Method for accurately identifying brain injury region based on multi-modal image

Abstract

The application discloses a brain injury region accurate identification method based on multi-modal images, relates to the technical field of brain injury region identification, and comprises the following steps: performing brain injury identification demand evaluation, determining adaptive identification accuracy and adaptive identification time limit; predicting a brain injury state according to basic information, current sign data and historical impact characteristics; performing multi-modal image combination scheme optimization based on the predicted brain injury state distribution, and determining an optimal image combination scheme; collecting data of a target patient to obtain a multi-modal image set, and identifying a brain injury region of the target patient according to the multi-modal image set. The technical problems that the existing brain injury region identification method relies on single-modal image data, resulting in limited accuracy, and lacks reasonable image combination planning, resulting in low identification efficiency are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain injury region identification, and particularly relates to a brain injury region accurate identification method based on multi-modal images. BACKGROUND

[0002] With the continuous development of medical technology, the diagnosis and treatment of brain injury also require higher. However, the existing brain injury region identification method relies on single modal image data, which limits the identification accuracy. Because different modal images have different advantages in displaying brain injury, a single modal cannot fully reflect the characteristics of the injury.

[0003] On the other hand, the brain injury region identification method lacks reasonable image combination planning, resulting in low identification efficiency and unnecessary image examination. SUMMARY

[0004] The embodiment of the present application provides a brain injury region accurate identification method based on multi-modal images, which solves the technical problems that the existing brain injury region identification method relies on single modal image data, which limits the accuracy, and lacks reasonable image combination planning, which reduces the identification efficiency.

[0005] The technical solution of the present application to solve the above technical problems is as follows:

[0006] The present application provides a brain injury region accurate identification method based on multi-modal images, which comprises:

[0007] Based on the basic information of the target patient, the current sign data, the historical impact characteristics and the current clinical scene, the brain injury identification requirement is evaluated, and the adaptive identification accuracy and the adaptive identification time limit are determined;

[0008] According to the basic information, the current sign data and the historical impact characteristics, the brain injury state is predicted, and the predicted brain injury state distribution is obtained;

[0009] Taking the adaptive identification accuracy greater than the adaptive identification time limit as a constraint, and taking the maximum identification accuracy and identification efficiency as a target, the multi-modal image combination scheme is optimized based on the predicted brain injury state distribution, and the optimal image combination scheme is determined;

[0010] According to the optimal image combination scheme, the data of the target patient is collected to obtain a multi-modal image set, and the brain injury region of the target patient is identified according to the multi-modal image set.

[0011] The present application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0012] The application embodiment provides a precise brain injury region identification method based on multi-modal images. First, the brain injury identification needs of a target patient are evaluated considering various information of the target patient, and the adaptive identification accuracy and time limit are determined to provide reasonable target guidance for subsequent image combination planning. Second, the brain injury state is predicted using basic information, physical data and impact characteristics, and the obtained prediction distribution can more accurately reflect the possible injury of the patient. Then, in the multi-modal image combination scheme optimization process, the adaptive requirements are used as constraints, and the accuracy and efficiency are improved to select the optimal scheme from numerous initial schemes, thereby avoiding unnecessary image examination and improving the identification efficiency. Finally, the multi-modal image set is collected according to the optimal scheme, and the injury region is identified, thereby fully utilizing the advantages of different modal images, comprehensively reflecting the injury characteristics, and effectively improving the identification accuracy.

[0013] Through the above technical solution, the application can provide more accurate and efficient brain injury region identification results for clinicians, which is helpful for timely formulating reasonable treatment plans and improving the treatment effect and prognosis of patients. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0015] Figure 1 FIG. 1 is a flowchart of a precise brain injury region identification method based on multi-modal images provided by the application embodiment. DETAILED DESCRIPTION

[0016] The application embodiment provides a precise brain injury region identification method based on multi-modal images, which is used to solve the technical problems that the existing brain injury region identification method relies on single modal image data, which limits the accuracy, and lacks reasonable image combination planning, which leads to low identification efficiency.

[0017] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0018] In the description of the present application, the terms "first", "second" are only for descriptive purpose, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0019] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. In order to enable any person skilled in the art to implement and use the present application, the following description is given. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope in accordance with the principles and characteristics disclosed in the present application.

[0020] Embodiments, such as Figure 1 As shown, the embodiments of the present application provide a precise identification method for brain injury area based on multi-modal images, which comprises:

[0021] S10: Based on the basic information of the target patient, the current sign data, the historical impact characteristics and the current clinical scene, the brain injury identification requirement is evaluated, and the adaptive identification accuracy and the adaptive identification time limit are determined;

[0022] Specifically, the basic information, current sign data, historical impact characteristics and current clinical scene of the target patient are acquired, wherein the basic information includes age, gender, medical history and medication history, the sign data includes consciousness level, pupil reaction, motor function, vital signs and signs of intracranial hypertension, the historical impact characteristics include impact type and impact speed, and the clinical scene includes at least scene type and core clinical appeal.

[0023] In the embodiments of the present application, the basic information, current sign data, historical impact characteristics and current clinical scene information of the target patient are first collected.

[0024] Among them, the basic information includes the age, gender, medical history and medication history of the patient, etc.

[0025] Current vital signs data include the patient's level of consciousness, pupillary response, motor function, and signs of increased intracranial pressure, reflecting the patient's current physical condition. Furthermore, the level of consciousness can be quantified using the Glasgow Coma Scale; pupillary response can be assessed by examining pupil size and light reflex sensitivity; motor function is obtained by evaluating limb movement ability and muscle strength; vital signs can be obtained by monitoring blood pressure, heart rate, respiration, and blood oxygenation; signs of increased intracranial pressure can be observed for projectile vomiting or seizures.

[0026] Historical impact characteristics record past impact events that may have caused traumatic brain injury to the patient, including impact type and impact velocity. Impact types include falls, car accidents, and falls.

