiliary diagnosis of cerebral hemorrhage using CT images according to the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

The invention provides a CT image cerebral hemorrhage auxiliary diagnosis positioning system, which utilizes a depth residual error neural network to build a craniocerebral CT fault classification model, and divides brain areas in layers through a mapping relation, thereby roughly positioning the cerebral hemorrhage areas and playing a role in auxiliary diagnosis. In the process of diagnosing the cerebral hemorrhage patient, a doctor can utilize a mobile phone or desktop end software to introduce a CT image, and the functions of positioning a cerebral region of a hemorrhage focus and the like are realized through a trained neural network model and a cerebral region division model, so that the doctor is helped to quickly and deeply know the cerebral hemorrhage details of the patient, and the diagnosis level is improved. Meanwhile, the mobile phone or desktop software has the advantages of being capable of effectively solving the problems of long waiting time of CT reports, delayed interpretation of CT information and the like due to the operability at any time and the like.

The invention provides a CT image cerebral hemorrhage auxiliary diagnosis positioning system, which comprises:

and the data acquisition module is used for acquiring a plurality of original CT images and marking the craniocerebral fault to which the CT images belong.

An image preprocessing module: in order to increase the speed of the network model in the testing process, in the using stage, the CT image obtained by the secondary shooting needs to be preprocessed, which includes: image center cropping, image compression, image graying, and the like.

An image enhancement module: in order to increase the size of a training set in a brain CT fault classification model and enhance the generalization capability of the model in practical application, necessary data enhancement needs to be carried out on an original CT image of the training set in a training stage, and the adopted data enhancement method comprises the following steps: the method comprises the following steps: random dithering with gray scale values in the range of 0.8-0.85, contrast dithering in the range of 0.5-0.9 (dithering blur possibly generated in analog shooting), addition of random gaussian noise, poisson noise, salt and pepper noise (noise interference in analog shooting), and random bright spot superposition (used for simulating reflection in shooting).

The model mainly comprises a craniocerebral CT fault rough classification model (built by using a residual error network), craniocerebral CT fault classification sequence optimization and intra-fault partition modules.

The adopted classification network structure is a residual error network, particularly ResNeXt101, and compared with ResNet, the network has a remarkable effect on the segmentation problem of medical images, and the prediction effect is greatly improved.

The network is made up of a series of residual blocks, and a single residual block can be represented as: x is the number of_l+1＝x_l+F(x_l,W_l). The residual block is divided into two parts: direct mapping part x_lAnd residual part F (x)_l,W_l) The residual part is generally composed of two or three convolution operations; loss reverse transmission: the loss function is cross entropy loss, calculates the loss between the prediction class vector and the label class, and uses the back propagation of the loss to update the network weight parameters. Using Adam as the optimization algorithm, Batch _ size is set to 2 and the learning rate is set to 0.001. And (3) network output: outputting a single CT image as a prediction category; for a sequence of CT images of a patient, the output is a classified sequence.

A model training module: defining a craniocerebral CT fault rough classification model, adopting near-path connection, and carrying out end-to-end learning and random gradient descent training through back propagation.

The first positioning module is used for acquiring all CT images of an individual to be positioned, inputting the CT images into a trained craniocerebral CT fault classification network to obtain the craniocerebral fault prediction category to which each CT image belongs, and sequentially representing the craniocerebral fault prediction categories as the craniocerebral fault classification sequences of the CT images;

further, the coarse classification of the brain CT fault utilizes a depth residual error neural network to classify the CT image set of the patient into 12 types of fault images of the sequential standard template, namely, a coarse classification result is obtained.

The second positioning module is used for sequentially carrying out the following steps on the brain fault classification sequence according to the characteristics of the real brain fault sequence:

1) denoising the sequence, namely removing noise points with classification confidence degrees lower than a threshold value;

2) increasing the sequence, namely performing non-strict increasing treatment on the classified sequence to ensure that the monotonicity of the whole sequence presents a non-strict increasing trend;

3) and (4) sequences are serialized, if the difference between the sequence numbers of the elements before and after the sequences is 2, the sequences are gradually degenerated to be less than 2 by sequence adjustment.

The main purpose of the step of optimizing the CT (computed tomography) fault classification sequence is to optimize the classified sequence on the basis of image-level coarse classification by using some prior information (confidence of network classification results) and limiting conditions, so that the problems of too small difference between adjacent CT fault images and too large difference between the CT images of different people are solved.

And the third positioning module is used for carrying out inner brain area division processing on each layer in the craniocerebral fault classification sequence of the CT image, carrying out brain area division on each layer of the CT image by utilizing a brain area division mapping map in the craniocerebral fault of the CT map, and outputting the brain area position where cerebral hemorrhage occurs, thereby realizing the positioning of the hemorrhage area.

And the intra-fault partition module performs intra-layer region partition on the optimized round CT fault classification sequence: and matching the classified images with corresponding templates, and positioning the segmented cerebral hemorrhage regions to one or more brain regions of the division result to obtain a diagnosis result with positioning information.

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.

