Three-dimensional medical image classification method and system based on deep neural network

Through deep neural network analysis of three-dimensional medical images, extract abnormal feature sequences and generate lesion labels, the problems of inaccurate and consistency of diagnosis results in traditional methods are solved, efficient risk assessment of reversible lesions and irreversible lesions, and improved image analysis Accuracy and comprehensiveness.

CN118968147APending Publication Date: 2024-11-15SHANGHAI BOHAO INFORMATION TECH CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411008737.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Traditional medical imaging analysis methods are time-consuming and labor-intensive, and are easily affected by subjective factors, making it difficult to guarantee the accuracy and consistency of diagnostic results, and it is especially difficult to effectively evaluate the risk of progress of reversible lesions.

Method used

A three-dimensional medical image classification method based on deep neural network is adopted to obtain abnormal image feature sequences, predict the lesion location area and positioning moment, generate initial reversible lesion labels and lesion sustained parameters, and combine pre-trained deep neural network to estimate irreversible lesion risk.

Benefits of technology

It significantly improves the accuracy and efficiency of identifying reversible lesions and potential irreversible lesions in three-dimensional medical imaging, provides the immediate status and development trend of the lesions, helps to formulate more accurate treatment and monitoring plans, and provides scientific basis for early intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968147A_ABST
    Figure CN118968147A_ABST
Patent Text Reader

Abstract

The invention provides a three-dimensional medical image classification method and system based on a deep neural network, and the method can comprehensively capture the dynamic change information of a lesion through precisely extracting an abnormal image feature sequence, including a lesion positioning region and a lesion positioning moment, in a target three-dimensional medical image, and then carries out the prediction of the lesion. The initial reversible lesion label and the lesion continuous parameter are generated, so that the immediate state of the lesion is provided for a medical institution, and the development trend of the lesion is disclosed. Through intelligent screening and arrangement of target reversible lesion labels, a reversible lesion progress vector is formed, the reversible lesion label vector is combined, and a pre-trained target deep neural network is utilized, so that accurate estimation of the irreversible lesion risk is realized, and a scientific basis is provided for early intervention and prevention. Therefore, the accuracy and comprehensiveness of lesion recognition in the three-dimensional medical image are improved, and deep fusion of medical image analysis and deep learning is promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and more specifically, to a three-dimensional medical image classification method and system based on deep neural network. Background Art

[0002] In the field of medical imaging diagnosis, three-dimensional medical images (such as CT, MRI, etc.) have become an important means for medical institutions to diagnose diseases. However, with the continuous development of medical imaging technology, the complexity and diversity of imaging data have increased significantly, and traditional imaging analysis methods can no longer meet the needs of efficient and accurate diagnosis. In particular, for the risk assessment of reversible lesions and their potential progression to irreversible lesions, more sophisticated and intelligent analysis techniques are needed.

[0003] Traditional medical image analysis methods mainly rely on the experience and judgment of medical institutions, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, making it difficult to ensure the accuracy and consistency of diagnostic results. Summary of the invention

[0004] In view of this, the purpose of this application is to provide a three-dimensional medical image classification method and system based on deep neural network.

[0005] According to a first aspect of the present application, a three-dimensional medical image classification method based on a deep neural network is provided, the method comprising:

[0006] Acquire an abnormal image feature sequence corresponding to the target three-dimensional medical image; the abnormal image feature sequence includes a lesion location area and a lesion location time corresponding to a reversible lesion appearing in the target three-dimensional medical image;

[0007] According to each lesion localization area and the corresponding lesion localization time in the abnormal image feature sequence, the lesion localization area is predicted to obtain an initial reversible lesion label and a lesion persistence parameter corresponding to the target three-dimensional medical image;

[0008] Determine a target reversible lesion label from each initial reversible lesion label, arrange each target reversible lesion label according to a lesion persistence parameter, output a reversible lesion progression vector corresponding to the target three-dimensional medical image, and generate a reversible lesion label vector corresponding to the target three-dimensional medical image according to each initial reversible lesion label;

[0009] The reversible lesion progression vector and the reversible lesion label vector are loaded into a target deep neural network for irreversible lesion risk estimation to obtain a target estimation result of the presence of irreversible lesions in the target three-dimensional medical image; the target deep neural network is generated by learning network parameters based on sample reversible lesion progression vectors, sample reversible lesion label vectors and irreversible lesion risk annotation data corresponding to sample three-dimensional medical images.

[0010] According to the second aspect of the present application, a three-dimensional medical imaging diagnosis system is provided, wherein the three-dimensional medical imaging diagnosis system comprises a machine-readable storage medium and a processor, wherein the machine-readable storage medium stores machine-executable instructions, and when the processor executes the machine-executable instructions, the three-dimensional medical imaging diagnosis system implements the aforementioned three-dimensional medical image classification method based on a deep neural network.

[0011] According to the third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed, the aforementioned three-dimensional medical image classification method based on deep neural network is implemented.

[0012] According to any one of the above aspects, the technical effect of the present application is:

[0013] The embodiments of the present application significantly improve the accuracy and efficiency of risk identification of reversible lesions and potential irreversible lesions in three-dimensional medical images. By accurately extracting abnormal image feature sequences in target three-dimensional medical images, including lesion location areas and lesion location moments, the dynamic change information of lesions can be fully captured. Furthermore, using this information to predict lesions, generate initial reversible lesion labels and lesion persistence parameters, not only provides medical institutions with the immediate status of lesions, but also reveals the development trend of lesions. By intelligently screening and arranging target reversible lesion labels, reversible lesion progression vectors are formed, presenting an intuitive lesion progression path to medical institutions, which helps to formulate more accurate treatment and monitoring plans. At the same time, combined with the reversible lesion label vector, using the pre-trained target deep neural network, an accurate estimation of the risk of irreversible lesions is achieved, providing a scientific basis for early intervention and prevention. As a result, the accuracy and comprehensiveness of lesion identification in three-dimensional medical images are improved, and the deep integration of medical image analysis and deep learning is promoted. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 A schematic diagram of the process of a three-dimensional medical image classification method based on a deep neural network provided in an embodiment of the present application is shown;

[0016] Figure 2 A schematic diagram of the component structure of a three-dimensional medical image diagnosis system for implementing the above-mentioned three-dimensional medical image classification method based on deep neural network provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0017] The embodiments of the present application are described below in conjunction with the drawings in the present application. It should be understood that the implementation methods described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0018] It will be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present application refer to that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude the implementation as other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the technical field. It should be understood that when an element is "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element, or may refer to the element and the other element establishing a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling, and the term "and / or" used herein indicates at least one of the items defined by the term, such as "A and / or B" may be implemented as "A", or as "B", or as "A and B".

[0019] In order to make the purpose, technical solution and advantages of the present application clearer, the implementation mode of the present application will be further described in detail with reference to the accompanying drawings. The technical solution of the embodiment of the present application and the technical effect produced by the technical solution of the present application are explained below by describing several exemplary implementation modes. It should be pointed out that the following implementation modes can refer to, draw lessons from or combine with each other, and the same terms, similar features and similar implementation steps in different implementation modes are not described repeatedly.

[0020] Figure 1The flowchart of the 3D medical image classification method and system based on a deep neural network provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the 3D medical image classification method based on a deep neural network of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the 3D medical image classification method based on a deep neural network include:

[0021] Step S110, obtaining an abnormal image feature sequence corresponding to the target three-dimensional medical image, wherein the abnormal image feature sequence includes a lesion location area and a lesion location time corresponding to a reversible lesion appearing in the target three-dimensional medical image.

[0022] In detail, the target 3D medical image refers to the 3D image data obtained by medical imaging equipment (such as CT, MRI, etc.), which has been pre-processed and reconstructed, and provides a 3D view of the patient's internal structure, which helps medical institutions to make accurate diagnosis and analysis. For example, a lung CT scan is a typical target 3D medical image, which contains the 3D structural information of the patient's lungs. Medical institutions can identify abnormal lesions in the lungs, such as nodules, ground-glass opacities, etc., by observing these images.

[0023] The abnormal image feature sequence refers to a feature set related to reversible lesions extracted from the target three-dimensional medical image, which may include the lesion localization area (i.e., the specific location of the lesion in the image) and the localization time of the lesion (i.e., the time point when the lesion was first detected). For example, when processing a lung CT scan, the server can identify a nodule as an abnormal image feature. The localization area of ​​the nodule is its specific location in the CT image (e.g., the left upper lobe), and the localization time is the time point when the nodule first appears in the CT scan (e.g., a certain time point in the first scan).

[0024] The reversible lesions are those that can be restored or improved under certain conditions. Such lesions usually do not pose a long-term threat to the patient's life, but if not treated in time, they can progress to irreversible lesions. For example, pneumonia is a common reversible lesion. With timely and effective treatment, the lung inflammation of pneumonia patients can gradually subside and restore normal lung function. However, if the treatment is not timely or inappropriate, pneumonia can progress to irreversible lesions such as pulmonary fibrosis.

[0025] That is, in this embodiment, the server first receives target three-dimensional medical image data from a medical imaging device. These target three-dimensional medical image data may be pre-processed and reconstructed to ensure their quality and consistency. Then, the server uses advanced image processing algorithms and machine learning techniques to conduct in-depth analysis of the target three-dimensional medical images to identify and extract abnormal image features therein.

[0026] The extraction of abnormal image feature sequences is the key to this step. The server focuses on features related to reversible lesions, such as the location of reversible lesions and the location time of lesions. For example, in a lung CT scan, the server can identify one or more nodules as potential lesion location areas and record the scan time when these nodules first appear as the lesion location time.

[0027] Step S120, predicting the lesion location area according to each lesion location area and the corresponding lesion location time in the abnormal image feature sequence, and obtaining the initial reversible lesion label and lesion persistence parameter corresponding to the target three-dimensional medical image.

[0028] Specifically, the initial reversible lesion label refers to the preliminary judgment result about the lesion type obtained by preliminary analysis of the abnormal image feature sequence. These labels may not be accurate enough and need further verification and correction. The lesion persistence parameter refers to a quantitative indicator that describes the speed and direction of lesion progression. These parameters are usually obtained based on time series analysis algorithms and can provide medical institutions with detailed information about lesion progression.