[0027] Current clinical scenarios include at least scenario types and core clinical needs, such as emergency resuscitation rooms, neuro-intensive care units, general wards, rehabilitation centers, and the core clinical needs within these scenarios, such as quickly ruling out surgical indications, accurately assessing prognosis, and developing rehabilitation plans. Different scenarios have different requirements for the time limit and accuracy of identification.

[0028] Based on the above information, a needs assessment for traumatic brain injury recognition is conducted. Combining clinical data and experience, the appropriate recognition accuracy and recognition time limit are determined, providing a basis for subsequent prediction of traumatic brain injury status and optimization of imaging combination schemes.

[0029] After determining the accuracy and time limit of the adaptive recognition, the system combines the patient's basic information, current vital signs data, and historical impact characteristics to predict the state of traumatic brain injury and optimize efficient multimodal image combination schemes, thereby improving the accuracy and efficiency of traumatic brain injury area identification.

[0030] Specifically, step S10 in the method includes:

[0031] The basic information and current vital sign data are extended according to a preset feature tolerance threshold to determine the basic information range and the vital sign data range.

[0032] Using the aforementioned basic information range, vital sign data range, historical impact characteristics, and current clinical scenario as search constraints, historical qualified traumatic brain injury identification records are retrieved, and a sample injury identification accuracy set and a sample injury identification duration set are obtained through screening. The mean sample injury identification accuracy and the mean sample injury identification duration are then calculated.

[0033] The average sample damage recognition accuracy is used as the adaptation recognition accuracy, and the average sample damage recognition time is used as the adaptation recognition time limit.

[0034] In this embodiment, firstly, since the patient's basic information and current vital signs data may fluctuate within a certain range, feature expansion can more comprehensively consider possible situations. Therefore, feature expansion is performed on the basic information and current vital signs data according to a preset feature tolerance threshold. The preset feature tolerance threshold is determined based on a large amount of clinical data and experience, and can expand the range of data while ensuring accuracy. For example, for the patient's age information, the preset feature tolerance threshold allows for a certain range of fluctuation above and below the actual age, thereby including patient data that may have similar injury conditions.

[0035] Secondly, after determining the basic information range and the vital sign data range, a search was conducted in historical qualified traumatic brain injury identification records using the range, historical impact characteristics, and current clinical scenario as search constraints. These historical qualified traumatic brain injury identification records, after screening and verification, contain identification information from different patients. Through the search, samples similar to the current target patient were selected, resulting in a sample injury identification accuracy set and a sample injury identification duration set.

[0036] Statistical analysis was performed on the selected sample data to calculate the mean accuracy and mean duration of sample injury identification. The mean calculation eliminated the influence of individual extreme data points, reflecting the overall situation. The mean accuracy of sample injury identification was used as the appropriate identification accuracy, and the mean duration of sample injury identification was used as the appropriate identification time limit. The determined appropriate values ​​considered both the current patient's specific situation and the experience gained from historical data, providing a scientific and reasonable basis for subsequent prediction of traumatic brain injury status and optimization of imaging combination plans.

[0037] The above steps dynamically determine the appropriate recognition accuracy and time limit based on the specific circumstances of different patients, avoiding the errors that may be caused by using fixed standards, thereby improving the accuracy and efficiency of identifying craniocerebral injury areas and better meeting clinical needs.

[0038] S20: Based on the basic information, current vital signs data and historical impact characteristics, predict the state of traumatic brain injury and obtain the predicted distribution of the state of traumatic brain injury.

[0039] In this embodiment, a machine learning model is used to analyze the patient's basic information, current vital signs, and historical impact characteristics. Specifically, the patient's age, gender, medical history, medication history, level of consciousness, pupillary response, motor function, vital signs, signs of intracranial hypertension, impact type, and impact velocity are used as input features and fed into a pre-trained machine learning model. This model establishes a mapping relationship between the input features and the state of traumatic brain injury through learning and training on a large amount of historical traumatic brain injury case data.

[0040] During model training, methods such as cross-validation are used to evaluate model performance and optimize model parameters to improve predictive accuracy. The trained model can output a predicted distribution of traumatic brain injury states based on input patient data. This distribution includes information such as the probability of different types of traumatic brain injury and the severity of the injury.

[0041] For example, the model might predict a 30% probability of a patient experiencing a concussion, a 20% probability of an intracranial hematoma, and a 50% probability of a cerebral contusion. It could further predict the severity level of different injury types, such as mild, moderate, or severe. This predictive information provides a basis for optimizing subsequent multimodal imaging combinations, helping doctors to more effectively select appropriate imaging methods and improve the accuracy and efficiency of identifying areas of traumatic brain injury.

[0042] Specifically, step S20 in the method includes:

[0043] Based on historical qualified traumatic brain injury identification records, a sample basic information set, a sample vital sign data set, and a sample impact feature set are collected. The distribution of the patient's sample traumatic brain injury status under different sample basic information, sample vital sign data, and sample impact features is collected to obtain the sample traumatic brain injury status distribution set.

[0044] Using the sample basic information set, sample vital signs dataset, and sample impact feature set as inputs, and the sample traumatic brain injury state distribution set as supervision, a deep learning model is trained until convergence to generate a traumatic brain injury state prediction plugin.

[0045] Using the aforementioned traumatic brain injury status prediction plugin, a predicted distribution of traumatic brain injury status is obtained based on the basic information, current vital signs data, and historical impact characteristics. The predicted distribution of traumatic brain injury status includes several predicted types of traumatic brain injury and several predicted injury intensities.

[0046] In this embodiment, firstly, a sample basic information set, a sample vital signs dataset, and a sample impact feature set are collected from historical qualified traumatic brain injury identification records. The sample basic information set covers the age, gender, past medical history, and medication history of different patients; the sample vital signs dataset includes the level of consciousness, pupillary response, motor function, vital signs, and signs of intracranial hypertension; the sample impact feature set records information such as impact type and impact velocity. Simultaneously, the distribution of the patient's traumatic brain injury state under different scenarios with different sample basic information, sample vital signs data, and sample impact features is collected to form a sample traumatic brain injury state distribution set.

