CN113314205A - Efficient medical image labeling and learning system - Google Patents
Efficient medical image labeling and learning system Download PDFInfo
- Publication number
- CN113314205A CN113314205A CN202110589696.6A CN202110589696A CN113314205A CN 113314205 A CN113314205 A CN 113314205A CN 202110589696 A CN202110589696 A CN 202110589696A CN 113314205 A CN113314205 A CN 113314205A
- Authority
- CN
- China
- Prior art keywords
- training
- model
- learning
- module
- sample
- 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.)
- Granted
Links
Images
Classifications
-
- 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
- G16H30/00—ICT specially adapted for the handling or processing of medical images
- G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/51—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- 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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Epidemiology (AREA)
- Medical Informatics (AREA)
- Primary Health Care (AREA)
- Public Health (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Radiology & Medical Imaging (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention realizes an efficient medical image labeling and learning system by a method in the field of medical image processing. The system comprises: the unsupervised contrast learning pre-training module introduces an unsupervised contrast learning method in the pre-training process based on a MoCo algorithm; a training module based on active learning and contrast learning acquires a pre-trained model and an actual target task data set, and further adjusts the model to adapt to actual requirements by combining difficult case mining of finding the most valuable sample in a non-labeled sample to acquire labels and giving higher weight to misclassified samples and a supervised contrast learning method introducing labeled data; the manual labeling and model training interactive control module provides a labeled interactive interface and a control module for model training. Through the system with the above framework, a universal training framework is realized.
Description
Technical Field
The invention relates to the field of medical image processing, in particular to an efficient medical image labeling and learning system.
Background
At present, various medical images are widely applied in clinic, such as X-ray, CT image, MRI and the like for imaging human body, and pathological images as pathological tissue analysis means.
Due to the characteristics of rapid, simple and popular acquisition of the CT image, the imaging performance based on the lung CT is taken as an important basis in the diagnosis of new coronary pneumonia and other lung diseases, and the important value of the CT image in the diagnosis of various lung diseases is self-evident; because MRI relies on the characteristic of hydrogen atom resonance imaging, the MRI can clearly image the soft tissue with rich water represented by the brain, meanwhile, CT is difficult to image the brain due to the shielding of the skull, and the MRI examination is most common when the brain is examined; the pathological image is obtained by observing the characteristics such as cell morphology after the pathological section is stained and the like, and then judging the type of the pathological change. Although the important value of medical images in clinical events is self-evident, diagnosis using medical images depends on manual interpretation by doctors, but the large data volume of medical images also brings huge burden to doctors, and in recent years, deep learning technology is increasingly widely applied in the field of medical image processing, so that the burden of doctors is effectively reduced and the accuracy of disease diagnosis is improved by means of preliminary screening of medical images or diagnosis assistance of doctors.
However, the main drawback of the deep learning medical image diagnosis method is that the conventional deep learning method depends on a large amount of high-quality labeled data, while labeling the medical image data is time-consuming and expensive, and in contrast, due to the rapid development of the electronic medical record, a large amount of examination result data covering various diseases and without fine labeling is stored in the database of the hospital. In addition to the above problems, the pre-training method based on labeled data still depends on a large amount of labeled data in a similar application scene, and on the other hand, migration learning has defects when problems such as unbalanced class, existence of new classes in a pre-training data set in practical application, deviation between a training set and a test set, and the like are handled.
Therefore, an attempt is made to construct a general learning framework suitable for various medical images based on a contrast learning method, a self-supervision learning method and an active learning method, the general learning framework has the capability of efficiently utilizing label-free data of a large number of related application scenes and less labels in target application scenes for feature extraction and classification learning, and compared with the traditional method, the method can fully explore the value of the label-free data and can reduce the requirement on manual labeling.
The main defects of the current deep learning medical image diagnosis method are that a large amount of finely labeled data is needed no matter the training is directly carried out on a target application scene from a randomly initialized model or a pre-trained model which is trained on other related application scenes is further finely adjusted and adapted to the current target application scene, and the acquisition of data labels is very expensive, time-consuming and labor-consuming; in addition, in the training process of the above mentioned pre-training-fine tuning transfer learning and training-testing-practical application, there may be a class imbalance problem, and the difference between the pre-training data set, the target task training data set and the practical application scenario affects the final use effect.
