CN116452851A - Training method and device for disease classification model, terminal and readable storage medium - Google Patents
Training method and device for disease classification model, terminal and readable storage medium Download PDFInfo
- Publication number
- CN116452851A CN116452851A CN202310260271.XA CN202310260271A CN116452851A CN 116452851 A CN116452851 A CN 116452851A CN 202310260271 A CN202310260271 A CN 202310260271A CN 116452851 A CN116452851 A CN 116452851A
- Authority
- CN
- China
- Prior art keywords
- feature
- classification model
- data
- network
- inputting
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012549 training Methods 0.000 title claims abstract description 170
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 title claims abstract description 151
- 238000013145 classification model Methods 0.000 title claims abstract description 145
- 201000010099 disease Diseases 0.000 title claims abstract description 56
- 238000000034 method Methods 0.000 title claims abstract description 55
- 239000013598 vector Substances 0.000 claims abstract description 145
- 238000000605 extraction Methods 0.000 claims abstract description 68
- 238000012216 screening Methods 0.000 claims abstract description 25
- 230000006870 function Effects 0.000 claims description 57
- 238000004458 analytical method Methods 0.000 claims description 30
- 238000011176 pooling Methods 0.000 claims description 18
- 238000004590 computer program Methods 0.000 claims description 9
- 238000007405 data analysis Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 5
- 230000006835 compression Effects 0.000 claims description 5
- 238000007906 compression Methods 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 5
- 238000013473 artificial intelligence Methods 0.000 abstract description 4
- 208000035475 disorder Diseases 0.000 description 68
- 206010028851 Necrosis Diseases 0.000 description 28
- 230000017074 necrotic cell death Effects 0.000 description 28
- 206010028980 Neoplasm Diseases 0.000 description 27
- 238000012360 testing method Methods 0.000 description 14
- 238000003745 diagnosis Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 11
- 238000005516 engineering process Methods 0.000 description 11
- 230000008569 process Effects 0.000 description 9
- 230000004927 fusion Effects 0.000 description 8
- 230000000694 effects Effects 0.000 description 6
- 238000010200 validation analysis Methods 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 5
- 238000011156 evaluation Methods 0.000 description 5
- 206010068319 Oropharyngeal pain Diseases 0.000 description 4
- 201000007100 Pharyngitis Diseases 0.000 description 4
- 230000001338 necrotic effect Effects 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 208000024891 symptom Diseases 0.000 description 4
- 230000000007 visual effect Effects 0.000 description 4
- 238000005481 NMR spectroscopy Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000035945 sensitivity Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 238000003759 clinical diagnosis Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000001788 irregular Effects 0.000 description 2
- 238000002595 magnetic resonance imaging Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000004393 prognosis Methods 0.000 description 2
- 238000001959 radiotherapy Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000000585 Mann–Whitney U test Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000009534 blood test Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000009535 clinical urine test Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000013506 data mapping Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 238000012333 histopathological diagnosis Methods 0.000 description 1
- 230000008595 infiltration Effects 0.000 description 1
- 238000001764 infiltration Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000003211 malignant effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000007723 transport mechanism Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The embodiment of the application relates to the technical field of artificial intelligence, and particularly provides a training method, device, terminal and readable storage medium for a disease classification model. The method comprises the following steps: obtaining image data, structured data and category labels; inputting the image data into a feature extraction network of a disease classification model to obtain a first feature vector; inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector; inputting the first feature vector and the second feature vector into a feature stitching network of the disease classification model to obtain a third feature vector of the target training sample; inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction category and a prediction probability of the prediction category, and constructing a loss function according to the prediction category, the prediction probability and a category label; and iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model, thereby improving the accuracy of disorder identification.
Description
Technical Field
The present application relates to the field of artificial intelligence technologies, and in particular, to a training method, device, terminal, and readable storage medium for a disease classification model.
Background
Nasopharyngeal necrosis is a serious complication of radiation therapy of benign and malignant diseases of the head and neck of the otorhinolaryngology, wherein about one third of patients suffering from nasopharyngeal necrosis are accompanied by tumor infiltration, namely tumor-infiltrated nasopharyngeal necrosis, and compared with patients without tumor-infiltrated nasopharyngeal necrosis, the patients suffer from poorer prognosis and more complex treatment.
At present, diagnosis of tumor invasive nasopharyngeal necrosis is mainly assisted by medical images of a radiologist on nuclear magnetic resonance technology, and a traditional imaging mode is used by the doctor to diagnose the tumor invasive nasopharyngeal necrosis, and the method mainly depends on judgment of experience of the radiologist. First, the nasopharyngeal necrosis has irregular outline, unclear edge and insignificant boundary with surrounding tissues on nuclear magnetic resonance. Secondly, tumor invasive nasopharyngeal necrosis only occupies a part of the whole necrotic area, and various structures are arranged around the necrotic area, which causes a certain difficulty for the doctor to judge.
With the rapid development of computer science and technology, a great number of artificial intelligence means are applied to the medical service industry, so that patients can enjoy safe, convenient and high-quality diagnosis and treatment services. However, in the prior art, the prediction accuracy of the diagnosis of tumor invasive nasopharyngeal necrosis is low, the efficacy is limited, and the prediction effect of the model is further improved by means of richer information, advanced data mining and data analysis.
Disclosure of Invention
The main purpose of the embodiments of the present application is to provide a training method, device, terminal and readable storage medium for a disease classification model, which aims to solve the problem of low prediction accuracy of diagnosis of tumor invasive nasopharyngeal necrosis, and further improve the accuracy of disease diagnosis by using the model.
In a first aspect, embodiments of the present application provide a method for training a disorder classification model, including:
obtaining a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient.
And inputting the image data to a feature extraction network of a disease classification model to be trained, and obtaining a first feature vector corresponding to the image data.
And inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data.
And inputting the first feature vector and the second feature vector into a feature stitching network of the disorder classification model to obtain a third feature vector of the target training sample.
And inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label.
And iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model.
In a second aspect, embodiments of the present application further provide a disease classification model training apparatus, including:
the data acquisition module is used for acquiring a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient;
the image feature extraction module is used for inputting the image data into a feature extraction network of a disease classification model to be trained to obtain a first feature vector corresponding to the image data;
the data feature extraction module is used for inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data;
The feature stitching module is used for inputting the first feature vector and the second feature vector into a feature stitching network of the disorder classification model to obtain a third feature vector of the target training sample;
the data analysis module is used for inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label;
and the data updating module is used for iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model.