[0047] Secondly, the deep learning model is trained using a set of basic sample information, a set of sample vital signs, and a set of sample impact features as input, and a set of sample traumatic brain injury state distributions as supervision. During training, the model continuously adjusts its parameters to minimize the error between the predicted results and the actual distribution of traumatic brain injury states. Through repeated training and optimization, the model converges, at which point a traumatic brain injury state prediction plugin that can accurately predict traumatic brain injury states is generated.

[0048] Finally, using a trained traumatic brain injury (TBI) state prediction plugin, the target patient's basic information, current vital signs, and historical impact characteristics are input into the plugin for analysis. Based on the learned knowledge and patterns, the plugin outputs a predicted TBI state distribution, which includes several predicted TBI types, such as skull fracture, intracranial hematoma, cerebral contusion, and subdural hematoma, as well as corresponding predicted injury intensities, such as mild, moderate, and severe.

[0049] For example, the steps for building and training a traumatic brain injury state prediction plugin based on a deep learning model are as follows:

[0050] First, data preparation involves collecting a set of basic information about the samples, a dataset of vital signs, a set of impact characteristics, and basic information about different samples, based on historical records of qualified traumatic brain injury identification. The collected data is then preprocessed to remove outliers and noisy data.

[0051] Secondly, the model is constructed based on a deep learning model, serving as the model architecture for the traumatic brain injury (TBI) state prediction plugin. The preprocessed sample basic information set, sample vital sign dataset, and sample impact feature set are used as inputs, and the predicted TBI state distribution is used as the output. The predicted TBI state distribution includes several predicted TBI types and several predicted injury intensities. The selected model is trained. The input layer has the number of nodes equal to the dimension of the input features. For example, if there are 10 features in total (sample basic information set, sample vital sign dataset, and sample impact feature set), the input layer contains 10 nodes. One to three hidden layers are set, with the number of nodes in each layer adjusted experimentally, such as 64 or 32. The ReLU activation function is used. The number of nodes in the output layer equals the number of predicted TBI state distributions. For example, if the evaluation time requires one node, the output layer generally does not use an activation function and directly outputs continuous values.

[0052] Finally, the model is trained, and its parameters are continuously adjusted to minimize the error between the predicted and actual values. Cross-validation is used to evaluate the trained model and verify its generalization ability and accuracy. In each training iteration, the sample brain injury state distribution set is used as supervision. The Adam optimizer and mean squared error (MSE) loss function are used to construct the training framework, and the model parameters are adjusted using the gradient descent algorithm. The batch size is set to 32, the total number of training epochs is 50, and an early stopping mechanism with a patience of 5 is introduced. When the validation set loss does not decrease for 5 consecutive epochs, the training process is automatically terminated, resulting in a trained brain injury state prediction plugin.

[0053] The prediction results obtained above provide detailed and targeted basis for subsequent optimization of multimodal imaging combination schemes, enabling more precise targeting of possible damage situations when selecting imaging examination methods, and further improving the accuracy and efficiency of craniocerebral injury area identification.

[0054] The distribution of patients' traumatic brain injury status under different scenarios of collecting basic sample information, vital sign data, and impact characteristics includes:

[0055] The collection of historical traumatic brain injury type sets and historical traumatic brain injury feature sets for patients in different scenarios of collecting basic information of samples, vital sign data of samples, and impact characteristics of samples, wherein the historical traumatic brain injury types and historical traumatic brain injury features correspond one-to-one.

[0056] Based on the historical traumatic brain injury feature set, the historical traumatic brain injury intensity set is determined, and the sample traumatic brain injury status distribution is constructed by combining the historical traumatic brain injury type set.

[0057] In this embodiment, firstly, from historical qualified traumatic brain injury identification records, a set of historical traumatic brain injury types and a set of historical traumatic brain injury features are collected for patients under different sample basic information, sample vital sign data, and sample impact characteristic scenarios. The set of historical traumatic brain injury types covers a variety of possible injury types, such as skull fracture, intracranial hematoma, cerebral contusion and laceration, and subdural hematoma; the set of historical traumatic brain injury features includes various features corresponding to the injury type, such as the location of the fracture, the size of the hematoma, and the extent of the cerebral contusion and laceration, and there is a one-to-one correspondence between historical traumatic brain injury types and historical traumatic brain injury features.

[0058] Secondly, the intensity of historical traumatic brain injury is assessed and determined based on a set of historical traumatic brain injury characteristics. Different assessment criteria apply to different injury characteristics. For example, with intracranial hematomas, if the hematoma is small and the compression of surrounding brain tissue is mild, the injury intensity is assessed as mild; if the hematoma is large, causing significant mass effect and serious conditions such as midline displacement, the injury intensity may be assessed as severe. Similarly, for skull fractures, if it is a simple linear fracture without significant displacement or damage to brain tissue, the injury intensity may be mild; however, a comminuted fracture accompanied by significant depression or severe compression of brain tissue indicates a severe injury.

[0059] Furthermore, after determining the historical set of traumatic brain injury intensities, this set is combined with the historical set of traumatic brain injury types to construct a sample distribution of traumatic brain injury states. This sample distribution reflects the traumatic brain injury status of patients under different sample baseline information, vital sign data, and impact characteristic scenarios. It provides accurate data for subsequent training of deep learning models, enabling the trained traumatic brain injury state prediction plugin to output a predicted distribution of traumatic brain injury states based on the target patient's baseline information, current vital sign data, and historical impact characteristics. This further improves the accuracy and efficiency of traumatic brain injury region identification, better meeting the needs of clinical diagnosis and treatment.

[0060] S30: With constraints of greater than the adaptive recognition accuracy and less than the adaptive recognition time limit, and with the goal of maximizing recognition accuracy and recognition efficiency, the optimal image combination scheme is determined by optimizing the combination scheme of multimodal images based on the predicted distribution of craniocerebral injury status.

[0061] The preset image type space includes CT images, T1-weighted images, T2-weighted images, diffusion-weighted images, magnetic susceptibility-weighted images, diffusion tensor images, magnetic resonance spectroscopy, liquid attenuation inversion recovery sequence, apparent diffusion coefficient map, and magnetic resonance perfusion-weighted images.

[0062] In this embodiment, the distribution of predicted traumatic brain injury states reveals information such as the probability of occurrence and severity of different types of traumatic brain injury. Multimodal imaging includes, but is not limited to, CT images, T1-weighted images, T2-weighted images, diffusion-weighted images, magnetic susceptibility-weighted images, diffusion tensor images, magnetic resonance spectroscopy, fluid attenuation inversion recovery sequences, apparent diffusion coefficient maps, and magnetic resonance perfusion-weighted images. Each imaging modality has different advantages and limitations in identifying different types and degrees of traumatic brain injury.

[0063] For example, CT images provide a clear view of acute traumatic brain injury, such as skull fracture and acute intracranial hemorrhage, and can quickly provide a general picture of the injury; T1-weighted and T2-weighted images can reflect the anatomical structure and lesion characteristics of brain tissue from different angles, which helps to detect lesions in the brain parenchyma; diffusion-weighted images are of great value in the diagnosis of lesions such as early cerebral infarction; and magnetic susceptibility-weighted images are excellent in detecting microbleeds.

[0064] In clinical diagnosis and research, the aforementioned imaging modalities are combined for the identification of traumatic brain injury regions. More image types generally result in higher recognition accuracy but also require longer recognition times and lower efficiency; conversely, fewer image types lead to higher recognition efficiency but lower accuracy. Specifically, the selection and optimization of possible image combination schemes are based on constraints of matching recognition accuracy and time limits. For example, for traumatic brain injury types with a high predicted probability and severe severity, imaging methods capable of accurately identifying this type of injury are prioritized; simultaneously, considering time constraints, combinations of examinations that are excessively time-consuming are avoided.

[0065] During the optimization process, factors such as the cost of imaging examinations and radiation dose are considered. For minor injuries, if a single imaging examination can meet the identification requirements, there is no need to use a combination of multiple examination methods, thereby reducing costs and the radiation dose received by the patient.

[0066] The determined optimal imaging combination scheme can identify the brain injury area with the highest possible accuracy within the specified identification time limit, providing clinicians with accurate diagnostic basis, thereby developing more effective treatment plans and further improving the treatment effect and prognosis of patients.

[0067] Specifically, step S30 in the method includes:

[0068] A preset image type space for recognizing traumatic brain injury is obtained, and a combination scheme of multimodal images is enumerated based on the preset image type space to generate several initial image combination schemes.

[0069] Based on the aforementioned initial image combination schemes, the recognition accuracy and recognition time are predicted respectively, and several predicted recognition accuracies and several predicted recognition times are output.

[0070] Based on the aforementioned prediction recognition accuracy and prediction recognition duration, and constrained by being greater than the adaptive recognition accuracy and less than the adaptive recognition time limit, the aforementioned initial image combination schemes are screened to obtain multiple qualified image combination schemes.

[0071] Based on the predicted distribution of traumatic brain injury status, several predicted recognition accuracies, and several predicted recognition durations, with the goal of maximizing recognition accuracy and efficiency, the quality of the multiple qualified image combination schemes is evaluated, and the qualified image combination scheme corresponding to the highest quality coefficient is selected as the optimal image combination scheme.

[0072] In this embodiment, firstly, a preset image type space for identifying traumatic brain injury is obtained. This space includes all possible image types for identifying traumatic brain injury, such as CT images, T1-weighted images, and T2-weighted images mentioned above. Based on this preset image type space, combination schemes for multimodal images are enumerated. The enumeration process considers various possible combinations of different image types, thereby generating several initial image combination schemes. For example, there may be combinations containing only one image type, or combinations containing two, three, or even more image types.

[0073] Secondly, for each of the generated initial image combination schemes, the recognition accuracy and recognition time are predicted. This prediction is achieved by establishing a prediction model. For each initial image combination scheme, the prediction model outputs the corresponding predicted recognition accuracy and predicted recognition time based on the types of images it contains and the performance of the images in recognizing different types and degrees of traumatic brain injury.

[0074] Then, based on several predicted recognition accuracies and several predicted recognition times, and with the constraint that the predicted recognition accuracy is greater than the adapted recognition accuracy and the predicted recognition time is less than the adapted recognition time, several initial image combination schemes are screened. Only initial image combination schemes with predicted recognition accuracy greater than the adapted recognition accuracy and predicted recognition time less than the adapted recognition time can be selected as qualified image combination schemes.

[0075] Finally, based on the predicted distribution of traumatic brain injury states, several predicted recognition accuracies, and several predicted recognition times, a scheme quality evaluation was conducted on multiple qualified image combination schemes with the goal of maximizing recognition accuracy and efficiency. During the evaluation process, the performance of different qualified image combination schemes in identifying various possible types and degrees of traumatic brain injury was considered, with the output of qualified recognition accuracy and recognition time.

[0076] By setting evaluation indicators and calculation methods, a quality coefficient is calculated for each qualified image combination scheme. The qualified image combination scheme with the highest quality coefficient is selected as the optimal image combination scheme. The optimal image combination scheme can maximize the recognition accuracy and efficiency while meeting the requirements of matching recognition accuracy and recognition time limit, providing strong support for clinicians to accurately diagnose traumatic brain injury, thereby helping to develop more scientific and effective treatment plans and improve patients' treatment outcomes and prognosis.

[0077] Furthermore, based on the aforementioned initial image combination schemes, the recognition accuracy and recognition time are predicted respectively, and several predicted recognition accuracies and several predicted recognition times are output, including:

[0078] Randomly select a first initial image combination scheme from the plurality of initial image combination schemes;

[0079] Based on historical qualified traumatic brain injury identification records, several first historical identification durations of the first initial image combination scheme are collected, and the average of the first historical identification durations is calculated as the first predicted identification duration.

[0080] A deep learning-based craniocerebral injury recognition accuracy predictor is constructed. Based on the basic information, current vital signs data, and historical impact characteristics, the accuracy of injury recognition is predicted for the first initial image combination scheme, and the first predicted recognition accuracy is output.

[0081] In this embodiment, firstly, one of several initial image combination schemes is randomly selected as the first initial image combination scheme. The selection process is completely random to ensure the objectivity and universality of subsequent predictions.

[0082] Secondly, based on historical records of qualified traumatic brain injury identification, several historical identification times for the first initial imaging combination scheme were collected. Historical identification times represent the time spent identifying traumatic brain injuries using this imaging combination scheme in previous practical applications. The average historical identification time was calculated and used as the first predicted identification time. The aim is to estimate the time required for future use of this imaging combination scheme based on historical data, providing a time-related reference for subsequent selection of qualified imaging combination schemes.

[0083] For example, assume the first initial image combination scheme is a combination of CT images and T2-weighted images. In historical qualified traumatic brain injury identification records, the durations of this combination in 10 identifications were collected as 20 minutes, 22 minutes, 18 minutes, 25 minutes, 21 minutes, 19 minutes, 23 minutes, 24 minutes, 20 minutes, and 22 minutes, respectively. By calculating the average duration, i.e., (20+22+18+25+21+19+23+24+20+22)÷10=21.4 minutes, the average duration of the first historical identification is obtained as 21.4 minutes, which is used as the first predicted identification duration.

[0084] Then, a traumatic brain injury (TBI) recognition accuracy predictor was constructed based on deep learning technology. After construction, the TBI recognition accuracy predictor was used to predict the injury recognition accuracy of the first initial image combination scheme based on the target patient's basic information, current vital signs data, and historical impact characteristics. During the prediction process, the predictor analyzes the TBI recognition capability of the image combination scheme under specific circumstances for the target patient based on learned knowledge and patterns. Finally, the first predicted recognition accuracy is output, reflecting the degree of accuracy that the image combination scheme may achieve in recognizing the target patient's TBI, providing an accuracy indicator for subsequent evaluation of the quality of the image combination scheme.

[0085] For example, the steps for constructing a prediction tool for the accuracy of traumatic brain injury recognition based on deep learning are as follows:

[0086] First, data preparation involves collecting basic patient information, current vital signs, historical impact characteristics, imaging techniques used, and corresponding identification results based on historical traumatic brain injury (TBI) identification data. The data is then preprocessed, including data cleaning, feature extraction, and normalization, to improve data quality and usability.

[0087] Secondly, the model is built based on a convolutional neural network. The preprocessed basic information, current vital signs data, historical impact features, and image combination scheme are input into the CNN model. The features of the data are extracted through convolutional layers, the feature dimensionality is reduced by pooling layers, and the fully connected layers are used for classification or regression. The predicted recognition accuracy is used as the output.

[0088] Next, for model training, historical data was divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training set was used to train the model, and the parameters were continuously adjusted using backpropagation to minimize the error between the predicted and actual results. During training, the validation set was used to monitor model performance and prevent overfitting. The Adam optimizer and Mean Squared Error (MSE) loss function were used to construct the training framework, and the model parameters were adjusted using gradient descent. The batch size was set to 32, the total number of training epochs to 50, and an early stopping mechanism (patience=5) was introduced. The training process was automatically terminated when the validation set loss did not decrease for five consecutive epochs, resulting in a trained, accurate predictor of traumatic brain injury.

[0089] Once the model is trained, it is evaluated using a test set, and metrics such as accuracy, recall, and F1 score are calculated to assess its performance. If the model's performance meets the requirements, it can be used to predict the accuracy of damage identification for the first initial image combination scheme.

[0090] Repeat the above steps to predict the recognition accuracy and recognition time for all initial image combination schemes, thereby obtaining several predicted recognition accuracies and several predicted recognition times, providing data support for subsequent screening of qualified image combination schemes and selection of the optimal image combination scheme.

[0091] Specifically, based on the predicted distribution of traumatic brain injury status, several prediction recognition accuracies, and several prediction recognition durations, and with the goal of maximizing recognition accuracy and efficiency, a scheme quality evaluation is performed on the multiple qualified image combination schemes, including:

[0092] Several damage influence weights are configured based on the several predicted damage intensities, wherein the damage influence weights are positively correlated with the predicted damage intensities;

[0093] Based on historical qualified traumatic brain injury identification records and a preset image type space, the identification correlation degree is evaluated for the several predicted traumatic brain injury types and the multiple preset image types respectively, resulting in several identification correlation degree sets.

[0094] Based on the aforementioned damage impact weights, the aforementioned identification correlation sets are compensated and adjusted to obtain several identification compensation correlation sets;

[0095] Based on the aforementioned sets of recognition compensation correlation degrees, the recognition compensation correlation degrees corresponding to multiple images included in each qualified image combination scheme are summed to obtain multiple total correlation degrees;

[0096] Based on the aforementioned predicted recognition accuracy and predicted recognition duration, with the goal of maximizing recognition accuracy and recognition efficiency, the multiple qualified image combination schemes are evaluated for scheme quality, and multiple scheme fitness values ​​are output. The scheme fitness value is positively correlated with the predicted recognition accuracy and negatively correlated with the predicted recognition duration.

[0097] Based on the multiple total correlation coefficients and multiple scheme fitness, the multiple qualified image combination schemes are evaluated for scheme quality, and multiple scheme quality coefficients are output.

[0098] In this embodiment, firstly, based on several predicted traumatic brain injury types and several predicted injury intensities in the predicted distribution of traumatic brain injury states, a corresponding injury influence weight is assigned to each predicted injury intensity. The injury influence weight is positively correlated with the predicted injury intensity; that is, the greater the predicted injury intensity, the higher the corresponding injury influence weight. For example, for a traumatic brain injury type with a high predicted injury intensity, its corresponding injury influence weight can be set to 0.8, while for a type with a low predicted injury intensity, the weight can be set to 0.2.

[0099] Secondly, based on historical records of qualified traumatic brain injury (TBI) identification and a pre-defined image type space, the identification correlation was evaluated for several predicted TBI types and multiple pre-defined image types. Identification correlation reflects the ability and effectiveness of a particular image type in identifying specific TBI types. Through analysis and statistics of historical data, the degree of correlation between each image type and different predicted TBI types was calculated, resulting in several sets of identification correlation. For example, if the predicted TBI type is cerebral infarction, the identification correlation of diffusion-weighted images may be high, while for skull fractures, the identification correlation of CT images is even higher.

[0100] Then, several damage impact weights are used to compensate and adjust several identification correlation sets by multiplying them. The adjusted identification correlation sets are called the identification compensated correlation sets. For example, for a certain type of traumatic brain injury with a high predicted damage intensity, its corresponding damage impact weight is larger. When adjusting the identification correlation, the identification correlation of this type of injury is given greater weight, thereby highlighting its importance in the assessment.

[0101] Next, based on several sets of recognition-compensation correlation scores, the recognition-compensation correlation scores corresponding to multiple images included in each qualified image combination scheme are summed. This summation operation yields the total correlation score for each qualified image combination scheme. The total correlation score reflects the comprehensive ability of the image combination scheme to identify all predicted types of traumatic brain injury. For example, a qualified image combination scheme includes CT images and T1-weighted images. The recognition-compensation correlation scores for various predicted types of traumatic brain injury are calculated for both CT and T1-weighted images, and then these correlation scores are summed to obtain the total correlation score for the combination scheme.

[0102] For example, a qualified image combination scheme includes CT images and T2-weighted images. The recognition compensation correlation coefficients of CT images for the three predictive types of traumatic brain injury—cerebral contusion, cerebral hemorrhage, and skull fracture—are 0.7, 0.8, and 0.9, respectively. The recognition compensation correlation coefficients of T2-weighted images for the three injury types are 0.6, 0.7, and 0.5, respectively. Adding the recognition compensation correlation coefficients of the CT images and T2-weighted images for the corresponding injury types, i.e., (0.7+0.6), (0.8+0.7), and (0.9+0.5), we obtain the total correlation coefficients of this combination scheme for cerebral contusion, cerebral hemorrhage, and skull fracture, which are 1.3, 1.5, and 1.4, respectively.

[0103] Subsequently, based on several predicted recognition accuracies and several predicted recognition times, and with the goal of maximizing recognition accuracy and efficiency, the quality of multiple qualified image combination schemes is evaluated, and multiple scheme fitness scores are output. Scheme fitness is positively correlated with predicted recognition accuracy and negatively correlated with predicted recognition time. This indicates that image combination schemes with higher recognition accuracy and shorter recognition times have higher scheme fitness. For example, an image combination scheme with high predicted recognition accuracy and short predicted recognition time will have relatively high scheme fitness.

[0104] Finally, based on multiple total correlation coefficients and multiple scheme fitness scores, a final scheme quality evaluation is performed on multiple qualified image combination schemes, outputting multiple scheme quality coefficients. The scheme quality coefficients comprehensively consider the image combination scheme's ability to identify different damage types, as well as its identification accuracy and efficiency. By comparing the scheme quality coefficients of each qualified image combination scheme, the qualified image combination scheme with the highest quality coefficient is selected as the optimal image combination scheme.

[0105] The optimal image combination scheme determined above can maximize the ability to identify the brain injury area while meeting the requirements of recognition accuracy and recognition time limit, providing clinicians with more accurate and efficient diagnostic basis, and thus enabling the formulation of more scientific and effective treatment plans for patients.

[0106] Furthermore, the multiple total correlation coefficients and multiple scheme fitness coefficients are processed dimensionlessly and weighted to obtain multiple scheme quality coefficients.

[0107] In this embodiment, multiple total relevance scores and multiple solution fitness scores are first processed to be dimensionless. Because the total relevance score and solution fitness scores have different dimensions, direct calculation would affect the accuracy of the results. A normalization method is adopted, for example, performing min-max normalization on the total relevance score and solution fitness scores respectively. For the total relevance score, the minimum and maximum values ​​among all total relevance scores are found. The minimum value is subtracted from each total relevance score, and then divided by the difference between the maximum and minimum values ​​to obtain the normalized total relevance score. For the solution fitness scores, the minimum and maximum values ​​are similarly found, and the same normalization operation is performed to obtain the normalized solution fitness scores.

[0108] For example, suppose there are three qualified image combination schemes with total correlation coefficients of 1.2, 1.5, and 1.8, and scheme fitness of 0.6, 0.8, and 0.9, respectively. The minimum total correlation coefficient is 1.2, and the maximum is 1.8.

[0109] The normalized total correlation coefficient of the first scheme is (1.2-1.2)÷(1.8-1.2)=0;

[0110] The normalized total correlation coefficient of the second scheme is (1.5-1.2)÷(1.8-1.2)=0.5;

[0111] The normalized total correlation coefficient of the third scheme is (1.8-1.2)÷(1.8-1.2)=1.

[0112] The fitness of the scheme has a minimum value of 0.6 and a maximum value of 0.9.

[0113] The fitness of the normalized scheme for the first scheme is (0.6-0.6)÷(0.9-0.6)=0;

[0114] The fitness of the normalized scheme for the second scheme is (0.8-0.6)÷(0.9-0.6)≈0.67;

[0115] The fitness of the normalized scheme for the third scheme is (0.9-0.6)÷(0.9-0.6)=1.