Disclosure of Invention
Therefore, the invention firstly provides an efficient medical image labeling and learning system, which comprises a pre-training module based on unsupervised contrast learning, a model fine-tuning module based on supervised contrast learning and active learning, and a control module comprising model interaction control and artificial labeling functions:
a first module: the pre-training module based on unsupervised contrast learning: inputting an unmarked data set from the outside, automatically generating positive and negative sample pairs by using input data through using a MoCo algorithm based on comparison learning in an unsupervised learning method, and learning representative characteristics in the samples by the model through comparison between the positive and negative sample pairs so as to realize unsupervised training without manual marking in a pre-training stage;
and a second module: model fine-tuning module based on supervised contrast learning and active learning: obtaining an unsupervised pre-training model through the module I pre-training module, and circularly performing the following processes under an active learning framework on the basis of label-free data until the expected performance index requirement is met: selecting a data sample by a difficult case mining method for marking, optimizing model characteristic representation by supervised comparative learning, adding a classifier at the end of the model, and training and fine-tuning the classifier and the model to realize a target classification task;
and a third module: the control module comprises model interaction control and manual labeling functions: the training process control and model parameter adjustment functions of the module I pre-training module and the module II training fine-tuning module are achieved, manual labeling is obtained through a visual user interface of the manual labeling in combination with the module II model fine-tuning module, and continuous iteration optimization of the model is achieved.
The pre-training module uses the MoCo algorithm, selects Info-NCE Loss as a Loss function and Adam Optimizer as an optimization method, uses all input label-free data to construct a sample pair and train an input model, trains characteristic expressions obtained by a coding part and an optimization model coding part of the model, and realizes pre-training of the model by fully utilizing a large amount of input label-free data.
The specific implementation mode of the MoCo algorithm is as follows: the method comprises the steps of carrying out different random transformation on each original picture to obtain two samples, forming a positive sample pair between the two samples from the same picture, forming a negative sample pair between the two samples from different pictures, carrying out feature extraction and coding on each sample by using the same network as a coder to obtain feature representation of the sample, and then optimizing a model to increase the similarity between the positive sample pairs and reduce the similarity between the negative sample pairs.
The model fine-tuning module repeatedly uses a difficult case mining method to find the most valuable sample with the highest prediction uncertainty in the unmarked data to obtain the label based on the active learning method framework, and further optimizes the feature extraction capability of the model by using the selected sample and the obtained label by using the supervised contrast learning method; and circularly performing the above processes, periodically performing performance evaluation in the training process of the above processes, accessing the model obtained by the above training as an encoder to a classifier, performing fine tuning training on the encoder and the classifier together, testing the fine tuned model by using test data to obtain a performance index, and continuing the above processes until the performance index meets the expectation.
The active learning method is based on a difficult case mining method for finding the most valuable samples in all the labeled samples to reduce generalized errors, and the process comprises the steps of firstly selecting a loss function for estimating future error rate, then respectively estimating error reduction brought by each sample in the unlabeled sample set to a basic classifier, and selecting the sample with the largest estimated value to label.
The unsupervised comparison method has the main problem that the selection and judgment criteria of the positive sample and the negative sample are from the same original sample, namely based on image characteristics, so that the possible problem is that the characteristics of two different types of images which are similar to each other are drawn close, and the classification task is not consistent with the target. The supervised contrast method changes the similarity of the image features into the difference of the image categories according to the optimization criterion by introducing the category labels of the data, thereby avoiding the problems in unsupervised contrast learning, and can ensure that the encoder of the coded image features can distinguish the features which are meaningful for distinguishing the image categories in the image and extract the features by optimizing the distribution of the extracted features in the feature space, drawing the similarity between the features of the images of the same category and increasing the feature similarity between the images of different categories.