[0029] That is, in this embodiment, once the abnormal image feature sequence is obtained, the server will use the information in the abnormal image feature sequence to further predict the lesion location area. This process involves estimating the type of lesion, the speed of progression and the possible direction of change.

[0030] To achieve this goal, the server pre-trains a classifier that can classify lesion localization areas based on their key features, such as shape, size, boundary clarity, and texture. For each lesion localization area, the server extracts these key features and uses the feature template of prior medical knowledge to make a preliminary judgment and generate a preliminary judgment result.

[0031] Next, the server uses a time series analysis algorithm (such as a dynamic time warping algorithm or a hidden Markov model algorithm) to analyze the development trend of each lesion location area at different lesion location times. This helps the server understand the progression speed and direction of the lesion and estimate its future development trend.

[0032] Finally, the server calculates the time difference from the appearance of each lesion location area to the current time point, and determines the corresponding lesion persistence parameters based on the development trend analysis results and the patient's clinical indicators. These lesion persistence parameters provide important quantitative information for subsequent lesion risk estimation.

[0033] Step S130, determining a target reversible lesion label from each initial reversible lesion label, arranging each target reversible lesion label according to a lesion persistence parameter, outputting a reversible lesion progression vector corresponding to the target three-dimensional medical image, and generating a reversible lesion label vector corresponding to the target three-dimensional medical image based on each initial reversible lesion label.

[0034] In detail, the target reversible lesion label refers to the label that is most relevant to the target three-dimensional medical image and is screened out from the initial reversible lesion label. These target reversible lesion labels are usually further verified and revised, and have high accuracy and clinical value. For example, when processing a lung CT scan, the server can screen out "early ground-glass changes in lung cancer" from the initial reversible lesion label as the target reversible lesion label, which more accurately describes the characteristics and type of the lesion, helping medical institutions to develop more appropriate treatment plans.

[0035] The reversible lesion progression vector refers to a vector obtained by sorting target reversible lesion labels according to lesion persistence parameters, reflecting the progression order and priority of the lesion from the initial state to the current state, and providing medical institutions with an intuitive representation of the disease progression.

[0036] That is, in this embodiment, after obtaining the initial reversible lesion labels, the server needs to determine the target reversible lesion labels from these initial reversible lesion labels. This process may involve screening and comparing multiple reference reversible lesion labels to select the label that is most relevant to the target three-dimensional medical image.

[0037] Once the target reversible lesion labels are determined, the server will rank them according to the lesion persistence parameter to reflect the progression speed and priority of the lesions. The ranked initial reversible lesion sequence constitutes the reversible lesion progression vector, which provides important information about the direction and speed of lesion progression.

[0038] At the same time, the server will also generate a reversible lesion label vector corresponding to the target three-dimensional medical image based on each initial reversible lesion label. The reversible lesion label vector is a binary vector, each element of which corresponds to a reversible lesion label. If the lesion corresponding to the reversible lesion label appears in the target three-dimensional medical image, the corresponding element is 1, otherwise it is 0.

[0039] Step S140, loading the reversible lesion progression vector and the reversible lesion label vector into a target deep neural network for irreversible lesion risk estimation, and obtaining a target estimation result of irreversible lesions in the target three-dimensional medical image. The target deep neural network is generated by learning network parameters based on sample reversible lesion progression vectors, sample reversible lesion label vectors and irreversible lesion risk annotation data corresponding to sample three-dimensional medical images.

[0040] Finally, the server will load the reversible lesion progression vector and the reversible lesion label vector into the target deep neural network for irreversible lesion risk estimation. The target deep neural network was previously trained with a large number of sample three-dimensional medical images and their corresponding annotated data.

[0041] During the training process, the server used multiple sample 3D medical image sequences, each of which contained multiple sample 3D medical images and their corresponding irreversible lesion risk annotation data and reversible lesion label sequences. By using these data as network learning data, the server was able to train a deep neural network that can accurately estimate the risk of irreversible lesions.

[0042] In actual applications, the server will input the reversible lesion progression vector and reversible lesion label vector of the target three-dimensional medical image into the target deep neural network, and obtain a risk estimation result about the occurrence of irreversible lesions in the target three-dimensional medical image. The risk estimation result can be a probability value, indicating the possibility of the target three-dimensional medical image developing into irreversible lesions in the future.

[0043] In order to further improve the accuracy of risk estimation, the server may also consider using multiple different neural networks for ensemble learning. These neural networks may be generated using different training strategies or network architectures, and they each perform risk estimation on the target three-dimensional medical image and provide a comprehensive risk assessment result.

[0044] Through the above steps, the reversible lesions in the target 3D medical images can be comprehensively and accurately analyzed, and the risk of their development into irreversible lesions can be estimated. This provides important auxiliary diagnostic information for medical institutions, helping them to formulate more appropriate treatment plans and monitoring plans.

[0045] Based on the above steps, the embodiment of the present application significantly improves the accuracy and efficiency of identifying reversible lesions and potential irreversible lesion risks in three-dimensional medical images. By accurately extracting abnormal image feature sequences in target three-dimensional medical images, including lesion location areas and lesion location moments, the dynamic change information of lesions can be fully captured. Further, using this information to predict lesions, generate initial reversible lesion labels and lesion persistence parameters, not only provides medical institutions with the immediate status of lesions, but also reveals the development trend of lesions. By intelligently screening and arranging target reversible lesion labels, reversible lesion progression vectors are formed, presenting an intuitive lesion progression path to medical institutions, which helps to formulate more accurate treatment and monitoring plans. At the same time, combined with the reversible lesion label vector, the pre-trained target deep neural network is used to achieve accurate estimation of the risk of irreversible lesions, providing a scientific basis for early intervention and prevention. As a result, the accuracy and comprehensiveness of lesion identification in three-dimensional medical images are improved, and the deep integration of medical image analysis and deep learning is promoted.

[0046] In a possible implementation, the method further includes:

[0047] Step A110, when the scale of the initial reversible lesion label is smaller than the target scale, the prior probability of irreversible lesions corresponding to the initial reversible lesion label is obtained as the target probability, and the target estimation result of the irreversible lesions appearing in the target three-dimensional medical image is determined according to the target probability. The target probability is obtained by performing a priori analysis on the prior three-dimensional medical images in which irreversible lesions appear in the prior three-dimensional medical images corresponding to the initial reversible lesion label.

[0048] Step A120, when the scale of the initial reversible lesion label is not less than the target scale, execute step S130.

[0049] In this embodiment, when processing a target three-dimensional medical image from a medical imaging device, the server has extracted the abnormal image feature sequence and predicted the initial reversible lesion label and its lesion persistence parameter through steps S110 and S120. However, in this case, the number of the predicted initial reversible lesion labels is small and does not reach the preset target scale, which may be due to unclear lesion features, insufficient data or other factors.

[0050] In detail, the server first calculates the number of predicted initial reversible lesion labels and compares it with the preset target size. If the number of initial reversible lesion labels is less than the target size, the server will adopt another strategy to estimate the risk of irreversible lesions.

[0051] For example, the server queries a priori database that stores the prior probabilities of irreversible lesions associated with the initial reversible lesion labels. These prior probabilities of irreversible lesions are obtained by analyzing a large number of prior 3D medical images with similar initial reversible lesion labels. For example, if the initial reversible lesion label is "micronodule", the server will find all prior 3D medical images labeled as "micronodule" and calculate the proportion of these prior 3D medical images that eventually develop into irreversible lesions (such as lung cancer).

[0052] The server uses the found prior probability of irreversible lesions as the target probability, and directly uses this target probability to estimate the risk of irreversible lesions in the target 3D medical image. For example, if the prior probability of irreversible lesions of "micronodules" is 5%, the server estimates the risk of irreversible lesions in the target 3D medical image to be 5%.

[0053] The server uses the target probability derived based on the prior probability as the target estimation result of the irreversible lesions in the target three-dimensional medical image, and records this result in the diagnosis report for reference by the medical institution.

[0054] In a specific example, suppose that when the server processes a lung CT scan, it predicts only one initial reversible lesion label, "micronodule". The preset target scale is at least two initial reversible lesion labels. Since only one initial reversible lesion label is predicted, the server will query the prior database and find that the prior probability of irreversible lesions for "micronodules" is 5%. Therefore, the server estimates the risk of irreversible lesions in the target 3D medical image to be 5% and reports this result to the medical institution.

[0055] In contrast to the previous example, when the server processes the target 3D medical image, the number of initial reversible lesion labels predicted by the server reaches or exceeds the preset target size, which means that the server has enough lesion information to further analyze the progression of the lesion and the risk of irreversible lesions.

[0056] For example, the server selects the label that is most relevant to the target three-dimensional medical image from the initial reversible lesion label as the target reversible lesion label. This usually involves screening and comparing multiple reference reversible lesion labels to select the label that best matches the current image characteristics. The server arranges the target reversible lesion labels according to lesion persistence parameters (such as the progression speed and change direction of the lesion). The result of the arrangement reflects the progression speed and priority of the lesion, providing an important basis for the subsequent estimation of the risk of irreversible lesions. The server uses the arranged target reversible lesion label sequence as a reversible lesion progression vector, which contains the progression direction and speed information of the lesion. At the same time, the server also generates a reversible lesion label vector based on all initial reversible lesion labels. The vector is a binary vector used to represent all reversible lesion labels that appear in the target three-dimensional medical image.

[0057] The server loads the reversible lesion progression vector and the reversible lesion label vector into a pre-trained target deep neural network, which is trained with a large amount of sample data and can accurately estimate the risk of irreversible lesions. After the server inputs the vector, the network will output a probability value, indicating the possibility that the target 3D medical image will develop into an irreversible lesion in the future.

[0058] The server uses the probability value output by the deep neural network as the target estimation result of the irreversible lesions in the target three-dimensional medical image, and records this result in the diagnosis report for reference by medical institutions.