[0116] Secondly, a weighted calculation is performed on the normalized total relevance and protocol fitness. Appropriate weights are assigned to the total relevance and protocol fitness based on the actual situation and clinical needs. For example, if the ability of the imaging combination protocol to identify different types of damage is more important, the weight of the total relevance can be appropriately increased; if recognition accuracy and efficiency are more important, the weight of the protocol fitness can be increased. Assuming the weight of the total relevance is 0.6 and the weight of the protocol fitness is 0.4, for each qualified imaging combination protocol, the normalized total relevance is multiplied by its weight, and then added to the normalized protocol fitness multiplied by its weight to obtain the protocol quality coefficient.

[0117] For example, the quality coefficient of the first scheme is 0×0.6+0×0.4=0;

[0118] The quality coefficient of the second scheme is 0.5×0.6+0.67×0.4=0.568;

[0119] The quality coefficient of the third scheme is 1×0.6+1×0.4=1.

[0120] Using the above method, a scheme quality coefficient is calculated for each qualified image combination scheme. Through dimensionless processing and weighted calculation, the ability of the image combination scheme to identify different damage types, as well as its identification accuracy and efficiency, can be comprehensively considered, resulting in a more scientific and reasonable scheme quality coefficient.

[0121] Ultimately, the optimal imaging combination scheme was selected from multiple qualified imaging combination schemes, based on its highest quality coefficient. This scheme can provide stronger support for clinicians to accurately diagnose traumatic brain injury while meeting the requirements of matching recognition accuracy and matching recognition time limit. It helps doctors to more accurately judge the patient's traumatic brain injury, thereby developing a treatment plan that is more in line with the patient's actual situation and improving the patient's treatment effect and prognosis.

[0122] S40: Collect data from the target patient according to the optimal image combination scheme to obtain a multimodal image set, and identify the craniocerebral injury area of ​​the target patient based on the multimodal image set.

[0123] In this embodiment, after determining the optimal image combination scheme, data is acquired from the target patient. The data acquisition process follows the image types and acquisition parameters specified in the scheme to ensure that an accurate and complete multimodal image set is obtained. For example, if the scheme includes CT images and MRI images, then when acquiring CT images, appropriate scanning parameters, such as tube voltage, tube current, and slice thickness, are set according to the patient's specific condition; when acquiring MRI images, appropriate sequences, such as T1-weighted images, T2-weighted images, and diffusion-weighted images, are selected, and the corresponding imaging parameters are adjusted.

[0124] When acquiring multimodal image sets, it is essential to ensure patient cooperation during the acquisition process to avoid artifacts caused by patient movement, which could affect image clarity and accuracy. Furthermore, the acquisition equipment should be maintained and calibrated to ensure stable and reliable performance.

[0125] After acquiring the multimodal image set, the brain injury region of the target patient was identified. First, the multimodal image set was preprocessed, including image denoising, enhancement, and registration. Noise removal improves image quality; enhancement highlights injury features for easier subsequent identification; and registration spatially aligns images from different modalities, ensuring correspondence between the same anatomical structures.

[0126] Then, deep learning algorithms are used to analyze the preprocessed multimodal images. CNNs can automatically extract features from the images and classify and locate the areas of traumatic brain injury. During the identification process, the experience and expertise of clinicians are combined to review and correct the identification results, improving the accuracy of the identification.

[0127] In summary, this application identifies the brain injury region of a target patient through the analysis and recognition of a multimodal image dataset. Based on the recognition results, the type, degree, and extent of the brain injury are determined, thereby selecting appropriate treatment methods, such as surgical treatment or drug treatment, to improve the patient's treatment outcome and prognosis.

[0128] In summary, the embodiments of this application have at least the following technical effects:

[0129] This application provides a method for accurate identification of traumatic brain injury regions based on multimodal imaging. First, it assesses the need for traumatic brain injury identification by considering various aspects of the target patient's information, clarifying the appropriate identification accuracy and timeframe, and providing a reasonable target guidance for subsequent image combination planning. Second, it predicts the state of traumatic brain injury using basic information, vital signs data, and impact characteristics; the resulting predicted distribution can more accurately reflect the patient's potential injury. Then, in the process of optimizing the multimodal image combination scheme, with adaptation requirements as constraints and improving accuracy and efficiency as the goal, the optimal scheme is selected from numerous initial schemes, avoiding unnecessary imaging examinations and improving identification efficiency. Finally, it collects a multimodal image set according to the optimal scheme and identifies the injury region, fully utilizing the advantages of different modalities of imaging to comprehensively reflect the injury characteristics and effectively improve identification accuracy. Through the above technical solution, this application can provide clinicians with more accurate and efficient results for identifying traumatic brain injury regions, helping to formulate reasonable treatment plans in a timely manner and improve patient treatment outcomes and prognosis.

[0130] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0131] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0132] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for precise identification of craniocerebral injury regions based on multimodal imaging, characterized in that, The methods include: Based on the target patient's basic information, current vital signs data, historical impact characteristics, and current clinical scenario, a needs assessment for traumatic brain injury identification is conducted to determine the appropriate identification accuracy and timeframe, including: The basic information and current vital sign data are extended according to a preset feature tolerance threshold to determine the basic information range and the vital sign data range. Using the aforementioned basic information range, vital sign data range, historical impact characteristics, and current clinical scenario as search constraints, historical qualified traumatic brain injury identification records are retrieved, and a sample injury identification accuracy set and a sample injury identification duration set are obtained through screening. The mean sample injury identification accuracy and the mean sample injury identification duration are then calculated. The average sample damage recognition accuracy is used as the adaptation recognition accuracy, and the average sample damage recognition time is used as the adaptation recognition time limit. Based on the aforementioned basic information, current vital signs data, and historical impact characteristics, the state of traumatic brain injury is predicted, and the predicted distribution of traumatic brain injury states is obtained. Constrained by a recognition accuracy greater than the specified accuracy and less than the specified recognition time limit, and aiming to maximize recognition accuracy and efficiency, an optimal image combination scheme is determined based on the predicted distribution of traumatic brain injury states. This includes: A preset image type space for recognizing traumatic brain injury is obtained, and a combination scheme of multimodal images is enumerated based on the preset image type space to generate several initial image combination schemes. Based on the aforementioned initial image combination schemes, the recognition accuracy and recognition time are predicted respectively, and several predicted recognition accuracies and several predicted recognition times are output, including: Randomly select a first initial image combination scheme from the plurality of initial image combination schemes; Based on historical qualified traumatic brain injury identification records, several first historical identification durations of the first initial image combination scheme are collected, and the average of the first historical identification durations is calculated as the first predicted identification duration. A brain injury recognition accuracy predictor is constructed based on deep learning. Based on the basic information, current vital signs data, and historical impact characteristics, the accuracy of injury recognition is predicted for the first initial image combination scheme, and the first predicted recognition accuracy is output. Based on the aforementioned prediction recognition accuracy and prediction recognition duration, and constrained by being greater than the adaptive recognition accuracy and less than the adaptive recognition time limit, the aforementioned initial image combination schemes are screened to obtain multiple qualified image combination schemes. Based on the predicted distribution of traumatic brain injury status, several predicted recognition accuracies, and several predicted recognition durations, with the goal of maximizing recognition accuracy and recognition efficiency, the quality of the multiple qualified image combination schemes is evaluated, and the qualified image combination scheme corresponding to the highest quality coefficient is selected as the optimal image combination scheme. Data is collected from the target patient according to the optimal image combination scheme to obtain a multimodal image set, and the brain injury area of ​​the target patient is identified based on the multimodal image set; Specifically, based on the predicted distribution of traumatic brain injury status, several prediction recognition accuracies, and several prediction recognition durations, and with the goal of maximizing recognition accuracy and efficiency, a scheme quality evaluation is performed on the multiple qualified image combination schemes, including: Several damage influence weights are configured based on several predicted damage intensities, wherein the damage influence weights are positively correlated with the predicted damage intensities; Based on historical qualified traumatic brain injury identification records and a preset image type space, the identification correlation degree is evaluated for several predicted traumatic brain injury types and multiple preset image types, resulting in several identification correlation degree sets. Based on the aforementioned damage impact weights, the aforementioned identification correlation sets are compensated and adjusted to obtain several identification compensation correlation sets; Based on the aforementioned sets of recognition compensation correlation degrees, the recognition compensation correlation degrees corresponding to multiple images included in each qualified image combination scheme are summed to obtain multiple total correlation degrees; Based on the aforementioned predicted recognition accuracy and predicted recognition duration, with the goal of maximizing recognition accuracy and recognition efficiency, the multiple qualified image combination schemes are evaluated for scheme quality, and multiple scheme fitness values ​​are output. The scheme fitness value is positively correlated with the predicted recognition accuracy and negatively correlated with the predicted recognition duration. Based on the multiple total correlation coefficients and multiple scheme fitness, the multiple qualified image combination schemes are evaluated for scheme quality, and multiple scheme quality coefficients are output. Among them, the multiple total correlation coefficients and multiple scheme fitness are processed in dimensionless and weighted to obtain multiple scheme quality coefficients.

2. The method for accurate identification of craniocerebral injury regions based on multimodal imaging according to claim 1, characterized in that, Acquire basic information, current vital signs data, historical impact characteristics, and current clinical scenario of the target patient. The basic information includes age, gender, past medical history, and medication history. The vital signs data includes level of consciousness, pupillary response, motor function, vital signs, and signs of intracranial hypertension. The impact characteristics include impact type and impact velocity. The clinical scenario includes at least scenario type and core clinical needs.

3. The method for accurate identification of craniocerebral injury regions based on multimodal imaging according to claim 1, characterized in that, Based on the aforementioned basic information, current vital signs data, and historical impact characteristics, the state of traumatic brain injury is predicted, and the predicted distribution of traumatic brain injury states is obtained, including: Based on historical qualified traumatic brain injury identification records, a sample basic information set, a sample vital sign data set, and a sample impact feature set are collected. The distribution of the patient's sample traumatic brain injury status under different sample basic information, sample vital sign data, and sample impact features is collected to obtain the sample traumatic brain injury status distribution set. Using the sample basic information set, sample vital signs dataset, and sample impact feature set as inputs, and the sample traumatic brain injury state distribution set as supervision, a deep learning model is trained until convergence to generate a traumatic brain injury state prediction plugin. Using the aforementioned traumatic brain injury status prediction plugin, a predicted distribution of traumatic brain injury status is obtained based on the basic information, current vital signs data, and historical impact characteristics. The predicted distribution of traumatic brain injury status includes several predicted types of traumatic brain injury and several predicted injury intensities.

4. The method for accurate identification of craniocerebral injury regions based on multimodal imaging according to claim 3, characterized in that, The distribution of traumatic brain injury status in patients under different scenarios of collecting basic sample information, vital sign data, and impact characteristics, including: The collection of historical traumatic brain injury type sets and historical traumatic brain injury feature sets for patients in different scenarios of collecting basic information of samples, vital sign data of samples, and impact characteristics of samples, wherein the historical traumatic brain injury types and historical traumatic brain injury features correspond one-to-one. Based on the historical traumatic brain injury feature set, the historical traumatic brain injury intensity set is determined, and the sample traumatic brain injury status distribution is constructed by combining the historical traumatic brain injury type set.

5. The method for accurate identification of craniocerebral injury regions based on multimodal imaging according to claim 1, characterized in that, The preset image type space includes CT images, T1-weighted images, T2-weighted images, diffusion-weighted images, magnetic susceptibility-weighted images, diffusion tensor images, magnetic resonance spectroscopy, liquid attenuation inversion recovery sequence, apparent diffusion coefficient map, and magnetic resonance perfusion-weighted images.

Images