In a third aspect, embodiments of the present application also provide a terminal device comprising a processor, a memory, a computer program stored on the memory and executable by the processor, and a data bus for enabling a connection communication between the processor and the memory, wherein the computer program, when executed by the processor, implements the steps of the training method of any of the disorder classification models as provided in the present application.
In a fourth aspect, embodiments of the present application also provide a storage medium for computer readable storage, wherein the storage medium stores one or more programs executable by one or more processors to implement the steps of a method of training a classification model of a condition as provided in any of the present application.
The embodiment of the application provides a training method, a device, a terminal and a storage medium of a disease classification model, wherein the training method comprises the steps of obtaining a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient; inputting the image data into a feature extraction network of a disease classification model to be trained, and obtaining a first feature vector corresponding to the image data; inputting the structured data into a feature screening network of the disease classification model to obtain a second feature vector corresponding to the structured data; inputting the first feature vector and the second feature vector into a feature stitching network of the disease classification model to obtain a third feature vector of the target training sample; inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label; and carrying out iterative updating on the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model. Therefore, the obtained target disorder classification model can be utilized to process the information of the patient to obtain the disorder classification result of the patient. In the training of the target disorder classification model, the multi-modal information of the patient is considered, and the information of different modes is subjected to feature fusion, so that the information of a plurality of modes can be effectively utilized, and the accuracy of disorder classification model identification is improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flow chart of a training method of a disease classification model according to an embodiment of the present application;
fig. 2 is a schematic data transmission diagram of a training method of a disease classification model according to an embodiment of the present application;
fig. 3 is a schematic diagram of training data distribution of a disease classification model according to an embodiment of the present application;
FIG. 4 is a schematic diagram showing the specific effects of a disease classification model according to an embodiment of the present disclosure;
FIG. 5 is a graph showing the results of analysis of a disease classification model, a deep learning model and an image histology model according to an embodiment of the present application;
FIG. 6 is a schematic diagram of a comparison of the prediction results of a clinical diagnosis and disease classification model according to an embodiment of the present application;
fig. 7 is a schematic block diagram of a training device for a disease classification model according to an embodiment of the present application;
Fig. 8 is a schematic block diagram of a structure of a terminal device according to an embodiment of the present application.
Detailed Description
The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure.
The flow diagrams depicted in the figures are merely illustrative and not necessarily all of the elements and operations/steps are included or performed in the order described. For example, some operations/steps may be further divided, combined, or partially combined, so that the order of actual execution may be changed according to actual situations.
It is to be understood that the terminology used in the description of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
The embodiment of the application provides a training method, device, terminal and readable storage medium for a disease classification model. The training method of the disease classification model can be applied to terminal equipment, wherein the terminal equipment can be a tablet personal computer, a notebook personal digital assistant, a wearable device or a server, and the server can be an independent server or a server cluster.
The embodiment of the application provides a training method, a training device, a training terminal and a readable storage medium of a disease classification model, wherein the training method comprises the steps of obtaining a target training sample, wherein the target training sample comprises image data, structured data and a class label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient; inputting the image data into a feature extraction network of a disease classification model to be trained, and obtaining a first feature vector corresponding to the image data; inputting the structured data into a feature screening network of the disease classification model to obtain a second feature vector corresponding to the structured data; inputting the first feature vector and the second feature vector into a feature stitching network of the disease classification model to obtain a third feature vector of the target training sample; inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label; and carrying out iterative updating on the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model. Therefore, the obtained target disorder classification model can be utilized to process the information of the patient to obtain the disorder classification result of the patient. In the training of the target disorder classification model, the multi-modal information of the patient is considered, and the information of different modes is subjected to feature fusion, so that the information of a plurality of modes can be effectively utilized, and the accuracy of disorder classification model identification is improved.
Some embodiments of the present application are described in detail below with reference to the accompanying drawings. The following embodiments and features of the embodiments may be combined with each other without conflict.
At present, diagnosis of tumor invasive nasopharyngeal necrosis is mainly assisted by medical images of radiologists on nuclear magnetic resonance technology, and finally histopathological diagnosis is obtained under an electronic nasopharyngoscope to judge whether the tumor invasive nasopharyngeal necrosis is caused or not.
Along with the development of computer computing power and optimization of algorithms, artificial intelligence becomes a research hot spot in the application of computer image recognition, voice recognition and natural language processing, and a convolutional neural network is widely accepted as a deep learning mainstream algorithm in the field of computer images in the field of machine learning research and shows good application capability. Along with the development of structural optimization of the convolutional neural network, the convolutional neural network has a great prospect in the aspect of intelligent diagnosis of medical images. Recent studies have revealed that deep learning has been successfully applied to cancer research, particularly tumor detection and prognosis of patients.
There are limits to the way doctors diagnose tumor-infiltrating nasopharyngeal necrosis using traditional imaging modalities, the performance of which is largely determined by the judgment of the radiologist's experience. Furthermore, nasopharyngeal necrosis is irregular in outline, unclear in margin, and not apparent to the boundary of surrounding tissues. Furthermore, tumor invasive nasopharyngeal necrosis only occupies a part of the whole necrotic area, and various structures are arranged around the necrotic area, which causes a certain difficulty for the doctor's judgment.
In order to solve the problem of low accuracy in diagnosis of tumor invasive nasopharyngeal necrosis in the prior art, the application provides a scheme of a disease classification model, and the accuracy of classification of the disease classification model is improved through feature fusion of different modes.
Referring to fig. 1, fig. 1 is a flowchart of a training method of a disease classification model according to an embodiment of the present application.
As shown in fig. 1, the training method of the disease classification model includes steps S1 to S6.
Step S1: obtaining a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient.
Illustratively, a target training sample is obtained, wherein the target training sample may contain data of multiple modalities, such as image data and personal structured data of the patient. The training of the disease classification model is supported by feature fusion of data of different modalities.
The image data is a medical image of the patient that is used to reflect features of the user's image details, such as an MRI (Magnetic Resonance Imaging, magnetic resonance) image; the structured data is patient personal structured data including, but not limited to, age, number of radiation therapy sessions, plasma EBV DNA, etc., which is used to reflect characteristics of the user's basic information. The method comprises the steps of obtaining data information under two different modes of user image characteristics and user personal basic characteristics, fusing the data information, enabling the model to know the overall state of a patient more comprehensively and efficiently, achieving complementation of various information, and further improving accuracy of the disease classification model.