[0059] For example, suppose that when the server processes a lung CT scan, it predicts two initial reversible lesion labels: "micronodules" and "ground-glass opacity". The number of these two labels reaches the preset target scale. The server first selects the labels that are most relevant to the target three-dimensional medical image as the target reversible lesion labels, assuming that they are all relevant. Then, the server arranges these two labels according to the lesion persistence parameters (such as the growth rate of nodules and the density change of ground-glass opacity). The arranged label sequence constitutes the reversible lesion progression vector. Next, the server generates a reversible lesion label vector, indicating that both labels exist. Finally, the server loads these vectors into the target deep neural network, and the network outputs a probability value, such as 20%, indicating that the target three-dimensional medical image has a 20% risk of developing irreversible lesions (such as lung cancer) in the future. The result will be recorded in the diagnosis report for medical institutions to formulate appropriate treatment and monitoring plans.

[0060] In a possible implementation, the method further includes:

[0061] Step B110, obtaining a three-dimensional medical image risk assessment list.

[0062] Step B120, extracting the risk assessment value corresponding to the target estimation result from the three-dimensional medical image risk assessment list as the target risk assessment value corresponding to the target three-dimensional medical image, and extracting the medical consultation guidance information corresponding to the target risk assessment value as the target medical consultation guidance information corresponding to the target three-dimensional medical image.

[0063] Step B130, obtain the image diagnosis data corresponding to the target three-dimensional medical image, update the target medical consultation guidance information based on the image diagnosis data, obtain the adjusted medical consultation guidance information, and generate the medical consultation report content corresponding to the target three-dimensional medical image based on the target risk assessment value and the adjusted medical consultation guidance information.

[0064] In this embodiment, in the previous steps, the server has completed in-depth analysis of the target three-dimensional medical image, including extracting abnormal image feature sequences, predicting initial reversible lesion labels and lesion persistence parameters, determining target reversible lesion labels and generating reversible lesion progression vectors and label vectors, and estimating the risk of irreversible lesions using deep neural networks. Now, the server will further process this information to generate a detailed medical report to provide comprehensive diagnostic recommendations and treatment guidance for medical institutions.

[0065] In detail, the server first accesses a database or file system that stores 3D medical image risk assessment information. The risk assessment list contains multiple 3D medical image types, corresponding irreversible lesion risk assessment values, and corresponding medical guidance information. This information is based on a large amount of historical data and expert experience, and is used to guide medical institutions to take different diagnosis and treatment measures for patients with different risk levels. For example, the server obtains the entire 3D medical image risk assessment list through an API call or file read operation, and loads it into memory for subsequent processing.

[0066] Based on the irreversible lesion risk estimation result (i.e., the target estimation result) obtained in the previous step, the server searches for the corresponding risk assessment value and medical guidance information in the 3D medical imaging risk assessment list. Since the risk assessment list is organized according to a certain logical structure (such as key-value pairs, tables, etc.), the server can quickly locate the risk assessment entry corresponding to the target estimation result.

[0067] The server traverses the risk assessment list, finds the risk assessment entry that matches the target estimation result, and extracts the target risk assessment value and target medical guidance information from it. For example, if the target estimation result is 20%, the server may find the entry corresponding to "irreversible lesion risk 20%-30%" in the risk assessment list, and extract the risk assessment value and medical guidance information under this entry.

[0068] After extracting the preliminary medical guidance information, the server also needs to consider the specific diagnostic situation of the target 3D medical image. The imaging diagnostic data usually contains a detailed description of the lesion, location, size, morphology and other information, which is crucial for formulating a personalized diagnosis and treatment plan.

[0069] For example, the server retrieves the image diagnosis data corresponding to the target three-dimensional medical image from the database. Then, based on these image diagnosis data, the server makes necessary updates or adjustments to the initially extracted medical guidance information. For example, if the image diagnosis data shows that the lesion is located in a key position in the lungs and has an irregular shape, the server can add medical guidance information about emergency surgery or further examinations.

[0070] After obtaining the target risk assessment value, target medical guidance information, and medical guidance information adjusted according to the imaging diagnosis data, the server integrates this information into a medical report. The medical report is an important basis for medical institutions to understand the patient's condition and formulate a diagnosis and treatment plan.

[0071] For example, the server first creates a medical report template, which contains the basic framework of the report and the content area that needs to be filled. Then, the server fills the target risk assessment value, target medical guidance information, and adjusted medical guidance information into the report template in a certain format. After filling, the server formats and proofreads the report content to ensure the accuracy and readability of the information. Finally, the server saves the generated medical report content in the form of an electronic document or sends it to the workstation of the medical institution.

[0072] In a specific example, suppose that when the server processes a lung CT scan, it obtains an estimated risk of irreversible lesions of 20%. The server first extracts the risk assessment value corresponding to "irreversible lesion risk 20%-30%" and preliminary medical guidance information (such as "regular review is recommended, pay attention to changes in the condition") from the three-dimensional medical imaging risk assessment list. Then, the server retrieves the imaging diagnostic data of the CT scan, showing that the lesion is located at the edge of the lung lobe and has a relatively regular morphology. Based on this information, the server determines that the current situation is relatively stable and no emergency measures are needed immediately, but it still needs to be closely monitored. Therefore, the server adds a reminder to the preliminary medical guidance information: "If symptoms such as coughing and chest pain worsen, please seek medical attention in time." Finally, the server integrates this information into the medical report template and generates a detailed medical report content for reference by medical institutions.

[0073] In a possible implementation, the method further includes:

[0074] Step C110, when the target three-dimensional medical image does not detect any lesion area or the lesion location area prediction of the target three-dimensional medical image is wrong, it is determined that the target three-dimensional medical image belongs to a conventional medical image, and the risk assessment value and medical consultation guidance information corresponding to the conventional medical image are extracted from the three-dimensional medical image risk assessment list to generate a target risk assessment value and target medical consultation guidance information corresponding to the target three-dimensional medical image.

[0075] Step C120, when the scale of the initial reversible lesion label corresponding to the target three-dimensional medical image is smaller than the target scale, it is determined that the target three-dimensional medical image belongs to a single reversible lesion label, and the risk assessment value and medical consultation guidance information corresponding to the single reversible lesion label are extracted from the three-dimensional medical image risk assessment list to generate a target risk assessment value and target medical consultation guidance information corresponding to the target three-dimensional medical image.

[0076] Step C130, when the scale of the initial reversible lesion label corresponding to the target three-dimensional medical image is not less than the target scale, it is determined that the target three-dimensional medical image belongs to a complex reversible lesion label, and the risk assessment value and medical consultation guidance information corresponding to the complex reversible lesion label are extracted from the three-dimensional medical image risk assessment list to generate a target risk assessment value and target medical consultation guidance information corresponding to the target three-dimensional medical image.

[0077] The risk assessment value in the three-dimensional medical image risk assessment list reflects the urgency of the three-dimensional medical image for medical consultation, and the consultation priority of the medical consultation guidance information in the three-dimensional medical image risk assessment list is inversely proportional to the risk assessment value.

[0078] In this embodiment, when processing the target three-dimensional medical image, various situations may be encountered, including no lesion area is detected, the lesion location area prediction is wrong, the number of initial reversible lesion labels is insufficient or reaches a certain scale, etc. These situations require the server to be able to respond flexibly, extract the corresponding risk assessment value and medical consultation guidance information from the three-dimensional medical image risk assessment list according to the specific characteristics of the target three-dimensional medical image, so as to generate accurate medical consultation report content.

[0079] Specifically, after an in-depth analysis of the target three-dimensional medical image, the server may find that the target three-dimensional medical image does not contain any obvious lesion area, or that although the lesion location prediction is performed, the result is wrong (such as misjudging normal tissue as a lesion area).

[0080] Therefore, the server first determines that the target three-dimensional medical image in this case belongs to a conventional medical image, that is, an image without obvious lesions or unreliable lesion prediction. Then, the server searches the risk assessment value and medical guidance information corresponding to the conventional medical image from the three-dimensional medical image risk assessment list. This information is usually more general, emphasizing regular review and attention to daily health management. The server uses the extracted risk assessment value and medical guidance information as the target risk assessment value and target medical guidance information corresponding to the target three-dimensional medical image.

[0081] When predicting the initial reversible lesion labels of the target three-dimensional medical image, the server may find that the number of labels is less than the preset target size, indicating that the characteristics of the lesions in the image are not obvious enough or the lesion type is relatively simple.

[0082] Therefore, the server determines that the target three-dimensional medical image in this case belongs to an image with a single reversible lesion label, that is, the lesion characteristics are relatively simple or not obvious. The server searches the three-dimensional medical image risk assessment list for the risk assessment value and medical guidance information corresponding to the single reversible lesion label. This information may focus more on monitoring and treatment recommendations for specific lesion types. The server integrates the extracted risk assessment value and medical guidance information into the target risk assessment value and target medical guidance information corresponding to the target three-dimensional medical image.

[0083] On the contrary, if the number of initial reversible lesion labels predicted by the server reaches or exceeds the preset target scale, it means that the characteristics of the lesions in the image are more complex or diverse. Therefore, the server determines that the target three-dimensional medical image in this case belongs to an image with a complex reversible lesion label, that is, the lesion characteristics are diverse and may be interrelated. The server searches the three-dimensional medical image risk assessment list for the risk assessment value and medical guidance information corresponding to the complex reversible lesion label. This information may be more comprehensive, covering the monitoring, treatment and comprehensive management recommendations for various types of lesions. Based on the extracted risk assessment value and medical guidance information, the server combines the complexity and diversity of the lesions to generate the target risk assessment value and target medical guidance information corresponding to the target three-dimensional medical image. This information may emphasize the importance of multidisciplinary collaboration, comprehensive treatment and long-term follow-up.

[0084] Regardless of the category of the target 3D medical image, the risk assessment value and the medical guidance information in the 3D medical image risk assessment list are closely related. The risk assessment value reflects the urgency of the 3D medical image, that is, the severity and progression rate of the lesion. The priority of the medical guidance information is inversely proportional to the risk assessment value, that is, the higher the risk assessment value (the higher the urgency), the higher the priority of the medical guidance information (faster and more aggressive diagnosis and treatment measures are required). This correlation ensures that medical institutions can quickly judge the urgency of the disease based on the risk assessment value and take corresponding diagnosis and treatment measures.

[0085] In a specific example, suppose that when the server processes a cardiac MRI scan, it fails to detect any obvious lesion areas in the image. In this case, the server determines that the MRI scan belongs to the category of conventional medical images, and extracts the risk assessment value (such as "low risk") and medical guidance information (such as "annual heart health check is recommended") corresponding to the conventional medical image from the three-dimensional medical image risk assessment list. The server then integrates this information into the target risk assessment value and target medical guidance information corresponding to the target MRI scan, and generates the medical report content for reference by the medical institution.

[0086] Assume that when the server processes another lung CT scan, it predicts multiple initial reversible lesion labels (such as "nodules", "ground glass shadows", etc.), and the number of labels reaches the preset target scale. The server determines that the CT scan belongs to the image category of complex reversible lesion labels, and extracts the risk assessment value corresponding to the complex reversible lesion label from the risk assessment list (such as "medium-high risk") and detailed consultation guidance information (such as "immediate further examination is recommended, and multidisciplinary consultation is considered to formulate a treatment plan"). Finally, the server generates a consultation report containing this information, providing comprehensive diagnosis and treatment recommendations for medical institutions.

[0087] In a possible implementation, step S120 includes:

[0088] Step S121 , using known reversible lesion label data in advance, a target classifier is trained, and the target classifier is used to classify any lesion location area.

[0089] Step S122, for each lesion localization area, extract the key features of the lesion localization area, and based on the extracted key features, use the prior medical knowledge feature template to make a preliminary judgment on the lesion type of the lesion localization area to generate a preliminary judgment result, wherein the key features include shape features, size features, boundary clarity features, and texture features.

[0090] Step S123, arrange the development process of each lesion localization area in a time series based on the lesion localization moment of each lesion localization area, and based on the arrangement sequence, use the time series analysis algorithm to analyze the development trend of each lesion localization area at different lesion localization moments to generate a development trend analysis result for each lesion localization area.

[0091] Step S124, according to the lesion localization moment, the abnormal image feature sequence is divided into different time periods, the abnormal image feature sequences divided into different time periods are input into the target classifier, and an initial reversible lesion label is assigned to each lesion localization area according to the input abnormal image feature sequence through the target classifier.

[0092] Step S125, for each lesion localization area, calculate the time difference from the appearance of the lesion localization area to the current time point, and, based on the development trend analysis results of each lesion localization area, estimate the progression speed and change direction of each lesion localization area, and combine the time difference and the progression speed and change direction of each lesion localization area, as well as the corresponding patient clinical indicators to determine the corresponding lesion persistence parameters.

[0093] When processing the target 3D medical image, the server needs to predict the lesion location area based on the lesion location area and the corresponding lesion location time in the abnormal image feature sequence to obtain the initial reversible lesion label and lesion persistence parameter. This process involves multiple steps, including training of the target classifier, extraction of key features, time series analysis, classification of lesion location areas, and determination of lesion persistence parameters.

[0094] In detail, the server first uses known reversible lesion label data to train a target classifier. These reversible lesion label data usually come from historical imaging data and expert annotations, and contain image features and corresponding labels of various reversible lesion types. The server reads these reversible lesion label data, extracts lesion features in the image, and matches them with corresponding labels. Then, the server uses these feature-label pairs to train a classification model, such as a support vector machine, random forest, or deep learning network. After the training is completed, the server will obtain a classifier that can identify different types of reversible lesions.

[0095] For each lesion location area in the target three-dimensional medical image, the server needs to extract key features and make a preliminary judgment using the prior medical knowledge feature template. The server first identifies the lesion location area in the target three-dimensional medical image, and then extracts the key features of these lesion location areas, including shape features (such as circular, elliptical, etc.), size features (such as diameter, volume, etc.), boundary clarity features (such as clear, blurred, etc.) and texture features (such as uniform, rough, etc.). Next, the server uses a predefined prior medical knowledge feature template to match these features with the features in the template to preliminarily determine the type of lesion. This process will generate a preliminary judgment result, but it may not be accurate enough and requires further verification.

[0096] The server then needs to analyze the development trend of each lesion location area at different lesion location times. For example, the server first arranges the development process of each lesion location area in a time series according to the lesion location time. Then, the server uses a time series analysis algorithm (such as a dynamic time warping algorithm, a hidden Markov model, etc.) to analyze these sequence data to predict the development trend of the lesion at different times. This process generates a development trend analysis result that describes how the size, shape, density and other characteristics of the lesion change over time.

[0097] After obtaining the time series analysis results, the server can input these time series analysis results into the target classifier and assign an initial reversible lesion label to each lesion location area. The server divides the abnormal image feature sequence into different time periods according to the lesion location moment. Then, the server inputs the feature sequence of each time period into the target classifier. The classifier will assign the most likely reversible lesion label to each lesion location area based on the input feature sequence and the model parameters obtained by previous training. This reversible lesion label is the initial reversible lesion label.

[0098] Finally, the server needs to calculate the lesion persistence parameters for each lesion localization area to describe the progression speed and change direction of the lesion. The server first calculates the time difference from the first appearance of each lesion localization area to the current time point. Then, based on the previously obtained development trend analysis results, the server estimates the progression speed and change direction of the lesion. The estimation process may involve calculating the rate of change of lesion size, the rate of morphological change, etc. Next, the server combines these estimates with the time difference and the corresponding patient clinical indicators (such as age, gender, underlying diseases, etc.), and determines the lesion persistence parameter through a certain algorithm (such as weighted summation, regression analysis, etc.). This parameter is a quantitative indicator that integrates the lesion progression speed, change direction and individual patient characteristics, and can provide medical institutions with detailed information on lesion progression.

[0099] In a specific example, suppose that when the server processes a liver CT scan, it finds a suspicious lesion area in the image. The server first extracts key features such as the shape, size, boundary clarity and texture of the area, and uses the feature template of prior medical knowledge to preliminarily judge that it may be a liver cyst. Then, the server arranges the development process of the area in a time series according to the location of the lesion at different scanning time points, and predicts its development trend using the time series analysis algorithm. The prediction results show that the size of the area has a tendency to increase slowly. Then, the server inputs the feature sequences of different time periods into the target classifier, and the classifier assigns an initial reversible lesion label of "liver cyst" to the area. Finally, the server calculates the time difference from the first appearance of the area to the current time point, and combines its progression speed and change direction with clinical indicators such as the patient's age to determine a lesion persistence parameter value, which indicates that the liver cyst progresses slowly, but still needs to be monitored regularly to prevent malignant transformation. The server integrates this information and generates a detailed diagnostic report for reference by medical institutions.

[0100] In a possible implementation, step S123 includes: analyzing the development trend of each lesion localization area at different lesion localization moments based on a dynamic time warping algorithm or a hidden Markov model algorithm, so as to model and predict the development state of the lesion localization area at different lesion localization moments according to the changes in morphological characteristics, size characteristics and / or density characteristics of the lesion localization area at different lesion localization moments, and obtain the change law of the lesion localization area over time, wherein the change law includes the growth rate change law and the morphological change law.

[0101] Step S125 includes:

[0102] Step S1251, based on the change pattern of the lesion localization area over time, estimate the progression speed and change direction of the lesion localization area, the progression speed is obtained by calculating the change amount of the size characteristics, morphological characteristics or density characteristics of the lesion localization area per unit time, and the change direction is determined by analyzing the spatial expansion trend of the lesion localization area.

[0103] Step S1252, after obtaining the progress speed and change direction of the lesion localization area, multiply the time difference by the progress speed as a weight factor to obtain a first value representing the overall development of the lesion localization area, and adjust the first value positively or negatively according to the change direction to generate a second value.

[0104] Step S1253, calculating the corresponding clinical weight factor according to the patient's clinical indicators, and performing weighted calculation on the clinical weight factor and the second value to generate the lesion persistence parameter.

[0105] In this embodiment, when processing the target three-dimensional medical image, the server needs to perform a detailed analysis of the development trend of each lesion location area at different lesion location moments in order to predict its future state and determine the persistence parameters of the lesion. This process involves the application of time series analysis algorithms, such as the dynamic time warping algorithm (DTW) or the hidden Markov model (HMM), and the estimation of the lesion progression speed and change direction based on the analysis results.

[0106] In detail, the server first selects a suitable time series analysis algorithm, such as DTW or HMM, based on the characteristics of the lesion localization area. These algorithms can effectively process time series data and identify patterns and trends in the data. Next, the image data of each lesion localization area at different lesion localization moments is read to extract key information such as morphological features, size features, and density features. The server then applies the selected algorithm (such as DTW) to analyze these feature data to identify the development trend of the lesion at different moments.

[0107] By analyzing the output of the algorithm, the server models the development status of the lesion location area at different times and predicts its future change trend. The server builds a state model of the lesion location area changing over time based on the similarity matrix calculated by the DTW algorithm or the state transition probability estimated by the HMM algorithm. The model describes the changes in the morphological characteristics, size characteristics, and density characteristics of the lesion at different times, as well as the correlation and regularity between them. Based on this model, the server can predict the possible state of the lesion at a certain moment in the future.

[0108] Then, the server calculates the progression rate of the lesion localization area based on the change pattern of the lesion localization area over time. For example, the server calculates the change in size characteristics, morphological characteristics or density characteristics of the lesion localization area per unit time by comparing the image data at different lesion localization moments, and the change reflects the progression rate of the lesion. For example, if the volume of the lesion area increases by 10% per unit time, the progression rate can be expressed as 10% / unit time. In addition to the progression rate, the server also needs to determine the direction of change of the lesion, that is, the spatial expansion trend of the lesion. For example, the server determines the direction of change of the lesion localization area by analyzing the changes in morphology and position of the lesion localization area at different times. For example, if the lesion area shows a trend of spreading to the surrounding area on the image, the direction of change can be described as "expanding outward."

[0109] The server uses the time difference as a weight factor and multiplies it by the progression speed to obtain a first value representing the overall development of the lesion localization area, and adjusts the first value positively or negatively according to the direction of change to generate a second value. For example, the server first calculates the time difference from the first appearance of the lesion localization area to the current time point. Then, this time difference is used as a weight factor and multiplied by the previously calculated progression speed to obtain a weighted progression speed value (first value). If the lesion is expanding outward, the server keeps the first value positive; if the lesion is shrinking inward, it is adjusted to a negative value (generating a second value), which reflects the development of the lesion over the overall time span.