The CT image cerebral hemorrhage auxiliary diagnosis system is further provided with: and a mobile phone end or a desktop end of a doctor in a consulting room is provided with an encapsulated auxiliary diagnostic program. The mobile phone end system can be in the form of a WeChat applet.

The mobile phone end or the desktop end is used as a front-end interaction part of the medical intelligent diagnosis service and is responsible for forming interface interaction with a user and collecting input information required by the model service, such as: basic information of the patient, CT image data of the patient, the past medical history and a complaint text of the patient are submitted to medical intelligent diagnosis service, are converted into input parameters which can be finally used by a medical intelligent diagnosis model through back-end service processing, immediately start an inference action of the model to generate a model output result, and then return the result to an applet front-end interface to be displayed to a user.

The system further comprises:

and the text information input module supports manual selection or filling of information such as clinical information, chief complaints and the like of patients.

The image input module supports that a doctor can stably shoot the CT image under a good illumination condition and has higher precision or directly import the CT image desktop file of the patient and transmit the CT image desktop file to the target detection module.

The target detection module carries out segmentation positioning processing, and then the positioning results of a plurality of detected CT images are imported into the diagnosis analysis module. Which contains information on disease characteristics from CT image sets of individual patients at different times (if any).

The diagnosis analysis module outputs the patient bleeding location information, yes/no diffusion and other conclusions according to the location information (if any) included by the CT images of the same patient in different periods, the auxiliary diagnosis information is returned to the mobile phone end or the desktop end, and the display module displays all the conclusions at the mobile phone end. The image database is structured by a plurality of professional neurosurgeons after inputting knowledge.

A bleeding amount estimation module for obtaining the actual area S corresponding to each pixel in the CT image_pixelThe fixed spacing h between adjacent slices of the craniocerebral CT image (the two are priori knowledge); for the segmented brain CT image, the number n of pixels included in the detected lesion region in the ith image is calculated_iThe size S of the actual area is obtained by the number of pixels_i＝S_pixel×n_i(ii) a The bleeding shape in the adjacent images is regarded as a pillar, and the bleeding volume between the adjacent CT slices is calculated by using the bleeding volume area of the imagesFinally, all the bleeding volumes are added to obtain the bleeding volume of the patient

The doctor starts the small program to detect the cerebral hemorrhage CT image information by the mobile phone end or the desktop end, and leads in a plurality of patient CT images; and finishing CT image preprocessing, eliminating redundant information in CT image data, compressing the image size, cutting the center, and performing graying and image enhancement after inputting the size of the image which is consistent with the size of the image during model training. After the mobile phone end or the desktop end is started, the basic information, the clinical data, the chief complaint and other contents of the patient are interactively input, and the information such as the cerebral hemorrhage focus position, the bleeding diffusion and the like of the patient are output according to the designed model.

The interface comprises a section for inputting or selecting personal information, medical history data and main complaint content of the patient; the interface should also contain a section that can take CT images of the patient directly or from a local file.

According to the interactive region selection of the small program page, different bleeding focus positions of the brain CT can be manually amplified for further observation.

As shown in fig. 1, the using process of the system for auxiliary diagnosis of cerebral hemorrhage using CT images according to the present invention includes: CT image shooting and inputting, CT image processing, text information inputting, core module training and applying and diagnosis conclusion outputting. Taking diagnosis and treatment of a patient with cerebral basal ganglia hemorrhage as an example, the processes of CT image shooting and inputting, CT image processing, text information inputting, core module training and applying and diagnosis conclusion outputting are as follows:

firstly, shooting an original image of a CT image by using a WeChat small program window, and performing primary processing, wherein the primary processing mainly comprises the following steps: image center cropping, image compression, image graying, and the like.

And then, putting the preprocessed image into a trained depth residual error network to realize fault classification: and classifying the CT image set of the patient by using a depth residual error network into 12 types of tomograms which are arranged in sequence to obtain a craniocerebral CT tomogram coarse classification sequence.

Next, in order to realize a more reasonable fault classification sequence, on the basis of a rough classification sequence, the classification sequence needs to be optimized by means of prior information and limiting conditions, so that the problems of too small difference between adjacent craniocerebral CT fault images and too large difference between craniocerebral CT images of different people are solved. The treatment processes are respectively as follows: sequence denoising, sequence non-strict increasing and sequence continuity.

Finally, carrying out inner brain region division processing on the optimized craniocerebral fault classification sequence, and carrying out brain region division on each layer of CT images by utilizing a brain region division mapping map in the craniocerebral fault of the CT map; and outputting the brain area part where the cerebral hemorrhage occurs, realizing the positioning function of the hemorrhage area and obtaining the diagnosis result with the positioning information.

In the process of constructing a neural network model in a system, particularly during CT image annotation, manual intervention can be performed in real time, specifically, a specialist doctor is invited to participate in the joint completion of the work, and the accuracy and the guidance value of the model are improved. In practical application, the invention can simultaneously carry out deep information mining on the disease of cerebral hemorrhage to form a more complete diagnosis conclusion, so that doctors can obtain all medical information of CT images in a short time.

It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.