The control module comprises a training controller in the model training process and a manual labeling function for obtaining labels required by model training;
the manual labeling function provides the labeling capacity of picture segmentation, a target area of a picture is cut by using polygonal, circular and rectangular tools, then a user-defined label is marked on the target area, the target area is automatically converted into a picture matrix in a tiff format, a target mask is generated, and the labeled picture is automatically added into a training set database; the training controller includes three functions: training, evaluating and predicting, wherein the training function loads training data into a memory or a video memory of a server, and then sets a training environment for the model to train; the evaluation function enables the test set to load evaluation data and an evaluation function and evaluates the model in the memory of the current server; the prediction function performs a prediction of the label for each sample in the predefined prediction set data and saves the prediction results to a database.
The technical effects to be realized by the invention are as follows:
the efficient medical image labeling and learning system provided by the invention completes a set of combined framework from labeling of medical images to model training, has the advantages of learning from a large amount of label-free data to how to extract useful characteristics in images, acquiring labels by selecting the most valuable samples and more effectively utilizing the existing labeled data, and realizes the purposes of reducing dependence on a large amount of labeled data and completing high-quality model training.
It has the following advantages in particular:
1. through an unsupervised contrast learning method, a large amount of label-free data can be utilized to learn an extraction method of image features, so that similar images have more similar feature representation compared with images with larger differences, the distribution of the extracted features in a feature space is optimized, and a foundation is provided for carrying out model fine tuning by using fewer label samples subsequently;
2. through an active learning method, the difficult cases which have the most value on model optimization in the unmarked data are selected in the learning process, and the labels are obtained and learned, so that the overall model labeling requirement is reduced by abandoning labels of a large number of simple cases; and through supervised contrast learning, the feature representation among the images of the same category is drawn, and feature clustering is completed explicitly, so that the learning effect on the existing labeled data is optimized;
3. the method comprises the steps of constructing a training platform for marking one-stop data to model training aiming at the deep learning medical image auxiliary diagnosis application, completing self-supervision pre-training, difficult sample selection needing marking and algorithm flow of model supervised training, and being capable of being flexibly embedded into the existing deep learning medical image auxiliary diagnosis platform and improving the existing algorithm effect;
4. the invention focuses on the training process of the algorithm, is used as a universal training frame, is not limited by specific application scenes and actual algorithm models, can be suitable for classification tasks aiming at different medical image types under different use scenes, can be suitable for more tasks such as segmentation, detection and the like by selecting and modifying different modules of the frame, and is also suitable for more and wider fields, such as defect detection and classification in industrial images, classification of universal natural images and the like.
Drawings
FIG. 1 is a medical image labeling and learning system design framework;
FIG. 2 is an unsupervised contrast learning pre-training module architecture;
FIG. 3 is a training module structure based on an active learning framework;
Detailed Description
The following is a preferred embodiment of the present invention and is further described with reference to the accompanying drawings, but the present invention is not limited to this embodiment.
The invention provides an efficient medical image labeling and learning system, which mainly comprises the steps of introducing unsupervised contrast learning in a pre-training process to utilize a large amount of label-free data; in the later formal training stage, active learning is introduced to select difficult cases from un-labeled data for labeling and training so as to reduce labeling requirements, efficient utilization of limited data is realized through contrast learning, a data labeling platform is built to complete a cyclic training process of labeling, training and re-extracting extracted key cases, and finally, a deep learning network model which is efficient and saves labeling cost is provided on different types of medical images including CT, MRI, pathological pictures and the like at different parts.
The method comprises the following parts and processes, as shown in fig. 1:
1. an unsupervised contrast learning pre-training module based on MoCo;
2. a training module based on active learning, supervised contrast learning and key sample extraction;
3. the manual labeling and model training interaction control module;
unsupervised contrast learning pre-training module
The method comprises the steps of comparing and learning, wherein the similarity between the features of similar images is drawn by optimizing the distribution of the extracted features in a feature space, the feature similarity between the images with larger difference is increased, the optimization of the model representation effect is realized, under the unsupervised setting, different samples obtained by different feature transformation methods from the same image are used as positive examples, and samples from different images are used as negative examples, so that the method can learn how to extract the useful features in the images without manual marking.
The unsupervised contrast learning pre-training module is composed as shown in fig. 2, based on the MoCo algorithm, different random transformations are performed on each original picture to obtain two samples, a positive sample pair is formed between the two samples from the same picture, a negative sample pair is formed between the two samples from different pictures, then the same network is used as an encoder to perform feature extraction and encoding on each sample to obtain feature representation of the sample, then an optimization model increases the similarity between the positive sample pairs and reduces the similarity between the negative sample pairs, and the extraction of valuable and differentiated features in the unsupervised samples is realized through the method. And designing and selecting Info-NCE Loss as a Loss function and Adam Optimizer as an optimization method, constructing a sample pair by using all input label-free data, training an input model, training a coding part of the model and optimizing feature representation obtained by the coding part of the model, and realizing pre-training of the model by fully utilizing a large amount of input label-free data, so that the feature representations of the model obtained from two samples corresponding to the same picture are more similar to the feature representations from different pictures.
Training module based on active learning and contrast learning
The training module based on active learning and comparative learning is shown in fig. 3, and combines a difficult case mining method and a supervised comparative learning method of the most valuable samples from all unlabeled samples based on an active learning framework, and combines with a labeling-training interaction module described later. Through active learning, the most valuable samples in all unmarked samples are found to be marked, the marking result is expressed by using the supervised contrast learning optimization feature, the coding part in the model is preliminarily trained, the overall process of the supervised contrast learning of the part is basically consistent with that of the above unsupervised contrast learning module, the difference is that a class label is introduced as a judgment basis of a positive example and a negative example instead of whether the samples are different enhanced examples from the same picture, and then the whole model including the final classifier is trained by using a normal model fine tuning method.
The active learning is a method for efficiently utilizing label-free data, and the core idea is to search for the least definite sample of the current model for labeling, so that the best effect is achieved by using the least labeled sample. The current labeling sample selection algorithm mainly comprises three methods, namely an uncertainty reduction method, a sample reduction method and a generalization error reduction method. The present invention employs a generalization error reduction-based approach that attempts to select those cases that will minimize future generalization errors. The method comprises the steps of firstly selecting a loss function for estimating future error rate, then respectively estimating error reduction brought by each sample in an unlabeled sample set, and selecting the sample with the largest estimation value for labeling.
In addition, after the important samples are selected and labeled through active learning, a supervised contrast learning method is used for training, so that the model not only focuses on prediction of a final classification result of each picture, but also enables the characteristics of samples of the same type to be similar, distribution of the characteristics in a characteristic space is optimized, difficulty of classification of extracted characteristics by a classifier at the last part of the model is reduced, and the training effect of the model is improved.
The unsupervised comparison method has the main problem that the selection and judgment criteria of the positive sample and the negative sample are from the same original sample, namely based on image characteristics, so that the possible problem is that the characteristics of two different types of images which are similar to each other are drawn close, and the classification task is not consistent with the target. The supervised contrast method changes the similarity of the image features into the difference of the image categories according to the optimization criterion by introducing the category labels of the data, thereby avoiding the problems in unsupervised contrast learning, and can ensure that the encoder of the coded image features can distinguish the features which are meaningful for distinguishing the image categories in the image and extract the features by optimizing the distribution of the extracted features in the feature space, drawing the similarity between the features of the images of the same category and increasing the feature similarity between the images of different categories.
The labeling-training control function in the manual labeling and model training interactive control module is a control module which provides labeled interactive interface design and training, wherein the labeling interface mainly provides labeling capability for picture segmentation, and comprises the steps of cutting a target area of a picture by using polygonal, circular and rectangular tools, then labeling the target area with a self-defined label, and automatically converting the area into a picture matrix in a tiff format and generating a target mask. The marked pictures are automatically added into a training set database, so that the model can be conveniently trained continuously.
The training controller includes three functions, training, evaluation and prediction. The training controller loads training data into a memory or a video memory of the server, and then sets a training environment for the model to train. And the evaluation controller test set loads evaluation data and an evaluation function and evaluates the model in the current server memory. The prediction controller predicts the label for each sample in the predefined prediction set data and stores the prediction result in the database for downloading.