[0110] Finally, the server calculates the corresponding clinical weight factor based on the patient's clinical indicators (such as age, gender, underlying diseases, etc.), and performs weighted calculation on this clinical weight factor and the second value to generate a lesion persistence parameter. For example, the server first queries the database or expert system to obtain the weight factor related to the current patient's clinical indicators. Then, the server multiplies this clinical weight factor by the second value to obtain a lesion persistence parameter value that comprehensively considers the lesion progression speed, change direction and individual characteristics of the patient. This parameter value can provide medical institutions with a quantitative assessment basis for the degree and duration of lesion progression.

[0111] In a specific example, suppose that when the server processes a lung CT scan, it finds a nodular lesion area in the image. The server first applies the DTW algorithm to analyze the morphological characteristics, size characteristics and density characteristics of the nodule at different scanning time points, and constructs a state model that changes over time. By analyzing the output results of the model, the server calculates the volume growth rate of the nodule per unit time as the progression rate (such as 5% / month), and determines that its change direction is outward expansion. Then, the server calculates the time difference from the first appearance of the nodule to the current scan (such as 6 months), and multiplies it with the progression rate as a weight factor to obtain a first value (30%). Since the nodule is expanding outward, the server keeps the first value as a positive value as the second value. Next, the server queries the database to obtain the corresponding clinical weight factor (such as 1.2) based on the patient's age (such as 60 years old) and smoking history (yes), and multiplies this clinical weight factor with the second value to obtain the lesion persistence parameter value (36%), which indicates that the nodule has grown at a faster rate in the past 6 months, and after considering clinical factors such as the patient's age and smoking history, its risk of continued progression is high. The server integrates this information and generates a detailed diagnostic report for reference by medical institutions.

[0112] In a possible implementation, step S130 includes:

[0113] Step S131, from each of the initial reversible lesion labels, an initial reversible lesion label associated with a target reversible lesion label is obtained as the target reversible lesion label. The target reversible lesion label is determined from reference reversible lesion labels in the field of medical imaging.

[0114] Step S132, arranging each of the target reversible lesion labels according to the lesion persistence parameter to obtain an arrangement order corresponding to each of the target reversible lesion labels.

[0115] Step S133, obtaining a progress vector template, loading the arrangement order corresponding to each of the target reversible lesion labels into the progress vector template, and obtaining the reversible lesion progress vector.

[0116] When processing the target 3D medical image, the server has predicted a series of initial reversible lesion labels through the previous steps and calculated the persistence parameters of each lesion. Next, the server needs to filter out the labels most relevant to the target medical image from these initial reversible lesion labels as the target reversible lesion labels, sort these labels according to the persistence parameters of the lesions, and finally generate a reversible lesion progression vector to reflect the progression order and priority of the lesions.

[0117] Specifically, the server first needs to filter out the labels most relevant to the target medical image from the initial reversible lesion labels according to certain rules or standards. These target reversible lesion labels usually come from a reference reversible lesion label set in the field of medical imaging.

[0118] In detail, the server traverses the list of initial reversible lesion labels, and for each label, checks whether it is associated with a reference reversible lesion label in the field of medical imaging. This association may be based on the name, description, features, or other attributes of the label. For example, if the initial reversible lesion label is "ground-glass opacity in the lung", the server can look for similar labels in the reference label set, such as "ground-glass opacity in early lung cancer". Through correlation analysis, the server selects the initial reversible lesion label that best matches the target medical imaging feature as the target reversible lesion label. These labels will be used for subsequent arrangement and generation of progression vectors.

[0119] Once the target reversible lesion labels are determined, the server then needs to sort these labels according to the lesion persistence parameters corresponding to each label. The lesion persistence parameters reflect the progression rate and potential risks of the lesion, so the sorting results are crucial for medical institutions to understand the progression of the disease and formulate treatment plans.

[0120] For example, the server traverses the target reversible lesion label list, and for each target reversible lesion label, extracts its corresponding lesion persistence parameter. Then, the server sorts the labels according to the size of the parameter value (or from small to large or from large to small according to specific needs). For example, if the larger the lesion persistence parameter value, the faster the lesion progresses and the higher the risk, the server will sort the parameter values ​​from large to small. After the sorting is completed, the server assigns an arrangement order to each target reversible lesion label, which will be used for the generation of subsequent progression vectors to reflect the progression order and priority of the lesions.

[0121] Finally, the server needs to use the previously determined target reversible lesion labels and their arrangement order, combined with a predefined progression vector template, to generate a reversible lesion progression vector corresponding to the target three-dimensional medical image. This vector will serve as part of the diagnostic report to provide medical institutions with an intuitive representation of lesion progression.

[0122] For example, the server first loads a predefined progression vector template, which can be a vector or array with a fixed length and structure, used to store the permutation order of the target reversible lesion labels. The server loads the permutation order corresponding to each target reversible lesion label into the progression vector template in sequence. During the loading process, the server may need to make appropriate adjustments or mappings to the permutation order based on the structure and length of the template. After loading, the server generates a reversible lesion progression vector containing the permutation order of all target reversible lesion labels. The reversible lesion progression vector reflects the progression order and priority of the lesion from the initial state to the current state, providing medical institutions with important information about the progression of the disease.

[0123] In a specific example, assume that the server predicts three initial reversible lesion labels when processing a liver MRI scan: "liver cyst", "liver hemangioma" and "focal nodular hyperplasia". Through correlation analysis, the server determines that these three labels are associated with the reference reversible lesion labels in the field of medical imaging, and uses them all as target reversible lesion labels. Next, the server sorts the lesion persistence parameters corresponding to each label (assuming that they have been calculated through the previous steps), and the order of arrangement is "liver hemangioma" (highest parameter value), "focal nodular hyperplasia" (second), and "liver cyst" (lowest). Finally, the server loads a predefined progression vector template and loads the three labels into the template in order of arrangement, generating a reversible lesion progression vector, which indicates that in the liver MRI scan, "liver hemangioma" is the lesion with the fastest progression and the highest risk, "focal nodular hyperplasia" is second, and "liver cyst" is relatively stable. Medical institutions can understand the overall progress of the disease based on this progression vector and formulate corresponding treatment plans.

[0124] In a possible implementation, the lesion localization region is composed of localization feature vectors of multiple image levels, the image accuracy corresponding to the image levels increases layer by layer, and corresponding reference reversible lesion labels appear at each image level.

[0125] The method further includes: obtaining, from various reference reversible lesion labels in the field of medical imaging, reference reversible lesion labels corresponding to various image levels whose image accuracy is greater than a set accuracy as target reversible lesion labels.

[0126] In medical image analysis, the determination of lesion localization areas often involves analysis of multiple image levels, and each image level corresponds to different image accuracy. As the image level goes deeper, the image accuracy increases layer by layer, which can reveal more detailed information about the lesion. At the same time, there are a large number of reference reversible lesion labels in the field of medical imaging, which are associated with features at different image levels. When processing the target three-dimensional medical image, the server needs to select a suitable label from a large number of reference labels as the target reversible lesion label based on the image accuracy requirements.

[0127] In detail, when processing the target 3D medical image, the server first divides it into multiple image levels. Each image level contains a set of positioning feature vectors, which describe the specific location, shape, size and other characteristics of the lesion in the image of that level. The server analyzes the image data layer by layer, extracts the positioning feature vectors of each level, and combines these vectors to form the lesion positioning area. This process may involve image processing technologies such as image segmentation and feature extraction.

[0128] Different image levels correspond to different image accuracies. Generally speaking, as the image level goes deeper, the image resolution and detail expression capabilities gradually improve, and the image accuracy also increases. The server understands the accuracy characteristics of each image level, and this information may be pre-stored in a database or configuration file. When processing an image, the server will select the appropriate image level as needed to achieve the required accuracy requirements.

[0129] There are a large number of reference reversible lesion labels in the field of medical imaging, which are associated with features at different image levels. The server needs to select appropriate labels from the reference labels as target reversible lesion labels according to the accuracy requirements of the target image.

[0130] In detail, the server first sets an image accuracy threshold according to actual needs, which represents the minimum accuracy requirement that the target image needs to achieve. The server traverses all reference reversible lesion labels in the field of medical imaging, and for each label, checks its associated image level and its corresponding image accuracy. If the image level accuracy associated with a label is greater than the set accuracy threshold, the label is used as a candidate target label. Among all candidate target labels, the server may also need to further screen according to other criteria (such as label specificity, sensitivity, clinical relevance, etc.) to finally determine a group or several target reversible lesion labels.

[0131] In a specific example, suppose a server needs to identify and analyze reversible lesions in a brain MRI scan when processing it. MRI scan data is divided into multiple image levels, ranging from low-resolution overviews to high-resolution detail images. The server first sets an image accuracy threshold, requiring the target image to be at least accurate enough to clearly show the edge of the lesion.

[0132] Next, the server starts screening from the reference reversible lesion label library in the field of medical imaging. These labels include but are not limited to "brain ischemic focus", "cerebral white matter lesions", "cerebral infarction", etc., and each label is associated with a specific imaging level and accuracy. The server checks these labels one by one and finds that the imaging level accuracy associated with the "cerebral infarction" label is higher than the set threshold, and the label is clinically highly correlated with the lesion characteristics observed in the current MRI scan.

[0133] Therefore, the server identifies "cerebral infarction" as the target reversible lesion label of the current target 3D medical image. In subsequent analysis and reporting, the server will focus on the lesion characteristics and development trends related to "cerebral infarction" to provide accurate diagnosis suggestions and treatment plans for medical institutions.

[0134] In a possible implementation, step S130 includes:

[0135] Step S1301, determining a target reversible lesion label cluster from a plurality of reference reversible lesion label clusters. The reference reversible lesion label cluster is generated by screening a plurality of reference reversible lesion labels.

[0136] Step S1302: acquiring, from each of the initial reversible lesion labels, an initial reversible lesion label associated with a target reversible lesion label cluster as a target reversible lesion label.

[0137] Step S140 includes: loading the reversible lesion progression vector and the reversible lesion label vector into a target deep neural network corresponding to the target reversible lesion label cluster to obtain a target estimation result of irreversible lesions appearing in the target three-dimensional medical image.

