Post-cerebral infarction image recognition method based on neural network

Abstract

The invention relates to the technical field of hemorrhage transformation prediction, in particular to a post-cerebral infarction image recognition method based on a neural network. The method has the advantages that the CNN and the residual blocks are used for automatically extracting complex deep features, end-to-end recognition of the CT image on the HT recognition result is achieved, preprocessed CT data are directly input into the CNN layer, shallow information is extracted, feature dimensions are increased, then the two kinds of residual blocks designed in the invention are sequentially stacked, depth features are continuously extracted, and the recognition accuracy is improved. And finally, further processing the depth features by using an HT predictor to obtain an identification result, constructing a prediction model by using CT plain films at different times and by means of deep learning, and predicting bleeding transformation possibly occurring in the future of the patient. Therefore, an emergency neurology department physician can better identify the patient in an early stage of the bleeding transformation risk which is not perfect in auxiliary examination, so that the attention of the patient with the high risk of bleeding transformation is increased, and the medication decision is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of hemorrhage transformation prediction, and particularly to an image recognition method for cerebral infarction based on a neural network. Background Art

[0002] The hemorrhagic transformation of ischemic stroke is defined as the absence of hemorrhage detected by the first CT examination after cerebral infarction. Hemorrhage is detected by a subsequent head CT examination, or hemorrhagic infarction HT can be detected by the first cranial CT. The occurrence of HT is one of the main events of concern in the treatment of acute ischemic stroke and is part of the natural process of acute cerebral infarction. The clinical incidence of HT is about 2.2%-44.0%, and the pathological incidence is as high as 70%. The treatment methods for ischemic stroke include intravenous thrombolysis, endovascular treatment, anticoagulation, and antiplatelet treatment. Almost all of these treatments increase the frequency and severity of hemorrhage transformation. The occurrence of HT is not only an important reason for the insufficient application of blood flow improvement treatments (such as thrombolysis, anticoagulation, antiplatelet, etc.), but also related to the poor prognosis of acute ischemic stroke. Neuroimaging is the main method for identifying the medical images of hemorrhagic transformation of ischemic stroke. Current clinical studies have been widely used in the imaging classification of hemorrhage transformation, such as the classification method of the European Cooperative Acute Stroke Study (ECASS). Many studies have attempted to identify hemorrhage transformation through non-contrast imaging, suggesting signs such as high-density arterial signs, white matter lesions, collateral circulation, and high-intensity acute injury markers; however, they still have limitations. Summary of the Invention

[0003] The purpose of the present invention aims to solve at least one of the technical defects.

[0004] For this reason, an object of the present invention is to propose an image recognition method for cerebral infarction based on a neural network to solve the problems mentioned in the background art and overcome the deficiencies existing in the prior art.

[0005] To achieve the above object, an embodiment of one aspect of the present invention provides an image recognition method for cerebral infarction based on a neural network, including the following steps: 1) Establish a deep neural network model; 2) Incorporate the head CT scan diagnostic review data during the patient's visit into the database; 3) Preprocess the data, randomly divide the database into a training set and a test set, and input the above data into the deep neural network model for model training; 4) Use CNN and residual blocks to extract the deep features of CT information and model the HT prediction system; 5) Directly input the preprocessed CT data into the CNN layer to extract shallow information and increase the feature dimension; 6) Sequentially stack the two types of residual blocks to continuously extract deep features; 7). Further process the depth features using the HT predictor to obtain the recognition result.

[0006] Preferably, according to any of the above solutions, the hemorrhagic transformation is defined as no hemorrhage being found in the first CT scan after ischemic stroke, and then hemorrhage being found in subsequent head CT scans. The screening of the patients is based on clinical and imaging diagnoses, and the imaging is reviewed by researchers.

[0007] Preferably, according to any of the above solutions, the head CT scan diagnostic review data during the patient's visit is the cases where the initial screening and re - screening results are inconsistent when the researchers review the imaging. The inconsistent cases are adjudicated by a committee composed of three experts, and the patients are included in the cohort when their clinical and imaging diagnoses are consistent with the researchers' judgments.

[0008] Preferably, according to any of the above solutions, the database also stores the imaging examination data received before the occurrence of hemorrhagic transformation, and in step 3), the database is randomly divided into a training set and a test set in a ratio of 9:1.

[0009] Preferably, according to any of the above solutions, when pre - processing the data, the thickness of each layer of the collected CT data is fixed at 5 mm, and the obtained CT data set is centrally cropped. After the data pre - processing, the dimension of the number of frames of each input is maintained at 28.

[0010] Preferably, according to any of the above solutions, the learning process of the residual block is shown as the following formula: ; ;

[0011] where X represents the input vector, represents the non - linear mapping function, represents the linear shortcut function, Y represents the output vector, W represents the weight vector, and b represents the bias vector.

[0012] Compared with the prior art, the advantages and beneficial effects of the present invention are: The image recognition method for post-cerebral infarction based on neural network uses CNN and residual blocks to automatically extract complex deep features, realizing end-to-end recognition of CT images for HT recognition results. The preprocessed CT data is directly input into the CNN layer to extract shallow information and increase the feature dimension. Then, the two types of residual blocks designed in the invention are sequentially stacked to continuously extract deep features, ensuring that the model performance does not decline. Finally, the HT predictor is used to further process the deep features to obtain the recognition result. CT plain films at different times are used to construct a prediction model with the help of deep learning to assist doctors in identifying medical images, so as to help doctors predict the possible hemorrhagic transformation of patients in the future, enabling emergency neurologists to better identify patients at an earlier stage of the risk of hemorrhagic transformation when auxiliary examinations are not yet perfect, thereby increasing the attention to patients at high risk of hemorrhagic transformation and assisting in drug use decisions. Description of the Drawings

[0013] Figure 1 Schematic diagrams of different residual blocks of the present invention; Figure 2 Schematic diagram of the architecture of CTNet of the present invention; Figure 3 Schematic diagram of the experimental example of the present invention; Figure 4 Schematic diagram of the model of the present invention; Figure 5 Schematic diagram of the image preprocessing of the present invention. Detailed Embodiments

[0014] The present invention will be further described below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0015] Embodiment 1: As Figures 1 to 5 shown, the image recognition method for post-cerebral infarction based on neural network includes the following steps: 1). Establish a deep neural network model; 2). Incorporate the head CT scan diagnostic review data during the patient's visit into the database; 3). Preprocess the data, and randomly divide the database into a training set and a test set, and input the above data into the deep neural network model for model training; 4). Use CNN and residual blocks to extract deep features of CT information and model the HT prediction system; 5). Directly input the preprocessed CT data into the CNN layer to extract shallow information and increase the feature dimension; 6). Sequentially stack the two types of residual blocks to continuously extract deep features; 7). Use the HT predictor to further process the deep features to obtain the recognition result.

[0016] The hemorrhagic transformation is defined as the absence of hemorrhage on the first CT scan after ischemic stroke, and subsequent hemorrhage is detected on subsequent head CT scans. The screening of the patients is based on clinical and imaging diagnoses, and the imaging is reviewed by the researchers.

[0017] CNN was proposed and further explored by Fukushima and Lecun et al. Due to its strong representativeness, CNN has been applied to different fields, such as mechanical metamaterial design, biomedical applications, and all-cause mortality prediction.

[0018] Next, the deep features of CT information are extracted using CNN and residual blocks to model the HT prediction system. CNN is a relatively common neural network structure, but as the number of layers increases, the performance of CNN will decay. The present invention uses a combination of CNN layers to propose residual CNN to avoid the problem of gradient explosion; In CT data, the morphology and location of lesions in each layer of the image are different, and subtle changes will affect the final recognition result; the present invention uses CNN and residual blocks to automatically extract complex deep features to achieve end-to-end recognition of the HT recognition result by CT images; the preprocessed CT data is directly input into the CNN layer to extract shallow information and increase the feature dimension; then, the two types of residual blocks designed in this paper are sequentially stacked to continuously extract deep features to ensure that the model performance does not decline; finally, the HT predictor is used to further process the deep features to obtain the recognition result.

[0019] This example retrospectively analyzed the data of 474 ischemic stroke patients from April 2014 to November 2022, collected the clinical and imaging data of the patients, and the hemorrhagic transformation was defined as the absence of hemorrhage on the first CT scan after ischemic stroke, and subsequent hemorrhage was detected on subsequent head CT scans.

[0020] All the patients included in the study did not show hemorrhage on the first CT scan at admission and received a series of head CT scans during the visit. The screening of the patients is based on clinical and imaging diagnoses, and the imaging is reviewed by the researchers.

[0021] For cases where the results of the initial screening and the review are inconsistent, a committee composed of three experts will make a ruling.

[0022] Only when the clinical and imaging diagnoses of the patients are consistent with the judgment of the researchers will they be included in the cohort.

[0023] Considering that patients may visit at different times, if a patient has received imaging examinations before the occurrence of hemorrhagic transformation, they can be included in the database.

[0024] We included patients who had not received HT treatment in the study at a ratio close to 1:1. The database was randomly divided into a training set and a test set at a ratio of 9:1. The model was trained with the above data.

[0025] When preprocessing the data, the thickness of each layer of the collected CT data was fixed at 5 mm, and the obtained CT dataset was centrally cropped. After the data preprocessing, the number of frames dimension of each input was maintained at 28.

[0026] The learning process of the residual block is shown as follows: ; ;

[0027] where X represents the input vector, represents the non-linear mapping function, represents the linear shortcut function, Y represents the output vector, W represents the weight vector, and b represents the bias vector; Increasing the number of CNN layers does not continuously improve the model performance, but instead reduces the model performance; therefore, the present invention designs different residual blocks to extract deep features.

[0028] The thickness of each layer of the CT data collected in this embodiment is fixed at 5 mm, and after data preprocessing, the number of frames dimension of each input is maintained at 28. This thickness and dimension can achieve the best prediction effect. After model prediction and machine learning, the final prediction accuracy reaches 74.52%, and it can accurately and effectively identify medical images.

[0029] Experimental example: As Figure 3 shown, the present invention conducted a statistical analysis to compare the performance of a deep learning model and a neurology expert in predicting the outcome of hemorrhage using non-contrast CT on a test dataset. The results are shown in the following table. The deep learning model is superior to the neurology expert in multiple metrics for identifying the outcome of hemorrhage. Table 2 shows the comparison between CTNet and clinicians. Specifically, the F1 score of the model is 78.94 (95% CI, 67.7 - 86.4), exceeding the F1 score of the neurology expert, which ranges from 43.92 to 66.26 (mean: 59.37).

[0030]

[0031] The AUC score of the model was 84.2 (95% CI, 75.8 - 92.1), the sensitivity was 71.55 (95% CI, 60.6 - 85.0), and the accuracy was 74.52 (95% CI, 63.9 - 83.2), all of which were better than the performance of neurologists. The AUC scores of neurologists were from 62.95 to 75.12 (mean: 67.66), the sensitivity was from 28.30 to 56.80 (mean: 45.48), and the accuracy was from 56.89 to 67.21 (mean: 60.41). Comparing the ROC curves of the model and neurologists, the model performed better in identifying the outcome of hemorrhage. The agreement rate of neurologists in predicting HT was between 0.61 - 0.87, while their agreement rate with the model was between 0.51 - 0.59. The confusion matrix showed the detailed identification results.

[0032] In summary, this image recognition method for post - cerebral infarction based on neural network uses CNN and residual blocks to automatically extract complex deep features, realizing the end - to - end recognition of the CT image for the HT recognition result. The pre - processed CT data is directly input into the CNN layer to extract shallow information and increase the feature dimension. Then, the two designed residual blocks of the invention are sequentially stacked to continuously extract deep features, ensuring that the model performance does not decline. Finally, the HT predictor is used to further process the deep features to obtain the recognition result. Using CT plain films at different times, a prediction model is constructed with the help of deep learning to assist doctors in identifying medical images, so as to help doctors predict the possible hemorrhage transformation of patients in the future, enabling emergency neurologists to better identify patients at an earlier stage of the hemorrhage transformation risk when the auxiliary examinations are not yet perfect, thus increasing the attention to patients at high risk of hemorrhage transformation and assisting in medication decisions.

[0033] Although the embodiments of the present invention have been shown and described above, it can be understood that the above - mentioned embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above - mentioned embodiments within the scope of the present invention without departing from the principles and purposes of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing images after cerebral infarction based on a neural network, characterized in that: The following steps are involved: 1) Establish a deep neural network model; 2) Incorporate the head CT scan diagnostic review data of the patient during the visit into the database; 3) Preprocess the data and randomly divide the database into training set and test set, and input the above data into the deep neural network model for model training; 4) Use CNN and residual blocks to extract deep features of CT information and model the HT prediction system; 5) Input the preprocessed CT data directly into the CNN layer to extract shallow information and increase the feature dimension; 6) The two residual blocks are sequentially superimposed to continuously extract deep features; 7) Use the HT predictor to further process the deep features to obtain the recognition results.

2. The method for recognizing post-cerebral infarction images based on a neural network according to claim 1, characterized in that: Hemorrhagic transformation was defined as the absence of hemorrhage on the initial CT scan after ischemic stroke followed by the presence of hemorrhage on a subsequent head CT scan; patients were screened based on clinical and radiographic diagnoses, with radiographs reviewed by investigators.

3. The method for recognizing post-cerebral infarction images based on a neural network according to claim 2, characterized in that: The head CT scan diagnostic review data during the patient's visit were cases in which the initial screening and re-examination results were inconsistent when the researchers reviewed the imaging. The inconsistent cases were adjudicated by a committee of three experts, and the patients were included in the cohort when their clinical and imaging diagnoses were consistent with the researchers' judgment.

4. The method for recognizing post-cerebral infarction images based on a neural network according to claim 3, characterized in that: The database also contains imaging examination data received before hemorrhagic transformation occurs. In step 3), the database is randomly divided into a training set and a test set at a ratio of 9:

1.

5. The method for recognizing post-cerebral infarction images based on a neural network according to claim 1, characterized in that: During the data preprocessing, the thickness of each layer of the acquired CT data is fixed at 5 mm, and the acquired CT data is centrally cropped. After the data preprocessing, the frame number dimension of each input is maintained at 28.

6. The method for recognizing post-cerebral infarction images based on a neural network according to claim 1, characterized in that: The residual block learning process is shown in the following formula: ； ； Where X represents the input vector, represents a nonlinear mapping function, represents the linear shortcut function, Y represents the output vector, W represents the weight vector, and b represents the bias vector.

Images