Claims (4)
1. An efficient medical image labeling and learning system is characterized in that: the system comprises the following three system modules:
a first module: the pre-training module based on unsupervised contrast learning: inputting an unmarked data set from the outside, automatically generating positive and negative sample pairs by using input data through using a MoCo algorithm based on comparison learning in an unsupervised learning method, and learning representative characteristics in the samples by the model through comparison between the positive and negative sample pairs so as to realize unsupervised training without manual marking in a pre-training stage;
and a second module: model fine-tuning module based on supervised contrast learning and active learning: obtaining an unsupervised pre-training model through the module I pre-training module, and circularly performing the following processes under an active learning framework on the basis of label-free data until the expected performance index requirement is met: selecting a data sample by a difficult case mining method for marking, optimizing model characteristic representation by supervised comparative learning, adding a classifier at the end of the model, and training and fine-tuning the classifier and the model to realize a target classification task;
and a third module: the control module comprises model interaction control and manual labeling functions: the training process control and the model parameter adjusting functions of the module I pre-training module and the module II fine-tuning module are realized, the module I and the module II fine-tuning module are combined to obtain manual labeling through a visual user interface of the manual labeling, and continuous iterative optimization of the model is realized.
2. The system for efficient labeling and learning of medical images of claim 1, wherein: the pre-training module uses the MoCo algorithm, selects Info-NCE Loss as a Loss function and Adam Optimizer as an optimization method, uses all input label-free data to construct a sample pair and trains an input model, trains characteristic expressions obtained by a coding part and an optimization model coding part of the model, and realizes pre-training of the model by fully utilizing a large amount of input label-free data;
the specific implementation mode of the MoCo algorithm is as follows: the method comprises the steps of carrying out different random transformation on each original picture to obtain two samples, forming a positive sample pair between the two samples from the same picture, forming a negative sample pair between the two samples from different pictures, carrying out feature extraction and coding on each sample by using the same network as a coder to obtain feature representation of the sample, and then optimizing a model to increase the similarity between the positive sample pairs and reduce the similarity between the negative sample pairs.
3. The system for efficient labeling and learning of medical images of claim 1, wherein: the model fine-tuning module repeatedly uses the difficult case mining method to find the most valuable sample acquisition label with the highest prediction uncertainty in the unlabeled data based on the active learning method framework, and further optimizes the feature extraction capability of the model by using the selected sample and the acquired label by using the supervised contrast learning method; performing the above processes in a circulating manner, performing performance evaluation periodically in the training process of the above processes, accessing the model obtained by the above training as an encoder to a classifier, performing fine tuning training on the encoder and the classifier together, testing the model after fine tuning by using test data to obtain performance indexes, and continuing the above processes until the performance indexes meet expectations;
the active learning method is based on a difficult case mining method for finding the most valuable samples in all the labeled samples to reduce generalized errors, and the process is that firstly a loss function is selected to estimate the future error rate, then each sample in the unlabeled sample set is respectively estimated to reduce the errors brought to a basic classifier, and the sample with the largest estimation value is selected to label.
4. The system for efficient labeling and learning of medical images of claim 1, wherein: the control module comprises a training controller in the model training process and a manual labeling function for obtaining labels required by model training;
the manual labeling function provides the labeling capacity of picture segmentation, a target area of a picture is cut by using polygonal, circular and rectangular tools, then a user-defined label is marked on the target area, the target area is automatically converted into a picture matrix in a tiff format, a target mask is generated, and the labeled picture is automatically added into a training set database;
the training controller includes three functions: training, evaluating and predicting, wherein the training function loads training data into a memory or a video memory of a server, and then sets a training environment for the model to train; the evaluation function enables the test set to load evaluation data and an evaluation function and evaluates the model in the memory of the current server; the prediction function performs a prediction of the label for each sample in the predefined prediction set data and saves the prediction results to a database.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110589696.6A CN113314205B (en) | 2021-05-28 | 2021-05-28 | Efficient medical image labeling and learning system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110589696.6A CN113314205B (en) | 2021-05-28 | 2021-05-28 | Efficient medical image labeling and learning system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113314205A true CN113314205A (en) | 2021-08-27 |
| CN113314205B CN113314205B (en) | 2022-05-17 |
Family
ID=77375844
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110589696.6A Active CN113314205B (en) | 2021-05-28 | 2021-05-28 | Efficient medical image labeling and learning system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113314205B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114267087A (en) * | 2022-02-28 | 2022-04-01 | 成都考拉悠然科技有限公司 | Action registration method and system based on hand sample machine learning model |
| CN114399465A (en) * | 2021-12-08 | 2022-04-26 | 紫东信息科技(苏州)有限公司 | Benign and malignant ulcer identification method and system |
| CN114612685A (en) * | 2022-03-22 | 2022-06-10 | 中国科学院空天信息创新研究院 | Self-supervision information extraction method combining depth features and contrast learning |
| CN114708353A (en) * | 2022-06-06 | 2022-07-05 | 中国科学院自动化研究所 | Image reconstruction method and device, electronic equipment and storage medium |
| CN114708465A (en) * | 2022-06-06 | 2022-07-05 | 中国科学院自动化研究所 | Image classification method and device, electronic equipment and storage medium |
| CN114882298A (en) * | 2022-07-11 | 2022-08-09 | 东声(苏州)智能科技有限公司 | Optimization method and device for confrontation complementary learning model |
| CN114970716A (en) * | 2022-05-26 | 2022-08-30 | 支付宝(杭州)信息技术有限公司 | Method and device for training representation model, readable storage medium and computing equipment |
| CN115146716A (en) * | 2022-06-22 | 2022-10-04 | 腾讯科技（深圳）有限公司 | Labeling method, device, equipment, storage medium and program product |
| CN117763348A (en) * | 2023-12-07 | 2024-03-26 | 成都市汇众天智科技有限责任公司 | Automatic driving data labeling system and method |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP4375883A1 (en) * | 2022-11-26 | 2024-05-29 | PreciPoint Group GmbH | Method for operating a distributed digital microscopy system and distributed digital microscopy system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108985334A (en) * | 2018-06-15 | 2018-12-11 | 广州深域信息科技有限公司 | The generic object detection system and method for Active Learning are improved based on self-supervisory process |
| CN111191732A (en) * | 2020-01-03 | 2020-05-22 | 天津大学 | Target detection method based on full-automatic learning |
| CN112069921A (en) * | 2020-08-18 | 2020-12-11 | 浙江大学 | Small sample visual target identification method based on self-supervision knowledge migration |
| CN112381116A (en) * | 2020-10-21 | 2021-02-19 | 福州大学 | Self-supervision image classification method based on contrast learning |
| US20210150710A1 (en) * | 2019-11-15 | 2021-05-20 | Arizona Board Of Regents On Behalf Of Arizona State University | Systems, methods, and apparatuses for implementing a self-supervised chest x-ray image analysis machine-learning model utilizing transferable visual words |
-
2021
- 2021-05-28 CN CN202110589696.6A patent/CN113314205B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108985334A (en) * | 2018-06-15 | 2018-12-11 | 广州深域信息科技有限公司 | The generic object detection system and method for Active Learning are improved based on self-supervisory process |
| US20210150710A1 (en) * | 2019-11-15 | 2021-05-20 | Arizona Board Of Regents On Behalf Of Arizona State University | Systems, methods, and apparatuses for implementing a self-supervised chest x-ray image analysis machine-learning model utilizing transferable visual words |
| CN111191732A (en) * | 2020-01-03 | 2020-05-22 | 天津大学 | Target detection method based on full-automatic learning |
| CN112069921A (en) * | 2020-08-18 | 2020-12-11 | 浙江大学 | Small sample visual target identification method based on self-supervision knowledge migration |
| CN112381116A (en) * | 2020-10-21 | 2021-02-19 | 福州大学 | Self-supervision image classification method based on contrast learning |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114399465A (en) * | 2021-12-08 | 2022-04-26 | 紫东信息科技(苏州)有限公司 | Benign and malignant ulcer identification method and system |
| CN114399465B (en) * | 2021-12-08 | 2022-11-25 | 紫东信息科技(苏州)有限公司 | Benign and malignant ulcer identification method and system |
| CN114267087A (en) * | 2022-02-28 | 2022-04-01 | 成都考拉悠然科技有限公司 | Action registration method and system based on hand sample machine learning model |
| CN114612685A (en) * | 2022-03-22 | 2022-06-10 | 中国科学院空天信息创新研究院 | Self-supervision information extraction method combining depth features and contrast learning |
| CN114612685B (en) * | 2022-03-22 | 2022-12-23 | 中国科学院空天信息创新研究院 | Self-supervision information extraction method combining depth features and contrast learning |
| CN114970716A (en) * | 2022-05-26 | 2022-08-30 | 支付宝(杭州)信息技术有限公司 | Method and device for training representation model, readable storage medium and computing equipment |
| CN114708353A (en) * | 2022-06-06 | 2022-07-05 | 中国科学院自动化研究所 | Image reconstruction method and device, electronic equipment and storage medium |
| CN114708353B (en) * | 2022-06-06 | 2022-09-06 | 中国科学院自动化研究所 | Image reconstruction method and device, electronic equipment and storage medium |
| CN114708465A (en) * | 2022-06-06 | 2022-07-05 | 中国科学院自动化研究所 | Image classification method and device, electronic equipment and storage medium |
| CN115146716A (en) * | 2022-06-22 | 2022-10-04 | 腾讯科技（深圳）有限公司 | Labeling method, device, equipment, storage medium and program product |
| CN115146716B (en) * | 2022-06-22 | 2024-06-14 | 腾讯科技（深圳）有限公司 | Labeling method, labeling device, labeling apparatus, labeling storage medium and labeling program product |
| CN114882298B (en) * | 2022-07-11 | 2022-11-01 | 东声(苏州)智能科技有限公司 | Optimization method and device for confrontation complementary learning model |
| CN114882298A (en) * | 2022-07-11 | 2022-08-09 | 东声(苏州)智能科技有限公司 | Optimization method and device for confrontation complementary learning model |
| CN117763348A (en) * | 2023-12-07 | 2024-03-26 | 成都市汇众天智科技有限责任公司 | Automatic driving data labeling system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113314205B (en) | 2022-05-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113314205B (en) | Efficient medical image labeling and learning system | |
| CN111444960A (en) | Skin disease image classification system based on multi-mode data input | |
| CN109544518B (en) | Method and system applied to bone maturity assessment | |
| CN111583285B (en) | Liver image semantic segmentation method based on edge attention strategy | |
| CN110992351B (en) | sMRI image classification method and device based on multi-input convolution neural network | |
| CN111738302B (en) | System for classifying and diagnosing Alzheimer's disease based on multi-modal data | |
| CN111476805A (en) | Cross-source unsupervised domain adaptive segmentation model based on multiple constraints | |
| CN111242233B (en) | Alzheimer disease classification method based on fusion network | |
| CN112132878B (en) | End-to-end brain nuclear magnetic resonance image registration method based on convolutional neural network | |
| CN112950480A (en) | Super-resolution reconstruction method integrating multiple receptive fields and dense residual attention | |
| CN116503668A (en) | Medical image classification method based on small sample element learning | |
| CN113781461A (en) | Intelligent patient monitoring and sequencing method | |
| CN116630289A (en) | Brain glioma CDKN2A/B gene state classification prediction method based on multi-mode MRI | |
| CN114417836A (en) | Deep learning-based Chinese electronic medical record text semantic segmentation method | |
| CN113920379A (en) | Zero sample image classification method based on knowledge assistance | |
| CN116434950B (en) | Diagnosis system for autism spectrum disorder based on data clustering and ensemble learning | |
| CN114972291B (en) | Medical image structured automatic labeling method and system based on hybrid enhanced intelligence | |
| WO2023226217A1 (en) | Microsatellite instability prediction system and construction method therefor, terminal device, and medium | |
| CN114463320B (en) | Magnetic resonance imaging brain glioma IDH gene prediction method and system | |
| CN114359194B (en) | Multimode cerebral apoplexy infarct region image processing method based on improved U-Net network | |
| CN112686912B (en) | Acute stroke lesion segmentation method based on gradual learning and mixed samples | |
| CN114283301A (en) | Self-adaptive medical image classification method and system based on Transformer | |
| CN112086174A (en) | Three-dimensional knowledge diagnosis model construction method and system | |
| CN117974693B (en) | Image segmentation method, device, computer equipment and storage medium | |
| CN117174261B (en) | Multi-type labeling flow integrating system for medical images |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |