Hemangioma analysis method, system and equipment based on deep learning and medium

Abstract

The invention belongs to the technical field of medical image processing, and discloses a hemangioma analysis method, system and device based on deep learning and a medium. The method comprises the following steps: acquiring a hemangioma image to be analyzed and corresponding clinical data; the hemangioma image is input into a constructed target detection model, a first recognition result is obtained, and the first recognition result comprises a first prediction boundary, a first lesion category and a first confidence coefficient of hemangioma; the first recognition result and the corresponding clinical data are input into a constructed comprehensive analysis model, a second recognition result is obtained, and the second recognition result comprises a second focus category and a second confidence coefficient; and inputting the second identification result into a constructed diagnosis and treatment suggestion generation model to obtain a diagnosis and treatment suggestion. Standard analysis of the hemangioma image is realized, and a more accurate diagnosis basis is provided for doctors.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a hemangioma analysis method, system, device and medium based on deep learning. Background Art

[0002] Hemangioma is a lesion caused by abnormal proliferation of vascular endothelial cells. Hemangiomas are divided into two categories: common hemangiomas and vascular malformation hemangiomas, each of which contains multiple subcategories. For example, common hemangiomas include infantile hemangiomas, congenital hemangiomas, Kaposi-type, plexiform hemangiomas, keratinizing hemangiomas, granulomatous hemangiomas, etc. Vascular malformation hemangiomas include venous malformations, arteriovenous malformations, microvenous malformations, lymphatic malformations, etc. In addition, there are mixed hemangiomas. Early diagnosis is crucial for discovery and diagnosis for effective treatment. At present, the diagnosis of hemangiomas depends on doctors' judgment based on their own experience with the help of imaging examinations such as ultrasound, angiography, and CT. Because the diagnosis and treatment needs of hemangioma diseases are distributed in different departments of medical institutions at all levels, and in hospitals with limited resources, the lack of professional doctors and equipment limits the accuracy of diagnosis. Summary of the invention

[0003] To this end, the embodiments of the present application provide a deep learning-based hemangioma analysis method, system, device and medium, which realize standard analysis of hemangioma images and provide doctors with more accurate diagnostic basis.

[0004] In a first aspect, the present application provides a hemangioma analysis method based on deep learning.

[0005] This application is achieved through the following technical solutions:

[0006] A hemangioma analysis method based on deep learning, comprising:

[0007] Acquire the hemangioma image to be analyzed and the corresponding clinical data;

[0008] Inputting the hemangioma image into a constructed target detection model to obtain a first recognition result, wherein the first recognition result includes a first predicted boundary of the hemangioma, a first lesion category, and a first confidence level, wherein the target detection model is constructed based on a YOLO model;

[0009] Inputting the first recognition result and the corresponding clinical data into a constructed comprehensive analysis model to obtain a second recognition result, wherein the second recognition result includes a second lesion category and a second confidence level, and the comprehensive analysis model is constructed based on a machine learning model;

[0010] The second recognition result is input into a constructed diagnosis and treatment suggestion generation model to obtain a diagnosis and treatment suggestion, wherein the diagnosis and treatment suggestion generation model is constructed based on a large language model.

[0011] In a preferred example of the present application, it can be further configured to include:

[0012] When the second lesion category belongs to the preset lesion category, transmitting the second recognition result and corresponding clinical data to a third-party remote diagnosis terminal;

[0013] Obtain diagnosis and treatment recommendations from third-party remote diagnosis terminals.

[0014] In a preferred example of the present application, it can be further configured that the process of constructing the target detection model includes:

[0015] Acquire multiple categories of hemangioma images and corresponding real annotation information, wherein the real annotation information includes the real boundary and real category of the hemangioma;

[0016] The hemangioma image and the real annotation information are input into the YOLO model for identification, and the predicted boundary, predicted category and predicted confidence of the hemangioma are obtained. The boundary positioning error is obtained based on the predicted boundary and the real boundary, and the category classification error is obtained based on the predicted category and the real category. A loss function is constructed based on the boundary positioning error and the category classification error, and the hyperparameters of the YOLO model are optimized. When the performance of the YOLO model reaches the preset indicators, it is used as a constructed target detection model.

[0017] In a preferred example of the present application, it can be further set that the loss function of the YOLO model is expressed as:

[0018]

[0019] Among them, λ coord Represents the weight coefficient of boundary positioning error, S 2 represents the number of all cells in the grid, B represents the number of bounding boxes, Indicates whether there is an object in the jth bounding box of the i-th grid, x i ,y i 、w i and h i Respectively represent the central abscissa, ordinate, width and height of the predicted bounding box of the i-th grid, represents the central abscissa, ordinate, width and height of the annotation bounding box of the i-th grid, λ noobj Represents the weight coefficient of the target-free confidence loss, C i represents the true confidence, represents the confidence of the prediction, Indicates whether there is no target in the jth bounding box of the i-th grid, p i (c) represents the predicted category label, Represents the true category label.

[0020] In a preferred example of the present application, it can be further configured that before the hemangioma image is input into the constructed target detection model, the method further includes: preprocessing the hemangioma image by any one or more of the following methods:

[0021] Image enhancement, image denoising, color normalization.

[0022] In a preferred example of the present application, it can be further configured that the construction process of the comprehensive analysis model includes:

[0023] Acquire historical clinical data of various types of hemangioma patients, wherein the clinical data includes patient age, patient gender, patient height, patient weight, patient medical history, disease stage, lesion type, and temperature difference between lesion skin and normal skin;

[0024] Inputting the historical clinical data into a machine learning model so that the machine learning model learns the relationship between lesion categories and clinical data;

[0025] Among them, the machine learning model uses any one of decision tree, random forest tree, and XGBoost.

[0026] In a preferred example of the present application, it can be further configured that the diagnosis and treatment suggestion generation model is constructed based on a large language model, including:

[0027] Obtain historical diagnostic suggestions for different categories of hemangiomas, input the hemangioma information and historical diagnostic suggestions into the large language model, and fine-tune the large language model; the large language model adopts the BERT model or the MedGPT model.

[0028] In a second aspect, the present application provides a hemangioma analysis system based on deep learning.

[0029] This application is achieved through the following technical solutions:

[0030] A deep learning-based hemangioma analysis system, used to execute the method described in the first aspect above, the system comprising:

[0031] A data acquisition module, used to acquire the hemangioma image to be analyzed and the corresponding clinical data;

[0032] A preliminary recognition module, used for inputting the hemangioma image into a constructed target detection model to obtain a first recognition result, wherein the first recognition result includes a first predicted boundary of the hemangioma, a first lesion category, and a first confidence level, wherein the target detection model is constructed based on a YOLO model;

[0033] An accurate recognition module, configured to input the first recognition result and the corresponding clinical data into a constructed comprehensive analysis model to obtain a second recognition result, wherein the second recognition result includes a second lesion category and a second confidence level, and the comprehensive analysis model is constructed based on a machine learning model;

[0034] The diagnosis and treatment suggestion generation module is used to input the second recognition result into the constructed diagnosis and treatment suggestion generation model to obtain diagnosis and treatment suggestions, wherein the diagnosis and treatment suggestion generation model is constructed based on the large language model.

[0035] On the third aspect, the present application is realized through the following technical solutions:

[0036] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the above-mentioned deep learning-based hemangioma analysis methods when executing the computer program.

[0037] In a fourth aspect, the present application provides a computer-readable storage medium.

[0038] This application is achieved through the following technical solutions:

[0039] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of any one of the above-mentioned hemangioma analysis methods based on deep learning.