[0138] In this embodiment, when processing the target three-dimensional medical image, the server has predicted a series of initial reversible lesion labels through the previous steps, and generated a reversible lesion progression vector and a reversible lesion label vector. In order to more accurately estimate the risk of irreversible lesions, the server needs to determine a label cluster that is most relevant to the target image from multiple reference reversible lesion label clusters, and filter out the labels associated with the label cluster from the initial reversible lesion labels as the target reversible lesion labels. Subsequently, the server will use the target deep neural network corresponding to the target label cluster to estimate the risk of irreversible lesions.

[0139] In detail, the server maintains a database or knowledge base containing multiple reference reversible lesion label clusters. Each reference reversible lesion label cluster is generated by screening and clustering a large number of reference reversible lesion labels, aiming to group labels with similar characteristics or clinical significance.

[0140] For example, the server first analyzes the predicted initial reversible lesion labels to understand their respective characteristics and clinical significance. Then, the server matches the initial reversible lesion labels with the reference reversible lesion label cluster. The matching process may be based on multiple factors such as the name, description, and similarity between feature vectors of the labels. During the matching process, the server will find the reference reversible lesion label cluster that is most similar or highly correlated with the initial reversible lesion label set as a whole, and determine it as the target reversible lesion label cluster.

[0141] Once the target reversible lesion label cluster is determined, the server needs to filter out labels associated with the label cluster from the initial reversible lesion labels as the target for further analysis and risk estimation.

[0142] For example, the server traverses the initial reversible lesion labels and checks whether each initial reversible lesion label is associated with one or some labels in the target reversible lesion label cluster. This association may be based on multiple criteria such as direct matching of labels, semantic similarity, similarity of feature vectors, etc. Through the association analysis, the server selects the initial reversible lesion labels associated with the target reversible lesion label cluster as the target reversible lesion labels.

[0143] The server has generated a reversible lesion progression vector and a reversible lesion label vector, and determined the target reversible lesion label cluster and its corresponding target deep neural network. Next, the server will use these vectors and neural networks to perform risk estimation of irreversible lesions.

[0144] For example, the server organizes the reversible lesion progression vector and the reversible lesion label vector into the input format required by the neural network. The server loads the target deep neural network corresponding to the target reversible lesion label cluster, which has been trained with a large amount of sample data and can accurately estimate the risk of irreversible lesions. The server loads the prepared input data into the target deep neural network, performs the forward propagation process, and obtains the risk estimation result of irreversible lesions in the target three-dimensional medical image. The server outputs the risk estimation result in the form of probability value or other forms for reference or further processing by medical institutions.

[0145] In a specific example, suppose the server predicts three initial reversible lesion labels, "renal cyst", "renal parenchymal calcification" and "hydronephrosis" when processing a kidney CT scan. The server first accesses the reference reversible lesion label cluster database, and through name matching and feature vector similarity analysis, finds that these three labels are most related to the "benign kidney lesion cluster", so it determines them as the target reversible lesion label cluster.

[0146] Next, the server selects labels that are directly associated with or semantically similar to the "benign renal lesion cluster" from the initial reversible lesion labels, namely "renal cyst" and "renal parenchymal calcification" (assuming that "hydronephrosis" also appears in the initial reversible lesion labels, but is not highly correlated with the cluster) as target reversible lesion labels.

[0147] Finally, the server generated a reversible lesion progression vector and a reversible lesion label vector containing "renal cyst" and "renal parenchymal calcification" and their progression information, and loaded them into the target deep neural network corresponding to the "renal benign lesion cluster". After a series of calculations, the neural network outputs a probability value (such as 5%) indicating the risk of irreversible lesions in the kidney CT scan, and the server records this result in the diagnosis report for reference by medical institutions.

[0148] In a possible implementation, the method further includes:

[0149] Step S101, obtaining a sample three-dimensional medical image sequence. The sample three-dimensional medical image sequence includes a plurality of sample three-dimensional medical images and irreversible lesion risk annotation data corresponding to the sample three-dimensional medical images, and a sample reversible lesion label sequence, wherein the irreversible lesion risk annotation data includes first annotation data and second annotation data, wherein the first annotation data reflects the presence of irreversible lesions in the three-dimensional medical image, and the second annotation data reflects the absence of irreversible lesions in the three-dimensional medical image.

[0150] Step S102, obtaining a basic deep neural network corresponding to each reference reversible lesion label cluster. The network learning data of the basic deep neural network includes a sample reversible lesion label vector corresponding to the sample three-dimensional medical image and a sample reversible lesion progression vector determined according to the corresponding reference reversible lesion label cluster.

[0151] Step S103, generating sample learning data sequences corresponding to each of the reference reversible lesion label clusters based on the sample 3D medical image sequence. The sample learning data sequence includes sample reversible lesion label vectors, sample reversible lesion progression vectors and irreversible lesion risk annotation data corresponding to the sample 3D medical image.

[0152] Step S104, based on the sample learning data sequence corresponding to the same reference reversible lesion label cluster, network parameter learning is performed on the corresponding basic deep neural network to obtain the target deep neural network corresponding to each of the reference reversible lesion label clusters.

[0153] When developing a deep neural network system for estimating the risk of irreversible lesions, the server needs to first obtain a large number of sample 3D medical image sequences, which contain rich lesion information and corresponding annotation data. Then, the server will prepare a basic deep neural network for each cluster based on the reference reversible lesion label cluster, and train these networks with sample data to generate a target deep neural network that can accurately estimate the risk of irreversible lesions.

[0154] In this embodiment, the server obtains sample 3D medical image sequences from a medical image database or a medical institution partner. These sample 3D medical image sequences include various types of 3D medical images, such as CT, MRI, etc., as well as corresponding irreversible lesion risk annotation data and sample reversible lesion label sequences.

[0155] In detail, the server downloads the sample 3D medical image files and related annotation data files through a secure data transmission protocol. The server organizes the collected data to ensure that each sample 3D medical image correctly matches its corresponding irreversible lesion risk annotation data and sample reversible lesion label sequence. The server parses the annotation data file and divides the irreversible lesion risk annotation data into the first annotation data (reflecting the presence of irreversible lesions in the 3D medical image) and the second annotation data (reflecting the absence of irreversible lesions in the 3D medical image).

[0156] The server has predefined multiple reference reversible lesion label clusters and prepared a basic deep neural network framework for each cluster. These basic deep neural network frameworks have not been fully trained and need to use sample data for parameter learning.

[0157] In detail, the server loads the basic deep neural network framework corresponding to each reference reversible lesion label cluster from the model library. Then, the parameters such as weights and biases of each basic deep neural network framework are initialized in preparation for training.

[0158] In order to train the deep neural network, the server needs to convert the sample 3D medical image sequence into a sample data sequence suitable for network learning. The process includes extracting the sample reversible lesion label vector and the sample reversible lesion progression vector, and combining them with the irreversible lesion risk annotation data.

[0159] In detail, the server processes each sample 3D medical image, extracts its reversible lesion features, and generates a sample reversible lesion label vector and a sample reversible lesion progression vector. Next, the sample reversible lesion label vector, sample reversible lesion progression vector, and irreversible lesion risk annotation data corresponding to each sample 3D medical image are combined into a sample learning data unit. Finally, all sample learning data units are combined into a sample learning data sequence in a certain order (such as time order, random order, etc.).

[0160] On this basis, the server uses the generated sample learning data sequence to train the basic deep neural network corresponding to each reference reversible lesion label cluster to generate a target deep neural network that can accurately estimate the risk of irreversible lesions.

[0161] For example, the server processes the sample learning data sequence in batches, each batch containing a certain number of sample data units. Then, each batch of sample data is input into the basic deep neural network, and the forward propagation process is performed to calculate the difference (loss) between the network output and the labeled data. Thus, the back propagation algorithm is executed according to the loss value to adjust the network's weights, biases and other parameters to reduce the loss value.

[0162] Repeat the forward propagation and back propagation process until the training stop condition is met (such as reaching a preset number of iterations, the loss value drops below a preset threshold, etc.).

[0163] After the training is completed, the server saves the trained target deep neural network to the model library for subsequent use.

[0164] In a specific example, suppose that the server is developing a deep neural network system for estimating the risk of irreversible lesions in lung CT scans. The server first collects sample 3D medical image sequences containing lung CT scans from multiple medical institutions. These sample 3D medical image sequences contain annotation data on whether lung cancer (irreversible lesions) is present and sample reversible lesion label sequences (such as nodule size, morphology, etc.).

[0165] Next, the server loaded the basic deep neural network framework prepared for "pulmonary nodule clusters" and initialized the network parameters. Then, the server extracted features from the sample 3D medical images, generated sample reversible lesion label vectors and sample reversible lesion progression vectors, and combined them with irreversible lesion risk annotation data to form a sample learning data sequence.

[0166] Finally, the server iteratively trained the basic deep neural network using the sample learning data sequence. During the training process, the server continuously adjusted the network's weights and bias parameters to reduce the difference between the network output and the labeled data. After multiple iterations, the server successfully trained a target deep neural network that can accurately estimate the risk of irreversible lesions in lung CT scans and saved it in the model library.

[0167] In a possible implementation, step S1301 includes:

[0168] Step S1301-1, obtaining the network performance index of the target deep neural network corresponding to each of the reference reversible lesion label clusters. The network performance index is generated by performing a network performance test on the target deep neural network corresponding to the reference reversible lesion label cluster based on a verification three-dimensional medical image sequence, wherein the verification three-dimensional medical image sequence includes a plurality of verification three-dimensional medical images and verification annotation data corresponding to the verification three-dimensional medical images, and a reversible lesion label verification sequence.

[0169] Step S1301 - 2 : determining the target reversible lesion label cluster from each of the reference reversible lesion label clusters according to the network performance indicator.

[0170] When developing a deep neural network system for estimating the risk of irreversible lesions, the server trains target deep neural networks for multiple reference reversible lesion label clusters. In order to determine which target network corresponding to which label cluster should be used when processing a specific target 3D medical image, the server needs to evaluate the performance of these networks, which involves testing the network with a validation 3D medical image sequence and generating network performance indicators based on the test results, and finally determining the target reversible lesion label cluster from multiple label clusters based on these indicators.

[0171] In detail, the server has trained multiple target deep neural networks corresponding to reference reversible lesion label clusters and is ready to perform performance tests on these target deep neural networks. The performance test requires the use of independent validation 3D medical image sequences, which contain unknown labeled image data to simulate actual application scenarios.

[0172] For example, the server obtains verification three-dimensional medical image sequences from a verification data set or an external data source. These sequences include multiple verification three-dimensional medical images and corresponding verification annotation data and reversible lesion label verification sequences. The verification annotation data is usually manually annotated by medical experts and is used to evaluate the accuracy of the network. Then, the verification three-dimensional medical image sequence is input into the target deep neural network corresponding to each reference reversible lesion label cluster, and the forward propagation process is performed to obtain the network's irreversible lesion risk estimation result for each image. Finally, the network's estimation results are compared with the verification annotation data, and a series of network performance indicators are calculated, such as accuracy, recall, F1 score, AUC value, etc. These indicators reflect the network's ability to identify irreversible lesions.

[0173] After obtaining the performance indicators of the target deep neural network corresponding to all reference reversible lesion label clusters, the server needs to determine an optimal target reversible lesion label cluster from multiple label clusters based on these indicators.

[0174] For example, the server conducts an in-depth analysis of the performance indicators of each network to understand their performance in different aspects. For example, a network may perform well in accuracy but slightly worse in recall, while another network may achieve a balance in F1 score. Then, a comprehensive evaluation of various performance indicators is performed according to the needs of specific application scenarios. If the application scenario places more emphasis on accuracy, the server can choose the network with the highest accuracy; if the ability to identify all potential lesions is more important, the server can choose a network with a higher recall rate. Based on the results of the comprehensive evaluation, the server determines an optimal target reversible lesion label cluster from multiple reference reversible lesion label clusters. The target deep neural network corresponding to this cluster performs best in performance indicators and is most suitable for processing the current target three-dimensional medical image.

[0175] In a specific example, suppose the server is processing a medical image analysis task on liver lesions and has trained target deep neural networks for three reference reversible lesion label clusters, namely "liver cyst cluster", "liver hemangioma cluster" and "liver cancer cluster". In order to determine which target network corresponding to which label cluster should be used when processing a specific liver CT scan, the server performs the following steps:

[0176] A set of validation 3D medical image sequences containing multiple liver CT scans, as well as the corresponding validation annotation data and reversible lesion label validation sequences were obtained from the validation dataset. The server then input these validation data into three target deep neural networks, performed performance tests respectively, and calculated performance indicators such as accuracy, recall, and F1 score for each network.

[0177] After an in-depth analysis of the performance indicators of the three networks, it was found that the target network corresponding to the "liver cancer cluster" performed well in terms of accuracy and F1 score, and the recall rate was also within an acceptable range. Considering that the focus of the current task is to accurately identify potential liver cancer lesions, the server decided to select the "liver cancer cluster" as the target reversible lesion label cluster. Therefore, when processing a specific liver CT scan, the server will use the target deep neural network corresponding to the "liver cancer cluster" for irreversible lesion risk estimation.

[0178] In a possible implementation, the target deep neural network corresponding to the same reference reversible lesion label cluster includes at least one, and each target deep neural network corresponding to the same reference reversible lesion label cluster is generated according to different neural network training strategies.

[0179] The step of loading the reversible lesion progression vector and the reversible lesion label vector into a target deep neural network corresponding to the target reversible lesion label cluster to obtain a target estimation result of irreversible lesions in the target three-dimensional medical image includes:

[0180] Step S141, determining the current deep neural network from each target deep neural network corresponding to the target reversible lesion label cluster according to the network performance indicator.

[0181] Step S142, loading the reversible lesion progression vector and the reversible lesion label vector into the current deep neural network to obtain a target estimation result of irreversible lesions appearing in the target three-dimensional medical image.

[0182] In this embodiment, the server has trained multiple target deep neural networks for a reference reversible lesion label cluster (such as a "lung cancer cluster"). These target deep neural networks use different training strategies to explore the impact of different model architectures and parameter settings on performance. For example, each target deep neural network is trained separately according to a predetermined training scheme. During the training process, the server continuously adjusts the weights and biases of the network to minimize the loss function and improve the prediction accuracy of the network.

[0183] After training is complete, the server needs to evaluate the performance of each target deep neural network to determine how they perform on the validation data.

[0184] For example, the server uses an independent validation three-dimensional medical image sequence (including multiple validation images and their annotated data) to test each network. During the test, the server inputs the validation image into the network to obtain an estimate of the risk of irreversible lesions, compares it with the annotated data, and calculates performance indicators such as accuracy, recall, and F1 score. When processing a specific target three-dimensional medical image, the server needs to select an optimal network from multiple target deep neural networks for risk estimation. For example, the server first analyzes the performance indicators of each network on the validation data to understand their performance in different aspects. According to the needs of the specific application scenario (such as whether accuracy or recall is more important), the server selects one or more networks with the best performance as candidates. If the application scenario has high requirements for model stability, the server can select multiple networks with similar performance for integrated learning. Finally, the server determines one from the candidate networks as the current deep neural network for processing the target three-dimensional medical image.

[0185] After determining the current deep neural network, the server loads the reversible lesion progression vector and reversible lesion label vector of the target three-dimensional medical image into the network to estimate the risk of irreversible lesions. For example, the server first extracts the reversible lesion features from the target three-dimensional medical image to generate a reversible lesion progression vector and a reversible lesion label vector. Next, the server loads these vectors into the current deep neural network and performs a forward propagation process. Finally, the server obtains the output value from the network as the target estimation result of the presence of irreversible lesions in the target three-dimensional medical image. The result can be a probability value indicating the risk of the lesion progressing to an irreversible state.

[0186] In a specific example, suppose the server is processing a medical image analysis task on lung nodules and has trained multiple target deep neural networks for "lung nodule clusters". These target deep neural networks use different initialization methods (such as random initialization, pre-training initialization), optimization algorithms (such as SGD, Adam) and regularization techniques (such as L1 regularization, Dropout).

[0187] When processing a specific lung CT scan, the server first analyzed the performance indicators of these networks on the validation data. After comprehensive evaluation, the server found that the network using pre-trained initialization, Adam optimization algorithm and Dropout regularization performed well in both accuracy and F1 score, so it decided to choose this network as the current deep neural network.

[0188] Next, the server extracted the reversible lesion features of the lung nodules from the lung CT scan, generated the reversible lesion progression vector and the reversible lesion label vector, and loaded them into the current deep neural network. After the network performed the forward propagation process, it output a probability value of 0.25, indicating that the risk of the lung nodules in the CT scan progressing to lung cancer is 25%. The server recorded this result in the diagnosis report for reference by medical institutions.

[0189] Figure 2 A three-dimensional medical imaging diagnosis system 100 provided in an embodiment of the present application is shown, including a processor 1001 and a memory 1003 and a program code stored in the memory 1003. The processor 1001 executes the above program code to implement the steps of a three-dimensional medical image classification method based on a deep neural network.

[0190] Figure 2 The three-dimensional medical imaging diagnosis system 100 shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the three-dimensional medical imaging diagnosis system 100 may also include a transceiver 1004, which can be used for data interaction between the three-dimensional medical imaging diagnosis system and other three-dimensional medical imaging diagnosis systems, such as data transmission and / or data reception. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the three-dimensional medical imaging diagnosis system 100 does not constitute a limitation on the embodiments of the present application.

[0191] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0192] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc.

[0193] The memory 1003 can be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to have or store program code and can be read by a computer, without limitation herein.

[0194] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0195] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0196] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders based on demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenario, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0197] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A three-dimensional medical image classification method based on deep neural network, characterized in that: The method comprises: Acquire an abnormal image feature sequence corresponding to the target three-dimensional medical image; the abnormal image feature sequence includes a lesion location area and a lesion location time corresponding to a reversible lesion appearing in the target three-dimensional medical image; According to each lesion localization area and the corresponding lesion localization time in the abnormal image feature sequence, the lesion localization area is predicted to obtain an initial reversible lesion label and a lesion persistence parameter corresponding to the target three-dimensional medical image; Determine a target reversible lesion label from each initial reversible lesion label, arrange each target reversible lesion label according to a lesion persistence parameter, output a reversible lesion progression vector corresponding to the target three-dimensional medical image, and generate a reversible lesion label vector corresponding to the target three-dimensional medical image according to each initial reversible lesion label; The reversible lesion progression vector and the reversible lesion label vector are loaded into a target deep neural network for irreversible lesion risk estimation to obtain a target estimation result of the presence of irreversible lesions in the target three-dimensional medical image; the target deep neural network is generated by learning network parameters based on sample reversible lesion progression vectors, sample reversible lesion label vectors and irreversible lesion risk annotation data corresponding to sample three-dimensional medical images.

2. The three-dimensional medical image classification method based on deep neural network according to claim 1, characterized in that: The method further comprises: When the scale of the initial reversible lesion label is smaller than the target scale, obtaining the prior probability of irreversible lesions corresponding to the initial reversible lesion label as the target probability, and determining the target estimation result of the irreversible lesions appearing in the target three-dimensional medical image according to the target probability; the target probability is obtained by performing a priori analysis on the prior three-dimensional medical images appearing with irreversible lesions in the prior three-dimensional medical images corresponding to the initial reversible lesion label; When the scale of the initial reversible lesion labels is not less than the target scale, the steps of determining the target reversible lesion labels from the initial reversible lesion labels, arranging the target reversible lesion labels according to the lesion persistence parameter, outputting the reversible lesion progression vector corresponding to the target three-dimensional medical image, and generating the reversible lesion label vector corresponding to the target three-dimensional medical image according to the initial reversible lesion labels are performed.

3. The three-dimensional medical image classification method based on deep neural network according to claim 1 or 2, characterized in that: The method further comprises: Obtain a 3D medical imaging risk assessment list; Extracting, from the three-dimensional medical image risk assessment list, a risk assessment value corresponding to the target estimation result as a target risk assessment value corresponding to the target three-dimensional medical image, and extracting medical consultation guidance information corresponding to the target risk assessment value as the target medical consultation guidance information corresponding to the target three-dimensional medical image; Acquire image diagnostic data corresponding to the target three-dimensional medical image, update the target medical consultation guidance information according to the image diagnostic data, obtain the adjusted medical consultation guidance information, and generate the medical consultation report content corresponding to the target three-dimensional medical image according to the target risk assessment value and the adjusted medical consultation guidance information; Wherein, the method further comprises: When the target three-dimensional medical image does not detect any lesion area or the lesion location area prediction of the target three-dimensional medical image is wrong, determining that the target three-dimensional medical image belongs to a conventional medical image, extracting the risk assessment value and medical consultation guidance information corresponding to the conventional medical image from the three-dimensional medical image risk assessment list to generate a target risk assessment value and target medical consultation guidance information corresponding to the target three-dimensional medical image; When the scale of the initial reversible lesion label corresponding to the target three-dimensional medical image is smaller than the target scale, determining that the target three-dimensional medical image belongs to a single reversible lesion label, extracting the risk assessment value and medical consultation guidance information corresponding to the single reversible lesion label from the three-dimensional medical image risk assessment list to generate a target risk assessment value and target medical consultation guidance information corresponding to the target three-dimensional medical image; When the scale of the initial reversible lesion label corresponding to the target three-dimensional medical image is not less than the target scale, determining that the target three-dimensional medical image belongs to a complex reversible lesion label, extracting the risk assessment value and medical consultation guidance information corresponding to the complex reversible lesion label from the three-dimensional medical image risk assessment list to generate a target risk assessment value and target medical consultation guidance information corresponding to the target three-dimensional medical image; The risk assessment value in the three-dimensional medical image risk assessment list reflects the urgency of the three-dimensional medical image for medical consultation, and the consultation priority of the medical consultation guidance information in the three-dimensional medical image risk assessment list is inversely proportional to the risk assessment value.

4. The three-dimensional medical image classification method based on deep neural network according to claim 1, characterized in that: The step of predicting the lesion location area based on each lesion location area and the corresponding lesion location time in the abnormal image feature sequence to obtain the initial reversible lesion label and lesion persistence parameter corresponding to the target three-dimensional medical image includes: Using known reversible lesion label data in advance, a target classifier is trained, where the target classifier is used to classify any lesion location area; For each lesion localization area, extract key features of the lesion localization area, and based on the extracted key features, use a priori medical knowledge feature template to make a preliminary judgment on the lesion type of the lesion localization area to generate a preliminary judgment result, wherein the key features include shape features, size features, boundary clarity features, and texture features; Arranging the development process of each lesion localization area in a time series based on the lesion localization moment of each lesion localization area, and analyzing the development trend of each lesion localization area at different lesion localization moments using a time series analysis algorithm based on the arrangement sequence, to generate a development trend analysis result of each lesion localization area; According to the lesion localization moment, the abnormal image feature sequence is divided into different time periods, the abnormal image feature sequences divided into different time periods are input into the target classifier, and an initial reversible lesion label is assigned to each lesion localization area according to the input abnormal image feature sequence by the target classifier; For each lesion localization area, the time difference from the appearance of the lesion localization area to the current time point is calculated, and based on the development trend analysis result of each lesion localization area, the progress speed and change direction of each lesion localization area are estimated, and the corresponding lesion persistence parameter is determined by combining the time difference and the progress speed and change direction of each lesion localization area, as well as the corresponding patient clinical indicators; The step of using a time series analysis algorithm to analyze the development trend of each lesion location area at different lesion location moments and generating a development trend analysis result for each lesion location area includes: Based on the dynamic time warping algorithm or the hidden Markov model algorithm, the development trend of each lesion localization area at different lesion localization moments is analyzed, so as to model and predict the development state of the lesion localization area at different lesion localization moments according to the morphological feature changes, size feature changes and / or density feature changes of the lesion localization area at different lesion localization moments, and obtain the change law of the lesion localization area over time, and the change law includes the growth rate change law and the morphological change law; The step of estimating the progression speed and change direction of each lesion localization area based on the development trend analysis result of each lesion localization area, and determining the corresponding lesion persistence parameter in combination with the time difference and the progression speed and change direction of each lesion localization area, and the corresponding patient clinical index, comprises: Based on the change rule of the lesion localization area over time, the progress speed and change direction of the lesion localization area are estimated, the progress speed is obtained by calculating the change amount of the size feature, morphological feature or density feature of the lesion localization area per unit time, and the change direction is determined by analyzing the expansion trend of the lesion localization area in space; After obtaining the progress speed and change direction of the lesion localization area, multiplying the progress speed by the time difference as a weight factor to obtain a first value representing the overall development of the lesion localization area, and performing a positive or negative adjustment on the first value according to the change direction to generate a second value; A corresponding clinical weight factor is calculated according to the patient's clinical indicator, and the clinical weight factor is weightedly calculated with the second value to generate the lesion persistence parameter.

5. The three-dimensional medical image classification method based on deep neural network according to claim 1, characterized in that: The step of determining a target reversible lesion label from each initial reversible lesion label, arranging each target reversible lesion label according to a lesion persistence parameter, and outputting a reversible lesion progression vector corresponding to the target three-dimensional medical image includes: From each of the initial reversible lesion labels, an initial reversible lesion label associated with a target reversible lesion label is obtained as the target reversible lesion label; the target reversible lesion label is determined from reference reversible lesion labels in the field of medical imaging; Arranging the target reversible lesion labels according to the lesion persistence parameter to obtain the arrangement order corresponding to each target reversible lesion label; Acquire a progress vector template, load the arrangement order corresponding to each of the target reversible lesion labels into the progress vector template, and obtain the reversible lesion progress vector; The lesion localization region is composed of localization feature vectors of multiple image levels, the image accuracy corresponding to the image levels increases layer by layer, and corresponding reference reversible lesion labels appear at each image level; The method further comprises: From various reference reversible lesion labels in the field of medical imaging, reference reversible lesion labels corresponding to various image levels whose image accuracy is greater than the set accuracy are obtained as target reversible lesion labels.

6. The three-dimensional medical image classification method based on deep neural network according to claim 1, characterized in that: The step of determining a target reversible lesion label from each initial reversible lesion label comprises: Determining a target reversible lesion label cluster from a plurality of reference reversible lesion label clusters; the reference reversible lesion label cluster is generated by screening a plurality of reference reversible lesion labels; From each of the initial reversible lesion labels, obtaining an initial reversible lesion label associated with a target reversible lesion label cluster as a target reversible lesion label; The step of loading the reversible lesion progression vector and the reversible lesion label vector into a target deep neural network to perform irreversible lesion risk estimation, and obtaining a target estimation result of irreversible lesions in the target three-dimensional medical image, comprises: The reversible lesion progression vector and the reversible lesion label vector are loaded into a target deep neural network corresponding to the target reversible lesion label cluster to obtain a target estimation result of irreversible lesions appearing in the target three-dimensional medical image.

7. The three-dimensional medical image classification method based on deep neural network according to claim 6, characterized in that: The method further comprises: Acquire a sample three-dimensional medical image sequence; the sample three-dimensional medical image sequence includes a plurality of sample three-dimensional medical images and irreversible lesion risk annotation data corresponding to the sample three-dimensional medical images, and a sample reversible lesion label sequence, the irreversible lesion risk annotation data includes first annotation data and second annotation data, the first annotation data reflects that irreversible lesions occur in the three-dimensional medical image, and the second annotation data reflects that irreversible lesions do not occur in the three-dimensional medical image; Obtaining a basic deep neural network corresponding to each reference reversible lesion label cluster; the network learning data of the basic deep neural network includes a sample reversible lesion label vector corresponding to the sample three-dimensional medical image and a sample reversible lesion progression vector determined according to the corresponding reference reversible lesion label cluster; Generate a sample learning data sequence corresponding to each of the reference reversible lesion label clusters according to the sample three-dimensional medical image sequence; the sample learning data sequence includes a sample reversible lesion label vector, a sample reversible lesion progression vector and irreversible lesion risk annotation data corresponding to the sample three-dimensional medical image; According to the sample learning data sequence corresponding to the same reference reversible lesion label cluster, the network parameters of the corresponding basic deep neural network are learned to obtain the target deep neural network corresponding to each of the reference reversible lesion label clusters.

8. The three-dimensional medical image classification method based on deep neural network according to claim 6, characterized in that: The step of determining a target reversible lesion label cluster from a plurality of reference reversible lesion label clusters comprises: Obtaining a network performance index of a target deep neural network corresponding to each of the reference reversible lesion label clusters; the network performance index is generated by performing a network performance test on the target deep neural network corresponding to the reference reversible lesion label cluster based on a verification three-dimensional medical image sequence, wherein the verification three-dimensional medical image sequence includes a plurality of verification three-dimensional medical images and verification annotation data corresponding to the verification three-dimensional medical images, and a reversible lesion label verification sequence; The target reversible lesion label cluster is determined from each of the reference reversible lesion label clusters according to the network performance indicator.

9. The three-dimensional medical image classification method based on deep neural network according to claim 8, characterized in that: The target deep neural network corresponding to the same reference reversible lesion label cluster includes at least one, and each target deep neural network corresponding to the same reference reversible lesion label cluster is generated according to a different neural network training strategy; The step of loading the reversible lesion progression vector and the reversible lesion label vector into a target deep neural network corresponding to the target reversible lesion label cluster to obtain a target estimation result of irreversible lesions in the target three-dimensional medical image includes: Determining a current deep neural network from each target deep neural network corresponding to the target reversible lesion label cluster according to the network performance indicator; The reversible lesion progression vector and the reversible lesion label vector are loaded into the current deep neural network to obtain a target estimation result of irreversible lesions appearing in the target three-dimensional medical image.

10. A three-dimensional medical imaging diagnostic system, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the three-dimensional medical image classification method based on deep neural network described in any one of claims 1 to 9 is implemented.

Citation Information

Cited By

  • Brain multiple sclerosis diseased region change detection method

    CN119850634A

  • Pulmonary nodule three-dimensional quantitative analysis and report generation system based on CT image

    CN121708057A

  • Three-dimensional quantitative analysis and report generation system for pulmonary nodules based on CT images

    CN121708057B