The source of the structured data may be from a patient's form-filling input, or a description of a word, so that unstructured data may be converted into structured data when the structured data is obtained, and technical means such as named entity recognition, relationship extraction, and triplet extraction may be used.
In some embodiments, the obtaining the target training sample comprises: acquiring patient data; and preprocessing the patient data to obtain structured data.
Illustratively, relevant data of the patient, such as image data, patient data, and category labels, are first collected as needed. The image data is a medical image describing a patient, and the patient data is personal data describing a sample patient, including personal information and related medical data, which may be other test index data of the patient, such as results of blood tests, results of urine tests, and so forth. The category white label is an accurate judgment result given by a doctor under the image data and the patient data of the current patient. And forming the target training sample by the image data, the patient data and the category labels.
The patient data may be structured data, unstructured data and semi-structured data, where the unstructured data and the semi-structured data need to be subjected to data preprocessing to obtain structured data, for example, technologies such as named entity identification, relation extraction, entity unification, and the like, and the unstructured data or the semi-structured data in the patient data is subjected to data processing to obtain the structured data.
For example, the patient data is a text description, such as "Zhang san, 35 years old, symptoms of existing sore throat", and the name of the person in the text description can be extracted according to the named entity recognition technology: zhang III, age: age 35, symptoms: sore throat; then, according to the relation extraction technique, the relationship among name, age and symptoms is obtained, such as Zhang San, age and 35 years. Unstructured representations of the original patient data may be converted to structured data in accordance with the techniques described above to support subsequent model training.
In addition, after the structured data is obtained, in order to unify the data, the data description needs to be normalized by using an entity unifying technical means, for example, the sore throat and sore throat are actually the same symptom, but the text presentation forms are different, so that the entity unifying means can be adopted to align the data. For example, a corresponding data mapping table is constructed according to historical experience or manual labeling mode, and then the unified effect of the entity is achieved.
For example, after obtaining the image data, the structured patient data and the category label of the user, the data needs to be screened for a second time, for example, relevant information with missing data is removed, so as to obtain a target training sample.
Step S2: and inputting the image data to a feature extraction network of a disease classification model to be trained, and obtaining a first feature vector corresponding to the image data.
The image data is input to a feature extraction network, and the image data and other image data can be distinguished from each other to form a feature vector to represent the image data. The first feature vector not only can well describe the image data, but also can distinguish the image data from other image data. Further, the obtained first feature vector can represent that the difference between images of the same category as the image data is small, and the difference between different categories of the images is large.
For example, the difference between the first feature vector of the patient a and the first feature vector of the patient B obtained through the feature extraction network of the disorder classification model to be trained should be large for the medical image corresponding to the patient a with or without tumor-infiltrating nasopharyngeal necrosis, and the difference between the first feature vectors of the medical image corresponding to the case where both the patient a and the patient B have or do not have tumor-infiltrating nasopharyngeal necrosis should be small.
For example, a neural network may be constructed by using a deep learning manner to perform feature extraction of image data by obtaining a first feature vector corresponding to the image data through a feature extraction network of a disorder classification model to be trained. However, when the deep learning method is used for feature extraction, large-scale training data are often required, but in the existing medical research, the large-scale training data are relatively missing, and if the deep learning is performed on a small sample, the feature extraction network may be over-fitted and cannot have better generalization capability. Based on this problem, the effect of image data feature extraction can be ensured by means of incremental learning.
For example, the feature extraction network is pre-trained based on the large visual database ImageNet, so that the feature extraction network can learn general characterization capability and bring better generalization performance, and convergence on target tasks is accelerated. After the pre-training is finished, the medical image is utilized to extract the visual characteristics of the medical image from the pre-training characteristic extraction network, so that the characteristic extraction network can learn the characteristics of the medical image better under the condition of retaining the generalization capability of large-scale data.
In some embodiments, the inputting the image data into the feature extraction network of the disorder classification model to be trained, before obtaining the first feature vector corresponding to the image data, further includes: obtaining training data, wherein the training data comprises a first picture and first characteristic information, and the first characteristic information is pixel characteristic information of a target object in the first picture; inputting the first picture into an initial feature extraction network to obtain second feature information; determining a first loss function according to the first characteristic information and the second characteristic information; and carrying out iterative updating on the initial feature extraction network according to the training data and the first loss function to obtain a feature extraction network of the disorder classification model to be trained.
For example, before training the image data feature extraction network, the feature extraction network learns the general representation capability of the image data on the large-scale basic data, brings better generalization performance to the feature extraction network, and further supports the better feature extraction capability on the small-sample medical image.
For example, pre-training data is obtained, the pre-training data includes a picture and pixel characteristic information of a target object corresponding to the picture, the target object can be a person or an object in the picture, and the category or the position information of the target object in the picture can be known through the pixel characteristic information. Inputting the picture into an initial feature extraction network to obtain pixel feature information predicted by a target object; comparing the pixel characteristic information with the predicted pixel characteristic information, judging the difference between the pixel characteristic information and the predicted pixel characteristic information, determining a loss value according to the difference value, and adding all the loss values of the training data to obtain a loss function; and (3) through parameter adjustment on the initial feature extraction network, the result of the loss function is continuously reduced, and when the loss value of the loss function meets the preset condition, the iterative update on the initial feature extraction network is stopped, so that the feature extraction network of the disorder classification model to be trained is obtained.
In addition, in order to ensure that the target feature extraction network is over-fitted, the loss value is not lower and better, and the selection needs to be performed according to the actual test result. Therefore, training rounds of the model can be set, and when the rounds meet preset round values, iterative updating of the initial feature extraction network can be stopped, so that the target feature extraction network is obtained.
For example, the feature extraction network structure may be set as an afflicientnet network structure, the large visual database ImageNet data is transferred to the afflicientnet network structure to perform pre-training, and the target feature extraction network is obtained after testing by a test set according to a preset value of a loss function or a set target training round. After the target feature extraction network is obtained through pre-training, performing incremental training on the target feature extraction network according to the medical image data to obtain a feature extraction network of a disease classification model to be trained, and further guaranteeing the accuracy and effectiveness of feature extraction on the image data to obtain a first feature vector.
Step S3: and inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data.
Illustratively, structured data contains multiple data types, but the data types are of different importance for disorder classification, even some data types have no effect on the outcome of the disorder classification model, and therefore, there is a need to quickly and accurately pick out useful features from the multiple data types, reduce the complexity of model training, reduce computational costs, and reduce the likelihood of overfitting.
For example, the clinical structured data of the patient is subjected to feature analysis to obtain the influence factors of each structured data type, and then the data types are screened according to the sizes of the influence factors to obtain useful data types or data types with larger influence factors so as to train a subsequent disease classification model.
In some embodiments, the feature screening network includes a feature analysis network and a feature selection network, the inputting the structured data into the feature screening network of the disorder classification model, obtaining a second feature vector corresponding to the structured data, includes: inputting the structured data into the feature analysis network for feature analysis to obtain a feature analysis result in the structured data; and inputting the feature analysis result into the feature selection network to perform feature screening to obtain a second feature vector corresponding to the structured data.
The method comprises the steps of inputting structured data into a feature analysis network for feature analysis to obtain feature importance corresponding to different features, screening according to the feature importance corresponding to the different features to obtain features meeting screening conditions, and vector expressing the screened features to support subsequent model training.
For example, the feature analysis network may employ a U-test, also known as a mann-whitney U-test, which is used to test whether the mean of the two distributions has a significant difference, perform feature analysis on the current clinical structured data by using the U-test to obtain analysis results of different features, set a threshold of the analysis results in the feature screening network, and when the features meet the set threshold, further determine the features meeting the set threshold as features of a subsequent disorder classification model.
Step S4: and inputting the first feature vector and the second feature vector into a feature stitching network of the disorder classification model to obtain a third feature vector of the target training sample.
The image data and the structured data are subjected to multi-mode information fusion to obtain patient information with different dimensions, and accuracy of a disease classification model of a patient is improved. The characteristics of different sources of the patient are obtained, and the advantages of the characteristics are fused by utilizing the complementarity among the characteristics, so that the performance of the model is improved.
The existing feature fusion technology includes feature stitching, feature summation and the like, wherein the feature summation can also be called bit-by-bit addition, namely, in order to fuse two feature vectors, the addition of corresponding elements is directly carried out, if the dimensions of the two feature vectors are different, the dimension vectors can be converted by linear transformation, and then the addition is carried out. The feature splicing can splice two feature vectors in a tail-end connection mode, and then the feature vectors are fused.
In some embodiments, the feature stitching network includes a feature pooling layer, a feature full connection layer, and a feature stitching layer, the inputting the first feature vector and the second feature vector to the feature stitching network of the condition classification model, obtaining a third feature vector of the target training sample, comprising: inputting the first feature vector to the feature pooling layer for feature compression to obtain a first pooling vector; inputting the second feature vector to the feature full-connection layer for vector dimension expansion to obtain a first expansion vector; and inputting the first pooling vector and the first expansion vector into the feature stitching layer to perform vector stitching, so as to obtain a third feature vector of the target training sample.
For example, after obtaining a first feature vector corresponding to an image and a second feature vector corresponding to structured data, feature fusion of different source data is performed by using a multi-mode feature stitching technology, and in order to achieve this objective, a feature pooling layer in a feature stitching network is first used, for example, feature compression is performed on the first feature vector by using a global average pooling technology to obtain a feature vector of a target dimension, for example, the feature dimension is set to 1024 dimensions. And then carrying out feature dimension increase on the structured data by utilizing a feature full-connection layer in the feature splicing network to obtain a dimension-increased feature vector, wherein the dimension-increased feature vector is set to be 64 dimensions, and then carrying out feature space feature splicing on the two feature vectors to obtain a multi-mode feature characterization vector, namely obtaining a 1088-dimension third feature vector.
Step S5: inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label; .
The feature representation of the training data is obtained, and then the feature representation is input to a feature classification network of the disorder classification model to perform feature classification, so that a probability value corresponding to a classification category processed by the feature classification network is obtained, and when the probability value corresponding to a true classification category is larger, the probability value corresponding to a non-true classification category is smaller, so that the classification result of the feature classification network is more correct.
The purpose of constructing the loss function is to hope that the prediction result of the feature classification network is more and more accurate, namely, the larger the probability value corresponding to the real classification category is trained, the smaller the probability value corresponding to the non-real classification category is trained, and when the loss function meets the set condition, the training of the feature classification network is considered to be completed.
The loss function is mainly used in the training phase of the model. After each batch of training data is sent into the model, a predicted value is output through forward propagation, and then a difference value between the predicted value and a true value, namely a loss value, is calculated by a loss function. After the loss value is obtained, the model updates each parameter through back propagation to reduce the loss between the real value and the predicted value, so that the predicted value generated by the model is close to the real value, and the learning purpose is achieved. After training the model, each parameter has been optimized after the model has been back-propagated. The result of prediction using the model must be close to the true result. Common loss functions include cross entropy loss functions, logistic loss, and the like.
In some implementations, the feature classification network may be a multi-layer perceptron that includes an input layer, a hidden layer, and an output layer.
The different layers of the multi-layer sensor are connected through full connection. The connection strength between neurons in a multi-layer perceptron is represented by weights, the magnitude of which represents the magnitude of the likelihood. The multi-layer sensor is also provided with a bias, and the bias is set for correctly classifying samples, which is an important parameter in the multi-layer sensor, namely ensuring that the output value calculated through input cannot be activated randomly. The multi-layer sensor also comprises an activation function which mainly plays a role of nonlinear mapping, and can limit the output amplitude of neurons to a certain range, and the output amplitude is generally limited to (-1) or (0-1). The most commonly used activation function is the Sigmoid function, which can change (- +. ++ infinity) is mapped to a range of (0 to 1).
For example, the disease classification model needs to classify disease and non-disease, and after the feature vector is input to the multi-layer perceptron, the probability corresponding to the disease and the non-disease respectively can be obtained, and the probability value corresponding to the classification class corresponding to the training data is as large as possible in the model training stage. If the training data 1-is ill and the training data 2-is not ill, the probability of the ill corresponding to the training data 1 is as close to 1 as possible after the feature vector passes through the multi-layer perceptron in the training stage, and the probability of the non-ill corresponding to the training data 2 is as close to 1 as possible. Wherein the sum of all class probabilities in one training data is equal to 1.
In some embodiments, the constructing a loss function from the prediction category, the prediction probability of the prediction category, and the category label comprises: obtaining a maximum value of the prediction probability of the prediction category of the target training sample according to the prediction category and the prediction probability of the prediction category; and utilizing log function conversion according to the maximum value of the prediction probability of the prediction category of the target training sample, and determining the loss function by combining the category label.
Illustratively, the loss function is: loss= - (1-p) t ) r log(p t ) Wherein p is t For training the probability that the sample belongs to the category t, r is a super parameter, and the purpose of the loss function is to make the model concentrate on the sample difficult to classify during training by reducing the weight of the sample easy to classify.
It is a dynamic scaling cross entropy penalty, by which the weights of easily distinguishable samples during training can be dynamically reduced, thereby rapidly focusing the center of gravity on those indistinguishable samples (possibly positive, or negative, but all helpful to the training network).
Step S6: and iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model.
Illustratively, parameters of the disorder classification model are continuously updated according to the training data, and when the model training round satisfies a preset round or the loss function satisfies a set threshold, the iterative updating is stopped. And performing model evaluation on the obtained disorder classification model, wherein when the model evaluation result meets the requirement, the obtained disorder classification model is a target disorder classification model, and when the model evaluation result does not meet the requirement, analyzing reasons, increasing training data or reducing a threshold value of a loss function according to the reasons, and retraining.
For example, according to analysis of the model evaluation result, the model lacks generalization capability, and then the target training data is required to be added or training rounds are required to be added; if the model overfitting is obtained through analysis according to the model evaluation result, the training round needs to be reduced at the moment.
Illustratively, a data transmission flow diagram of a method of training a disorder classification model is shown in FIG. 2. 434 cases of nasopharyngeal necrosis patients diagnosed at the university of middle mountain tumor prevention center, including 201 cases of tumor-infiltrating nasopharyngeal necrosis patients and 233 cases of tumor-infiltrating-free nasopharyngeal necrosis patients, were screened from 3 months in 2012 to 9 months in 2020, and the cohorts were randomly divided into a training set, a validation set, and a test set at a ratio of 7:2:1.
The baseline clinical profile for the training set, validation set and test set is shown in figure 3. In the training, validation and test sets, 49.1%, 40.9% and 37.2% of patients developed tumor-infiltrating nasopharyngeal necrosis, respectively.
The performance of the disorder classification model using the present application in the training set, validation set, and test set is shown in fig. 4, as compared to the performance of the deep learning model, the image histology model, and the radiologist diagnosis. Through a deep learning model, the prediction accuracy of tumor invasive nasopharyngeal necrosis is 0.725, the sensitivity is 0.732 and the specificity is 0.718 in the test set. When the diagnosis is carried out, when the EBV DNA in the structured data is incorporated into a disease classification model, the accuracy, the sensitivity and the specificity can be respectively improved to 0.818, 0.859 and 0.779 by multi-modal data fusion, and compared with the diagnosis of an image histology model and a radiologist, the disease classification model has the highest detection accuracy and the highest detection sensitivity when the Area Under a working characteristic Curve (AUC) of a subject in the aspects of distinguishing tumor-invasive nasopharyngeal necrosis from non-tumor-invasive nasopharyngeal necrosis is 0.876, as shown in figure 5.
Using the training set, validation set and test set data described above, the last layer of the model is subjected to t-SNE (t-distributed stochastic neighbor embedding) before the disorder classification model is classified
The conversion realizes dimension reduction, and then whether obvious characteristic differences exist between the image without tumor invasive nasopharyngeal necrosis and the image with tumor invasive nasopharyngeal necrosis or not is compared. the effectiveness of the disease classification model algorithm constructed by the method is visually verified through the t-SNE. The visual results in the training set, the validation set and the test set can show that the effect of disease classification model prediction is generally better than clinical diagnosis. As shown in fig. 6, the disorder classification model can more effectively distinguish between tumor-free and tumor-infiltrated nasopharyngeal necrosis.
Referring to fig. 7, fig. 7 is a schematic diagram of a disease classification model training device 200 provided in an embodiment of the present application, where the disease classification model training device 200 includes a data acquisition module 201, an image feature extraction module 202, a data feature extraction module 203, a feature stitching module 204, a data analysis module 205, and a data update module 206, and the data acquisition module 201 is configured to obtain a target training sample, where the target training sample includes image data, structured data, and a class label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient; the image feature extraction module 202 is configured to input the image data to a feature extraction network of a disorder classification model to be trained, and obtain a first feature vector corresponding to the image data; the data feature extraction module 203 is configured to input the structured data to a feature screening network of the disorder classification model, and obtain a second feature vector corresponding to the structured data; a feature stitching module 204, configured to input the first feature vector and the second feature vector to a feature stitching network of the disorder classification model, to obtain a third feature vector of the target training sample; a data analysis module 205, configured to input the third feature vector to a feature classification network of the condition classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and construct a loss function according to the prediction class, the prediction probability of the prediction class, and the class label; and a data updating module 206, configured to iteratively update the disorder classification model based on the training samples and the loss function, so as to obtain a target disorder classification model.
In some embodiments, the data acquisition module 201 performs, in obtaining the target training samples:
acquiring patient data;
and preprocessing the patient data to obtain structured data.
In some embodiments, the image feature extraction module 202 further performs, before inputting the image data to a feature extraction network of a disorder classification model to be trained, obtaining a first feature vector corresponding to the image data:
obtaining training data, wherein the training data comprises a first picture and first characteristic information, and the first characteristic information is pixel characteristic information of a target object in the first picture;
inputting the first picture into an initial feature extraction network to obtain second feature information;
determining a first loss function according to the first characteristic information and the second characteristic information;
and carrying out iterative updating on the initial feature extraction network according to the training data and the first loss function to obtain a feature extraction network of the disorder classification model to be trained.
In some embodiments, the feature screening network includes a feature analysis network and a feature selection network, and the data feature extraction module 203 performs, in inputting the structured data into the feature screening network of the disorder classification model, to obtain a second feature vector corresponding to the structured data:
Inputting the structured data into the feature analysis network for feature analysis to obtain a feature analysis result in the structured data;
and inputting the feature analysis result into the feature selection network to perform feature screening to obtain a second feature vector corresponding to the structured data.