[0040] In summary, compared with the prior art, the technical solution provided in the embodiment of the present application has at least the following beneficial effects:

[0041] This application obtains the hemangioma image to be analyzed and the corresponding clinical data; inputs the hemangioma image into the constructed target detection model to obtain a first recognition result, the first recognition result includes the first predicted boundary of the hemangioma, the first lesion category and the first confidence; inputs the first recognition result and the corresponding clinical data into the constructed comprehensive analysis model to obtain a second recognition result, the second recognition result includes the second lesion category and the second confidence; inputs the second recognition result into the constructed diagnosis and treatment suggestion generation model to obtain diagnosis and treatment suggestions. This application uses a pre-trained target detection model to perform preliminary recognition and diagnosis on the image data to obtain a preliminary first recognition result; then uses the pre-trained comprehensive analysis model, and combines the corresponding clinical data for further accurate classification to obtain a second recognition result, further ensuring the accuracy of the recognition result; finally, a diagnosis and treatment result is obtained by using the diagnosis and treatment suggestion generation model. The automated image recognition and analysis process reduces the dependence on professional doctors, significantly reduces the inconsistency caused by differences in human judgment, and realizes the standardization of the pre-diagnosis process; diagnostic suggestions can be obtained quickly to improve the efficiency of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of a flow chart of a hemangioma analysis method based on deep learning provided in one embodiment of the present application;

[0043] Figure 2 A schematic diagram of the structure of a hemangioma analysis system based on deep learning provided in one embodiment of the present application;

[0044] Figure 3 A schematic diagram of the structure of a hemangioma analysis system based on deep learning provided in one embodiment of the present application;

[0045] Description of reference numerals:

[0046] Data acquisition module-01, preliminary identification module-02, accurate identification module-03, diagnosis and treatment suggestion generation module-04, remote diagnosis and treatment suggestion module-05. DETAILED DESCRIPTION

[0047] This specific embodiment is merely an explanation of the present application and is not a limitation of the present application. After reading this specification, those skilled in the art may make modifications to the present embodiment without any creative contribution as needed, but such modifications are protected by the patent law as long as they are within the scope of the claims of the present application.

[0048] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0049] In addition, the term "and / or" in this application is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this application, unless otherwise specified, generally indicates that the associated objects before and after are in an "or" relationship.

[0050] In this application, the terms "first", "second", etc. are used to distinguish identical or similar items with basically the same effects and functions. It should be understood that there is no logical or temporal dependency between "first", "second", and "nth", nor is there any limitation on quantity and execution order.

[0051] The following is a further detailed description of the embodiments of the present application in conjunction with the accompanying drawings. Figure 1 As shown, the method steps of the present application include:

[0052] S10: Acquire the hemangioma image to be analyzed and the corresponding clinical data;

[0053] The image of the lesion on the surface of the patient's hemangioma is collected by a high-definition camera, and the collected image is stored in the image database according to the patient ID. The resolution of the image is generally not less than 640×640 pixels, and the image is ensured not to be out of focus. At the same time, the clinical data corresponding to the patient's hemangioma image is also obtained. The clinical data at least includes the patient's age, the lesion area in the image, and the temperature difference between the lesion area and the surrounding skin, but the clinical data does not include the data of imaging examinations. In the actual implementation process, after obtaining the clinical data, the clinical data needs to be preprocessed, specifically, the clinical data is cleaned and standardized, such as filling missing values, deleting duplicate values, and correcting obviously erroneous data; normalizing the data, etc.

[0054] S20: Input the hemangioma image into the constructed target detection model to obtain a first recognition result. The first recognition result includes a first predicted boundary of the hemangioma, a first lesion category, and a first confidence level. The target detection model is constructed based on the YOLO model.

[0055] The hemangioma image obtained in step S10 is input into the constructed target detection model, and the target detection model outputs the first recognition result of the hemangioma in the hemangioma image, and the first recognition result includes the location, category and probability of the lesion in the hemangioma image. Specifically, the first predicted boundary in the first recognition result is a bounding box predicted to contain the lesion area, and the first predicted boundary is in the form of (x, y, w, h), where x is the horizontal coordinate of the bounding box, y is the vertical coordinate of the bounding box, w is the width of the bounding box, and h is the height of the bounding box. The first lesion category is the estimated lesion category of the hemangioma, and the first confidence is the category confidence, which indicates the confidence that the hemangioma belongs to each category. The category confidence is equal to the product of the confidence and the category probability, where the confidence indicates the probability of whether a certain object is contained in the bounding box, and the category probability indicates the probability that the object belongs to each category.

[0056] S30: Input the first recognition result and the corresponding clinical data into the constructed comprehensive analysis model to obtain a second recognition result, wherein the second recognition result includes a second lesion category and a second confidence level, and the comprehensive analysis model is constructed based on the machine learning model.

[0057] The constructed comprehensive analysis model is used to analyze and judge the first recognition result based on the first recognition result and combined with clinical data to obtain a second recognition result of the hemangioma in the hemangioma image. The first lesion category is the final lesion category of the hemangioma, and the second confidence is the final adjusted category confidence of the hemangioma.

[0058] S40: Input the second recognition result into the constructed diagnosis and treatment recommendation generation model to obtain diagnosis and treatment recommendations. The diagnosis and treatment recommendation generation model is constructed based on the large language model.

[0059] This application uses a pre-trained target detection model to perform preliminary recognition and diagnosis on image data to obtain a preliminary first recognition result; then uses a pre-trained comprehensive analysis model and combines it with the corresponding clinical data for further accurate classification to obtain a second recognition result, further ensuring the accuracy of the recognition result; finally, a diagnosis and treatment result is obtained by using a diagnosis and treatment recommendation generation model. The automated image recognition and analysis process reduces dependence on professional doctors, significantly reduces the inconsistency caused by differences in human judgment, and achieves standardization of the pre-diagnosis process; diagnostic recommendations can be obtained quickly, improving the efficiency of diagnosis.

[0060] In some embodiments, when the method of the present application obtains the second recognition result, it also includes: when the second lesion category belongs to the preset lesion category, transmitting the second recognition result and the corresponding clinical data to the third-party remote diagnosis terminal; obtaining the diagnosis and treatment suggestions of the third-party remote diagnosis terminal. Specifically, when the final lesion category of the hemangioma obtained by the comprehensive analysis model belongs to the preset lesion category, the preset lesion category is a specific complex hemangioma category. At this time, it is necessary to transmit the second recognition result and the corresponding clinical data to the third-party remote diagnosis terminal to obtain the diagnosis and treatment suggestions of the third-party remote diagnosis terminal. In the actual implementation process, the preset lesion category can be set according to actual needs. After the data is transmitted to the third-party remote diagnosis terminal, it is analyzed and judged by experts in this field and the diagnosis and treatment suggestions are uploaded.

[0061] In some embodiments, data quality is ensured so that it can be subsequently input into the target detection model for analysis. Before the hemangioma image is input into the constructed target detection model, it also includes: preprocessing the hemangioma image in any one or more of the following ways: image enhancement, image denoising, and color normalization. Specifically, image enhancement can be performed by contrast enhancement, brightness adjustment, rotation, scaling, or translation through the contrast of the feature image. Image noise is reduced by spatial domain filtering and frequency domain filtering. Color normalization is performed by adjusting the color balance of the image to improve the generalization ability of the model.

[0062] In some embodiments, the construction process of the target detection model includes: obtaining multiple categories of hemangioma images and corresponding real annotation information, the real annotation information includes the real boundary and real category of the hemangioma. The network structure of the YOLO model includes an input layer, multiple convolutional layers and pooling layers, a fully connected layer, etc., which are used to extract image features. The hemangioma image and the real annotation information are input into the basic YOLO model for recognition, and the predicted boundary, predicted category and predicted confidence of the hemangioma are obtained. The boundary positioning error is obtained based on the predicted boundary and the real boundary, and the category classification error is obtained based on the predicted category and the real category. The loss function is constructed based on the boundary positioning error and the category classification error, and the hyperparameters of the YOLO model are optimized. When the performance of the YOLO model reaches the preset index, it is used as a constructed target detection model.

[0063] Among them, the loss function of the YOLO model is expressed as:

[0064]

[0065] Among them, λ coord Represents the weight coefficient of boundary positioning error, S 2 represents the number of all cells in the grid, B represents the number of bounding boxes, Indicates whether there is an object in the jth bounding box of the i-th grid. If there is an object The value is 1 when there is no target The value of x is 0, i ,y i 、w i and h i Respectively represent the central abscissa, ordinate, width and height of the predicted bounding box of the i-th grid, represents the central abscissa, ordinate, width and height of the annotation bounding box of the i-th grid, λ noobj represents the weight coefficient of the targetless confidence loss, C represents the confidence, C i represents the true confidence, represents the confidence of the prediction, p(c) represents the category probability, and p i (c) represents the predicted category label, represents the true category label, Indicates whether there is no target in the jth bounding box of the i-th grid. If there is no target, The value is 1 when there is a target The value of is 0.

[0066] Specifically, a large number of hemangioma images containing multiple hemangioma categories are obtained from public data sets, and the above-mentioned hemangioma images cover hemangiomas of various categories, different stages and different parts. And obtain the real annotation information obtained by annotating these hemangioma images using annotation software. Among them, the real annotation information includes the real boundary and real category of the hemangioma in the hemangioma image, and can also include the size information of the hemangioma. After the hemangioma image and the real annotation information are paired, a data set is formed, and the data set is divided into a training set, a validation set and a test set according to a certain ratio. The data in the training set is input into the YOLO model, and the YOLO model is trained. The YOLO model learns the patterns, features and laws of the data input into the training set to obtain the predicted boundary, predicted category and predicted confidence of the hemangioma. Based on the boundary positioning error and the category classification error, the hyperparameters of the YOLO model are optimized, the validation set is used to evaluate the performance of the model on unseen data, and the test set is used to finally evaluate the performance of the model. When the performance of the YOLO model reaches the preset index, it is used as a constructed target detection model.

[0067] Among them, the principle of the YOLO model for target detection on the image is: the YOLO model first divides the input hemangioma image into S×S grids, and each grid is responsible for detecting whether the target "hemangioma" exists in the grid. If the center of the target falls within a certain grid, the grid is responsible for detecting the target. The size of the grid can be selected as S=7~S=14. When the area covered by the grid is large, the fine-grained detection capability is weak, which may lead to a decrease in detection accuracy; when the area covered by the grid is small, the fine-grained detection capability is enhanced, but the amount of calculation increases, which may lead to a longer detection time. At the same time, there are B bounding boxes and their corresponding confidence levels in each grid. In this application, the value of B is 2~5, which can ensure that while capturing more targets, it can reduce target omissions in complex scenes. The confidence reflects the accuracy of the target and whether the bounding box does contain the target. The formula for confidence is expressed as:

[0068]

[0069] P(Object) represents the probability of the existence of an object in a grid cell. Represents the intersection-over-union ratio of the predicted bounding box and the true box.

[0070] Each grid predicts the probability of C categories, which means the probability of the target belonging to each category when there is a target in the grid. The final output is the category probability of each bounding box multiplied by the confidence, which means the confidence that the bounding box contains a certain category of target, expressed as:

[0071] Among them, N i represents the number of pixels belonging to category i in the grid unit, and N represents the total number of pixels in the grid unit. The output formula is as follows:

[0072] Output=P(Class i |Object)×Confidece,

[0073] In the specific implementation process, the performance of the model is evaluated through the following specific indicators: accuracy, recall, and F1 value. Among them, accuracy refers to the proportion of samples correctly classified by the model in the test set to the total number of samples; recall refers to the proportion of samples correctly identified as positive examples by the model among all samples that are actually positive examples; F1 value (F1Score) refers to the harmonic mean of accuracy and recall, which is used to measure the overall performance of the model, especially in the case of class imbalance. Usually, these indicators reach the preset thresholds, such as accuracy reaching more than 90%, recall rate and F1 value reaching more than 85%. In several consecutive trainings, the fluctuations of various indicators are within an acceptable range, indicating that the model performance is stable.

[0074] In some embodiments, the process of constructing a comprehensive analysis model includes: obtaining historical clinical data of various categories of hemangioma patients, the clinical data including patient age, patient gender, patient height, weight, patient medical history, disease stage, lesion category, and temperature difference between lesion skin and normal skin; inputting the historical clinical data into a machine learning model so that the machine learning model learns the relationship between lesion category and clinical data; wherein the machine learning model adopts any one of a decision tree, a random forest tree, and XGBoost.

[0075] The decision tree model represents the decision process through a tree structure and can handle classification and regression problems. Its core is information gain, which represents the purity improvement brought by selecting a certain feature for splitting. When the decision tree model is selected as the machine learning model in this application, the information gain formula is set as follows:

[0076]

[0077] H(D) is the entropy of the data set D, |D| is the number of samples in the data set, and D v is the subset of feature A whose value is v, and Values(A) is all possible values ​​of feature A.

[0078] Random forest improves the accuracy and stability of the model by integrating multiple decision trees. Its core is bagging and random feature selection. When the random forest tree model is selected as the machine learning model in this application, the formula is set as:

[0079]

[0080] Where N is the number of decision trees, h i (x) is the prediction result of the i-th decision tree for sample x.

[0081] XGBoost uses gradient boosting trees to train models, which can effectively handle the complex relationship between features and labels. When XGBoost is selected as the machine learning model in this application, the objective function of XGBoost is set to:

[0082]

[0083] Where n is the number of samples, l is the loss function, Ω is the regularization term, and f k is the kth tree, and K is the number of trees.

[0084] In some embodiments, the diagnosis and treatment recommendation generation model is constructed based on a large language model, including: obtaining historical diagnosis recommendations for different categories of hemangiomas, inputting hemangioma information and historical diagnosis recommendations into the large language model, and fine-tuning the large language model; wherein the large language model adopts a BERT model or a MedGPT model. By pre-training and fine-tuning the large language model, efficient medical text understanding and generation are achieved to provide auxiliary treatment recommendations. Specifically, the large language model is pre-trained, and language comprehension capabilities are acquired by pre-training the masked language model task on a large-scale medical text dataset. The masked language model loss function is as follows:

[0085]

[0086] Where T is the sequence length, x t is the masked word, x -t It's some other word.

[0087] Furthermore, the large language model is fine-tuned based on the preset treatment recommendation data, so that the model can generate treatment recommendations for specific hemangioma types. The fine-tuning loss function is:

[0088]

[0089] Where N is the number of samples, x i is the input text, y i is the generated treatment recommendation.

[0090] In another embodiment of the present application, a deep learning-based hemangioma analysis system is provided, which is used to perform the above-mentioned analysis method, such as Figure 2 As shown, the analysis system includes: a data acquisition module 01, a preliminary identification module 02, an accurate identification module 03, and a diagnosis and treatment suggestion generation module 04;

[0091] Data acquisition module 01, used to acquire the hemangioma image to be analyzed and the corresponding clinical data;

[0092] A preliminary recognition module 02 is used to input the hemangioma image into a constructed target detection model to obtain a first recognition result, wherein the first recognition result includes a first predicted boundary of the hemangioma, a first lesion category, and a first confidence level, wherein the target detection model is constructed based on a YOLO model;

[0093] An accurate identification module 03, used for inputting the first identification result and the corresponding clinical data into a constructed comprehensive analysis model to obtain a second identification result, wherein the second identification result includes a second lesion category and a second confidence level, and the comprehensive analysis model is constructed based on a machine learning model;

[0094] The diagnosis and treatment suggestion generation module 04 is used to input the second recognition result into the constructed diagnosis and treatment suggestion generation model to obtain diagnosis and treatment suggestions, wherein the diagnosis and treatment suggestion generation model is constructed based on the large language model.

[0095] In some embodiments, Figure 3 As shown, the analysis system of the present application also includes: a remote diagnosis and treatment recommendation module 05, which is used to transmit the second identification result and the corresponding clinical data to a third-party remote diagnosis terminal when the second lesion category belongs to a preset lesion category, and obtain the diagnosis and treatment recommendations of the third-party remote diagnosis terminal.

[0096] In some embodiments, the data acquisition module 01 also includes data preprocessing, which is used to preprocess the hemangioma image: image enhancement, image denoising, and color normalization.

[0097] In some embodiments, the preliminary recognition module 02 further includes a YOLO model training unit, which is used to train the YOLO model and construct a target detection model. The specific training process is: obtaining multiple categories of hemangioma images and corresponding real annotation information, where the real annotation information includes the real boundary and real category of the hemangioma;

[0098] The hemangioma image and the real annotation information are input into the YOLO model for identification, and the predicted boundary, predicted category and predicted confidence of the hemangioma are obtained. The boundary positioning error is obtained based on the predicted boundary and the real boundary, and the category classification error is obtained based on the predicted category and the real category. A loss function is constructed based on the boundary positioning error and the category classification error, and the hyperparameters of the YOLO model are optimized. When the performance of the YOLO model reaches the preset indicators, it is used as a constructed target detection model.

[0099] In some embodiments, the precise identification module 03 also includes a machine learning model training unit for obtaining historical clinical data of various categories of hemangioma patients, the clinical data including patient age, patient gender, patient height, weight, patient medical history, disease stage, lesion category, and temperature difference between lesion skin and normal skin; the historical clinical data is input into the machine learning model so that the machine learning model learns the relationship between lesion category and clinical data; wherein the machine learning model adopts any one of decision tree, random forest tree, and XGBoost.

[0100] The specific definition of the hemangioma analysis system based on deep learning provided in this embodiment can be found in the embodiment of the hemangioma analysis method based on deep learning above, and will not be repeated here. Each module in the above-mentioned hemangioma analysis system based on deep learning can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0101] An embodiment of the present application provides a computer device, which may include a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the processor executes the steps of the hemangioma analysis method based on deep learning as in any of the above embodiments.

[0102] The working process, working details and technical effects of the computer device provided in this embodiment can be found in the embodiment of the hemangioma analysis method based on deep learning above, and will not be described in detail here.

[0103] The embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the hemangioma analysis method based on deep learning as in any of the above embodiments are implemented. The computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive and / or a memory stick, etc. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0104] The working process, working details and technical effects of the computer-readable storage medium provided in this embodiment can be found in the above embodiment of the method for tracking cranial and brain lesions in ultrasound images, and will not be described in detail here.

[0105] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0106] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system described in the present application is divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A hemangioma analysis method based on deep learning, It is characterized in that include: Acquire the hemangioma image to be analyzed and the corresponding clinical data; Inputting the hemangioma image into a constructed target detection model to obtain a first recognition result, wherein the first recognition result includes a first predicted boundary of the hemangioma, a first lesion category, and a first confidence level, wherein the target detection model is constructed based on a YOLO model; Inputting the first recognition result and the corresponding clinical data into a constructed comprehensive analysis model to obtain a second recognition result, wherein the second recognition result includes a second lesion category and a second confidence level, and the comprehensive analysis model is constructed based on a machine learning model; The second recognition result is input into a constructed diagnosis and treatment suggestion generation model to obtain a diagnosis and treatment suggestion, wherein the diagnosis and treatment suggestion generation model is constructed based on a large language model.

2. The hemangioma analysis method based on deep learning according to claim 1, It is characterized in that Also includes: When the second lesion category belongs to the preset lesion category, transmitting the second recognition result and corresponding clinical data to a third-party remote diagnosis terminal; Obtain diagnosis and treatment recommendations from third-party remote diagnosis terminals.

3. The hemangioma analysis method based on deep learning according to claim 1, It is characterized in that The process of building a target detection model includes: Acquire multiple categories of hemangioma images and corresponding real annotation information, wherein the real annotation information includes the real boundary and real category of the hemangioma; The hemangioma image and the real annotation information are input into the YOLO model for identification, and the predicted boundary, predicted category and predicted confidence of the hemangioma are obtained. The boundary positioning error is obtained based on the predicted boundary and the real boundary, and the category classification error is obtained based on the predicted category and the real category. A loss function is constructed based on the boundary positioning error and the category classification error, and the hyperparameters of the YOLO model are optimized. When the performance of the YOLO model reaches the preset indicators, it is used as a constructed target detection model.

4. The hemangioma analysis method based on deep learning according to claim 3, It is characterized in that The loss function of the YOLO model is expressed as: Among them, λ coord Represents the weight coefficient of boundary positioning error, S 2 represents the number of all cells in the grid, B represents the number of bounding boxes, Indicates whether there is an object in the jth bounding box of the i-th grid, x i ,y i 、w i and h i Respectively represent the central abscissa, ordinate, width and height of the predicted bounding box of the i-th grid, represents the central abscissa, ordinate, width and height of the annotation bounding box of the i-th grid, λ noobj Represents the weight coefficient of the target-free confidence loss, C i represents the true confidence, represents the confidence of the prediction, Indicates whether there is no target in the jth bounding box of the i-th grid, p i (c) represents the predicted category label, Represents the true category label.

5. The hemangioma analysis method based on deep learning according to claim 1, It is characterized in that Before the hemangioma image is input into the constructed target detection model, the method further includes: preprocessing the hemangioma image by any one or more of the following methods: Image enhancement, image denoising, color normalization.

6. The hemangioma analysis method based on deep learning according to claim 1, It is characterized in that The construction process of the comprehensive analysis model includes: Acquire historical clinical data of various types of hemangioma patients, wherein the clinical data includes patient age, patient gender, patient height, patient weight, patient medical history, disease stage, lesion type, and temperature difference between lesion skin and normal skin; Inputting the historical clinical data into a machine learning model so that the machine learning model learns the relationship between lesion categories and clinical data; Among them, the machine learning model uses any one of decision tree, random forest tree, and XGBoost.

7. The hemangioma analysis method based on deep learning according to claim 1, It is characterized in that The diagnosis and treatment recommendation generation model is constructed based on a large language model, including: Obtain historical diagnostic suggestions for different categories of hemangiomas, input the hemangioma information and historical diagnostic suggestions into the large language model, and fine-tune the large language model; the large language model adopts the BERT model or the MedGPT model.

8. A hemangioma analysis system based on deep learning, It is characterized in that Used to perform the method according to any one of claims 1 to 7, comprising: A data acquisition module, used to acquire the hemangioma image to be analyzed and the corresponding clinical data; A preliminary recognition module, used for inputting the hemangioma image into a constructed target detection model to obtain a first recognition result, wherein the first recognition result includes a first predicted boundary of the hemangioma, a first lesion category, and a first confidence level, wherein the target detection model is constructed based on a YOLO model; An accurate recognition module, configured to input the first recognition result and the corresponding clinical data into a constructed comprehensive analysis model to obtain a second recognition result, wherein the second recognition result includes a second lesion category and a second confidence level, and the comprehensive analysis model is constructed based on a machine learning model; The diagnosis and treatment suggestion generation module is used to input the second recognition result into the constructed diagnosis and treatment suggestion generation model to obtain diagnosis and treatment suggestions, wherein the diagnosis and treatment suggestion generation model is constructed based on the large language model.

9. A computer device, It is characterized in that The method comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images