In some embodiments, the feature stitching network includes a feature pooling layer, a feature full connection layer, and a feature stitching layer, and the feature stitching module 204 performs, in inputting the first feature vector and the second feature vector to the feature stitching network of the disorder classification model, to obtain a third feature vector of the target training sample:
inputting the first feature vector to the feature pooling layer for feature compression to obtain a first pooling vector;
inputting the second feature vector to the feature full-connection layer for vector dimension expansion to obtain a first expansion vector;
and inputting the first pooling vector and the first expansion vector into the feature stitching layer to perform vector stitching, so as to obtain a third feature vector of the target training sample.
In some implementations, the feature classification network in the data analysis module 205 is a multi-layer perceptron.
In some embodiments, the data analysis module 205 performs, in the constructing a loss function from the prediction category, the prediction probability of the prediction category, and the category label:
obtaining a maximum value of the prediction probability of the prediction category of the target training sample according to the prediction category and the prediction probability of the prediction category;
and utilizing log function conversion according to the maximum value of the prediction probability of the prediction category of the target training sample, and determining the loss function by combining the category label.
It should be noted that, for convenience and brevity of description, specific working processes of the above-described apparatus may refer to corresponding processes in the foregoing training method embodiments of the disease classification model, and are not described herein again.
Referring to fig. 8, fig. 8 is a schematic block diagram of a structure of a terminal device according to an embodiment of the present application.
As shown in fig. 8, the terminal device 300 includes a processor 301 and a memory 302, and the processor 301 and the memory 302 are connected by a bus 303, such as an I2C (Inter-integrated Circuit) bus.
In particular, the processor 301 is used to provide computing and control capabilities, supporting the operation of the entire server. The processor 301 may be a central processing unit (Central Processing Unit, CPU), the processor 301 may also be other general purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate arrays (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. Wherein the general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
Specifically, the Memory 302 may be a Flash chip, a Read-Only Memory (ROM) disk, an optical disk, a U-disk, a removable hard disk, or the like.
It will be appreciated by those skilled in the art that the structure shown in fig. 8 is merely a block diagram of a portion of the structure related to the embodiment of the present application and does not constitute a limitation of the terminal device to which the embodiment of the present application is applied, and that a specific terminal device may include more or less components than those shown in the drawings, or may combine some components, or have a different arrangement of components.
The processor 301 is configured to execute a computer program stored in a memory, and implement the training method of the disease classification model provided in any embodiment of the present application when the computer program is executed.
In some embodiments, the processor 301 is configured to run a computer program stored in a memory, apply to a terminal device, and implement the following steps when executing the computer program:
obtaining a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient;
Inputting the image data to a feature extraction network of a disease classification model to be trained, and obtaining a first feature vector corresponding to the image data;
inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data;
inputting the first feature vector and the second feature vector into a feature stitching network of the disorder classification model to obtain a third feature vector of the target training sample;
inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label;
and iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model.
In some embodiments, the processor 301 performs, in the obtaining the target training sample:
acquiring patient data;
and preprocessing the patient data to obtain structured data.
In some embodiments, the processor 301 further performs, before inputting the image data to the feature extraction network of the disorder classification model to be trained, obtaining the first feature vector corresponding to the image data:
obtaining training data, wherein the training data comprises a first picture and first characteristic information, and the first characteristic information is pixel characteristic information of a target object in the first picture;
inputting the first picture into an initial feature extraction network to obtain second feature information;
determining a first loss function according to the first characteristic information and the second characteristic information;
and carrying out iterative updating on the initial feature extraction network according to the training data and the first loss function to obtain a feature extraction network of the disorder classification model to be trained.
In some embodiments, the feature screening network includes a feature analysis network and a feature selection network, and the processor 301 performs, in the step of inputting the structured data into the feature screening network of the disorder classification model, to obtain a second feature vector corresponding to the structured data:
inputting the structured data into the feature analysis network for feature analysis to obtain a feature analysis result in the structured data;
And inputting the feature analysis result into the feature selection network to perform feature screening to obtain a second feature vector corresponding to the structured data.
In some embodiments, the feature stitching network includes a feature pooling layer, a feature full connection layer, and a feature stitching layer, and the processor 301 performs, in the inputting the first feature vector and the second feature vector into the feature stitching network of the disorder classification model, to obtain a third feature vector of the target training sample:
inputting the first feature vector to the feature pooling layer for feature compression to obtain a first pooling vector;
inputting the second feature vector to the feature full-connection layer for vector dimension expansion to obtain a first expansion vector;
and inputting the first pooling vector and the first expansion vector into the feature stitching layer to perform vector stitching, so as to obtain a third feature vector of the target training sample.
In some implementations, the feature classification network in processor 301 is a multi-layer perceptron.
In some implementations, the processor 301 performs, in said constructing a loss function from said prediction category, a prediction probability of said prediction category, and said category label:
Obtaining a maximum value of the prediction probability of the prediction category of the target training sample according to the prediction category and the prediction probability of the prediction category;
and utilizing log function conversion according to the maximum value of the prediction probability of the prediction category of the target training sample, and determining the loss function by combining the category label.
It should be noted that, for convenience and brevity of description, specific working processes of the terminal device described above may refer to corresponding processes in the foregoing training method embodiments of the disease classification model, and are not described herein again.
Embodiments of the present application also provide a storage medium for computer readable storage, the storage medium storing one or more programs executable by one or more processors to implement the steps of the training method of any of the disorder classification models as provided in the embodiments of the present application.
The storage medium may be an internal storage unit of the terminal device of the foregoing embodiment, for example, a memory of the terminal device. The storage medium may also be an external storage device of the terminal device, such as a plug-in hard disk provided on the terminal device, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), or the like.
Those of ordinary skill in the art will appreciate that all or some of the steps of the methods, functional modules/units in the apparatus disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. In a hardware embodiment, the division between the functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed cooperatively by several physical components. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). The term computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as known to those skilled in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer. Furthermore, as is well known to those of ordinary skill in the art, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media.
It should be understood that the term "and/or" as used in this specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations. It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or system. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or system that comprises the element.
The foregoing embodiment numbers of the present application are merely for describing, and do not represent advantages or disadvantages of the embodiments. The foregoing is merely illustrative of the embodiments of the present application, but the scope of the present application is not limited thereto, and any equivalent modifications or substitutions will be apparent to those skilled in the art within the scope of the present application, and these modifications or substitutions are intended to be included in the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.
Claims (10)
1. A method of training a disorder classification model, the method comprising:
obtaining a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient;
inputting the image data to a feature extraction network of a disease classification model to be trained, and obtaining a first feature vector corresponding to the image data;
inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data;
inputting the first feature vector and the second feature vector into a feature stitching network of the disorder classification model to obtain a third feature vector of the target training sample;
inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label;
And iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model.
2. The method of training a disorder classification model according to claim 1, wherein obtaining a target training sample comprises:
acquiring patient data;
and preprocessing the patient data to obtain structured data.
3. The method for training a disorder classification model according to claim 1, wherein before inputting the image data to a feature extraction network of a disorder classification model to be trained to obtain a first feature vector corresponding to the image data, further comprises:
obtaining training data, wherein the training data comprises a first picture and first characteristic information, and the first characteristic information is pixel characteristic information of a target object in the first picture;
inputting the first picture into an initial feature extraction network to obtain second feature information;
determining a first loss function according to the first characteristic information and the second characteristic information;
and carrying out iterative updating on the initial feature extraction network according to the training data and the first loss function to obtain a feature extraction network of the disorder classification model to be trained.
4. The method of training a disorder classification model according to claim 1, wherein the feature screening network comprises a feature analysis network and a feature selection network;
the step of inputting the structured data to a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data, including:
inputting the structured data into the feature analysis network for feature analysis to obtain a feature analysis result in the structured data;
and inputting the feature analysis result into the feature selection network to perform feature screening to obtain a second feature vector corresponding to the structured data.
5. The method of training a disorder classification model according to claim 1, wherein the feature stitching network comprises a feature pooling layer and a feature full connection layer; the inputting the first feature vector and the second feature vector into the feature stitching network of the disorder classification model, to obtain a third feature vector of the target training sample, includes:
inputting the first feature vector to the feature pooling layer for feature compression to obtain a first pooling vector;
inputting the second feature vector to the feature full-connection layer for vector dimension expansion to obtain a first expansion vector;
And inputting the first pooling vector and the first expansion vector into the feature stitching layer to perform vector stitching, so as to obtain a third feature vector of the target training sample.
6. The method of claim 1, wherein the feature classification network is a multi-layer perceptron.
7. The method of training a classification model of a condition according to claim 1, wherein said constructing a loss function from said predicted class, a predicted probability of said predicted class, and said class label comprises:
obtaining a maximum value of the prediction probability of the prediction category of the target training sample according to the prediction category and the prediction probability of the prediction category;
and utilizing log function conversion according to the maximum value of the prediction probability of the prediction category of the target training sample, and determining the loss function by combining the category label.
8. A training device for a disorder classification model, comprising:
the data acquisition module is used for acquiring a target training sample, wherein the target training sample comprises image data, structured data and a category label, the image data is a medical image of a sample patient, and the structured data is personal structured data of the sample patient;
The image feature extraction module is used for inputting the image data into a feature extraction network of a disease classification model to be trained to obtain a first feature vector corresponding to the image data;
the data feature extraction module is used for inputting the structured data into a feature screening network of the disorder classification model to obtain a second feature vector corresponding to the structured data;
the feature stitching module is used for inputting the first feature vector and the second feature vector into a feature stitching network of the disorder classification model to obtain a third feature vector of the target training sample;
the data analysis module is used for inputting the third feature vector into a feature classification network of the disorder classification model to obtain a prediction class of the target training sample and a prediction probability of the prediction class, and constructing a loss function according to the prediction class, the prediction probability of the prediction class and the class label;
and the data updating module is used for iteratively updating the disorder classification model based on the training sample and the loss function to obtain a target disorder classification model.
9. A terminal device, characterized in that the terminal device comprises a processor and a memory;
The memory is used for storing a computer program;
the processor is configured to execute the computer program and to implement a training method of the disorder classification model according to any one of claims 1 to 7 when the computer program is executed.
10. A computer-readable storage medium, which when executed by one or more processors, causes the one or more processors to perform the training method steps of the disorder classification model of any one of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310260271.XA CN116452851A (en) | 2023-03-17 | 2023-03-17 | Training method and device for disease classification model, terminal and readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310260271.XA CN116452851A (en) | 2023-03-17 | 2023-03-17 | Training method and device for disease classification model, terminal and readable storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116452851A true CN116452851A (en) | 2023-07-18 |
Family
ID=87132850
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310260271.XA Pending CN116452851A (en) | 2023-03-17 | 2023-03-17 | Training method and device for disease classification model, terminal and readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116452851A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117174298A (en) * | 2023-11-01 | 2023-12-05 | 中国人民解放军总医院 | Emergency critical illness clinical decision system based on multi-mode pre-training large model |
| CN117218433A (en) * | 2023-09-13 | 2023-12-12 | 珠海圣美生物诊断技术有限公司 | Household multi-cancer detection device and multi-mode fusion model construction method and device |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490251A (en) * | 2019-03-08 | 2019-11-22 | 腾讯科技（深圳）有限公司 | Prediction disaggregated model acquisition methods and device, storage medium based on artificial intelligence |
| CN111291817A (en) * | 2020-02-17 | 2020-06-16 | 北京迈格威科技有限公司 | Image recognition method and device, electronic equipment and computer readable medium |
| CN111583217A (en) * | 2020-04-30 | 2020-08-25 | 深圳开立生物医疗科技股份有限公司 | Tumor ablation curative effect prediction method, device, equipment and computer medium |
| CN111783519A (en) * | 2020-05-15 | 2020-10-16 | 北京迈格威科技有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN111814868A (en) * | 2020-07-03 | 2020-10-23 | 苏州动影信息科技有限公司 | Model based on image omics feature selection, construction method and application |
| CN112949772A (en) * | 2021-04-12 | 2021-06-11 | 中国科学院自动化研究所 | Stomach cancer multidimensional feature extraction and analysis system based on image omics |
| CN113888532A (en) * | 2021-11-09 | 2022-01-04 | 推想医疗科技股份有限公司 | Medical image analysis method and device based on flat scanning CT data |
| CN113888470A (en) * | 2021-09-06 | 2022-01-04 | 清华大学 | Diagnosis method and device based on convolutional neural network and multi-modal medical image |
| CN114266693A (en) * | 2021-12-16 | 2022-04-01 | 阿里巴巴（中国）有限公司 | Image processing method, model generation method and equipment |
| CN114398983A (en) * | 2022-01-14 | 2022-04-26 | 腾讯科技（深圳）有限公司 | Classification prediction method, classification prediction device, classification prediction equipment, storage medium and computer program product |
| CN114693971A (en) * | 2022-03-29 | 2022-07-01 | 深圳市大数据研究院 | Classification prediction model generation method, classification prediction method, system and platform |
| CN114864076A (en) * | 2022-05-07 | 2022-08-05 | 扬州大学 | Multi-modal breast cancer classification training method and system based on graph attention network |
| CN115131630A (en) * | 2022-07-20 | 2022-09-30 | 元码基因科技(苏州)有限公司 | Model training method, microsatellite state prediction method, electronic device and storage medium |
| WO2022268102A1 (en) * | 2021-06-22 | 2022-12-29 | 安翰科技(武汉)股份有限公司 | Deep learning-based cancer prognosis survival prediction method and device, and storage medium |
-
2023
- 2023-03-17 CN CN202310260271.XA patent/CN116452851A/en active Pending
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490251A (en) * | 2019-03-08 | 2019-11-22 | 腾讯科技（深圳）有限公司 | Prediction disaggregated model acquisition methods and device, storage medium based on artificial intelligence |
| CN111291817A (en) * | 2020-02-17 | 2020-06-16 | 北京迈格威科技有限公司 | Image recognition method and device, electronic equipment and computer readable medium |
| CN111583217A (en) * | 2020-04-30 | 2020-08-25 | 深圳开立生物医疗科技股份有限公司 | Tumor ablation curative effect prediction method, device, equipment and computer medium |
| CN111783519A (en) * | 2020-05-15 | 2020-10-16 | 北京迈格威科技有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
| CN111814868A (en) * | 2020-07-03 | 2020-10-23 | 苏州动影信息科技有限公司 | Model based on image omics feature selection, construction method and application |
| CN112949772A (en) * | 2021-04-12 | 2021-06-11 | 中国科学院自动化研究所 | Stomach cancer multidimensional feature extraction and analysis system based on image omics |
| WO2022268102A1 (en) * | 2021-06-22 | 2022-12-29 | 安翰科技(武汉)股份有限公司 | Deep learning-based cancer prognosis survival prediction method and device, and storage medium |
| CN113888470A (en) * | 2021-09-06 | 2022-01-04 | 清华大学 | Diagnosis method and device based on convolutional neural network and multi-modal medical image |
| CN113888532A (en) * | 2021-11-09 | 2022-01-04 | 推想医疗科技股份有限公司 | Medical image analysis method and device based on flat scanning CT data |
| CN114266693A (en) * | 2021-12-16 | 2022-04-01 | 阿里巴巴（中国）有限公司 | Image processing method, model generation method and equipment |
| CN114398983A (en) * | 2022-01-14 | 2022-04-26 | 腾讯科技（深圳）有限公司 | Classification prediction method, classification prediction device, classification prediction equipment, storage medium and computer program product |
| CN114693971A (en) * | 2022-03-29 | 2022-07-01 | 深圳市大数据研究院 | Classification prediction model generation method, classification prediction method, system and platform |
| CN114864076A (en) * | 2022-05-07 | 2022-08-05 | 扬州大学 | Multi-modal breast cancer classification training method and system based on graph attention network |
| CN115131630A (en) * | 2022-07-20 | 2022-09-30 | 元码基因科技(苏州)有限公司 | Model training method, microsatellite state prediction method, electronic device and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 梁蒙蒙 等: "基于随机化融合和CNN的多模态肺部肿瘤图像识别", 《南京大学学报(自然科学)》, vol. 54, no. 4, pages 775 - 785 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117218433A (en) * | 2023-09-13 | 2023-12-12 | 珠海圣美生物诊断技术有限公司 | Household multi-cancer detection device and multi-mode fusion model construction method and device |
| CN117174298A (en) * | 2023-11-01 | 2023-12-05 | 中国人民解放军总医院 | Emergency critical illness clinical decision system based on multi-mode pre-training large model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kavitha et al. | Deep learning based capsule neural network model for breast cancer diagnosis using mammogram images | |
| Salahuddin et al. | Transparency of deep neural networks for medical image analysis: A review of interpretability methods | |
| US11416747B2 (en) | Three-dimensional (3D) convolution with 3D batch normalization | |
| US9934567B2 (en) | Methods and means of CAD system personalization to reduce intraoperator and interoperator variation | |
| KR102153920B1 (en) | System and method for interpreting medical images through the generation of refined artificial intelligence reinforcement learning data | |
| CN109308495B (en) | Apparatus and system for automatically predicting physiological condition from medical image of patient | |
| CN116452851A (en) | Training method and device for disease classification model, terminal and readable storage medium | |
| Penzkofer et al. | ESUR/ESUI position paper: developing artificial intelligence for precision diagnosis of prostate cancer using magnetic resonance imaging | |
| CN111709485B (en) | Medical image processing method, device and computer equipment | |
| Yue et al. | Retinal vessel segmentation using dense U-net with multiscale inputs | |
| Cao et al. | Supervised contrastive pre-training formammographic triage screening models | |
| Tajudin et al. | Deep learning in the grading of diabetic retinopathy: A review | |
| Sun et al. | Right for the Wrong Reason: Can Interpretable ML Techniques Detect Spurious Correlations? | |
| Chen et al. | Diagnostic accuracy study of automated stratification of Alzheimer’s disease and mild cognitive impairment via deep learning based on MRI | |
| Sivakumar et al. | Deep learning in skin lesion analysis for malignant melanoma cancer identification | |
| CN113379770B (en) | Construction method of nasopharyngeal carcinoma MR image segmentation network, image segmentation method and device | |
| Guney et al. | An overview of deep learning algorithms and their applications in neuropsychiatry | |
| CN115187512A (en) | Hepatocellular carcinoma great vessel invasion risk prediction method, system, device and medium | |
| Bakasa et al. | Stacked ensemble deep learning for pancreas cancer classification using extreme gradient boosting | |
| Khozama et al. | Study the Effect of the Risk Factors in the Estimation of the Breast Cancer Risk Score Using Machine Learning | |
| Prasad et al. | Lung cancer detection and classification using deep neural network based on hybrid metaheuristic algorithm | |
| Mall et al. | A deep (learning) dive into visual search behaviour of breast radiologists | |
| Kovalev et al. | Automatic detection of pathological changes in chest X-ray screening images using deep learning methods | |
| Biswas et al. | DFU_XAI: a deep learning-based approach to diabetic foot ulcer detection using feature explainability | |
| Allen et al. | Robust automated lymph node segmentation with random forests |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |