CN111414946B - Artificial intelligence-based medical image noise data identification method and related device - Google Patents
Artificial intelligence-based medical image noise data identification method and related device Download PDFInfo
- Publication number
- CN111414946B CN111414946B CN202010170514.7A CN202010170514A CN111414946B CN 111414946 B CN111414946 B CN 111414946B CN 202010170514 A CN202010170514 A CN 202010170514A CN 111414946 B CN111414946 B CN 111414946B
- Authority
- CN
- China
- Prior art keywords
- data
- label
- recognition
- noise
- model
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computing Systems (AREA)
- Evolutionary Biology (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Epidemiology (AREA)
- Medical Informatics (AREA)
- Primary Health Care (AREA)
- Public Health (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the application discloses a medical image noise data identification method and a related device based on artificial intelligence, aiming at a medical image with a label, if the medical image needs to be identified whether noise data exists, the medical image can be used as label data to be identified, and the medical image can be identified through a first identification model. Because the first recognition model has better anti-noise capability, the confidence degree of the first recognition model can effectively express the possibility that the data of the label to be recognized belongs to the noise data according to the recognition result obtained by the first recognition model, so that the medical image with the wrong label can be effectively screened out from the medical image with the label, and the accuracy of determining the noise data is improved.
Description
Technical Field
The present application relates to the field of data processing, and more particularly, to a method and related apparatus for recognizing noise data of medical images based on artificial intelligence.
Background
When the network model is trained, the label data can be used as a training sample to carry out supervised training or semi-supervised training on the network model. The label data may be understood as data carrying a label for identifying specific contents embodied in the data, for example, in the label data related to identification of a lesion area in the medical image, the label of the medical image may indicate whether the medical image has a lesion area, or a type of a lesion in the lesion area, and the like.
However, the label of not all the data is accurate, and if the label of one label data is actually incorrect, i.e., the content identified by the label is not embodied in the data, or the data does not relate to the content identified by the label, then the one label data belongs to the noise data. If the network model is trained by using the noise data, the network model is overfitting to the noise, and finally a model with a poor effect is obtained. For example, in the field of medical imaging, even experts cannot give accurate labeling information due to the complexity of lesion identification.
In the related art, noise data is generally screened by a statistical method, for example, noise data is taken as a statistical abnormal value, and noise data is screened from a large amount of tag data by a method using abnormal value detection. However, many noise data do not belong to an abnormal value under statistical rules, for example, a medical image in which a lesion is identified by a label, which is supposed to belong to noise data, but is statistically undeterminable as an abnormal value, and a lesion formed after a surgery on a lesion area is shown although the lesion is not shown.
Therefore, in the related art, the accuracy of the noise data screening method for medical images is not high, and the current noise data identification requirement is difficult to meet.
Disclosure of Invention
In order to solve the technical problem, the application provides a medical image noise data identification method and a related device based on artificial intelligence, and the determination accuracy of noise data is improved.
The embodiment of the application discloses the following technical scheme:
in one aspect, an embodiment of the present application provides a method for identifying noise data of a medical image, where the method includes:
determining an identification result of the tag data to be identified according to the first identification model; the first recognition model is obtained by performing N rounds of training on an initial network model, N is less than the total number M of rounds required for completing the complete training of the initial network model, and the model parameters of the first recognition model are not converged;
and determining whether the tag data to be recognized is noise data or not according to the confidence degree embodied by the recognition result.
In another aspect, an embodiment of the present application provides a medical image noise data identification apparatus, including a first determination unit and a second determination unit:
the first determining unit is used for determining the identification result of the tag data to be identified according to the first identification model; the first recognition model is obtained by performing N rounds of training on an initial network model, N is less than the total number M of rounds required for completing the complete training of the initial network model, and the model parameters of the first recognition model are not converged;
and the second determining unit is used for determining whether the tag data to be identified is noise data according to the confidence degree embodied by the identification result.
In another aspect, an embodiment of the present application provides a noise data identification apparatus for medical imaging, the apparatus including a processor and a memory:
the memory is used for storing program codes and transmitting the program codes to the processor;
the processor is configured to perform the method of the above aspect according to instructions in the program code.
In another aspect, the present application provides a computer-readable storage medium for storing a computer program for executing the method of the above aspect.
According to the technical scheme, if the medical image with the label needs to be identified whether noise data exists, the medical image can be used as the label data to be identified, and the medical image can be identified through the specific first identification model. The first recognition model is obtained by performing N rounds of training on the initial network model, because N is less than the total number of rounds M required for completing the complete training on the initial network model, model parameters in the trained first recognition model are not converged, the initial training is just completed, the first recognition model learns common characteristics of most similar data in a training sample through training, and because non-noise data in a medical image serving as the training sample, namely the occupation ratio of a correct label is large, the common characteristics of the label data are mainly learned by the first recognition model, the noise data are not fitted, and the noise resisting capability is good. Due to the anti-noise capability of the first recognition model, the confidence degree embodied by the recognition result obtained by the first recognition model can effectively express the possibility that the label data to be recognized belongs to the noise data, so that the medical image with the wrong label can be effectively screened out from the medical images with the labels, and the accuracy of determining the noise data is improved.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to these drawings without inventive exercise.
Fig. 1 is a schematic view of an application scenario of a noise data identification method for medical images according to an embodiment of the present application;
fig. 2 is a schematic flowchart of a method for recognizing noise data of medical images according to an embodiment of the present disclosure;
fig. 3 is a schematic flowchart of another method for recognizing noise data of medical images according to an embodiment of the present application;
fig. 4 is a schematic structural diagram of a noise data identification apparatus for medical images according to an embodiment of the present disclosure;
fig. 5 is a schematic structural diagram of a server according to an embodiment of the present application;
fig. 6 is a schematic structural diagram of a terminal device according to an embodiment of the present application.
Detailed Description
Embodiments of the present application are described below with reference to the accompanying drawings.
In order to more effectively screen out medical images with wrong labels, the embodiment of the application provides a medical image noise data identification method based on artificial intelligence.
The method for recognizing the noise data of the medical image is realized based on Artificial Intelligence (AI), which is a theory, a method, a technology and an application system for simulating, extending and expanding human Intelligence by using a digital computer or a machine controlled by the digital computer, sensing the environment, acquiring knowledge and obtaining the best result by using the knowledge. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.
The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence base technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.
In the embodiment of the present application, the artificial intelligence software technology mainly involved includes the above-mentioned computer vision technology, machine learning/deep learning, and the like.
For example, Image Processing (Image Processing), Image Semantic Understanding (ISU), video Processing (video Processing), Video Semantic Understanding (VSU), face recognition (face recognition), and the like in Computer Vision (Computer Vision) may be involved.
For example, Deep Learning (Deep Learning) in Machine Learning (ML) may be involved, including various types of Artificial Neural Networks (ANN).
In order to facilitate understanding of the technical solution of the present application, the following describes a method for recognizing noise data of medical images provided in the embodiments of the present application with reference to an actual application scenario.
The method for recognizing the noise data of the medical image can be applied to noise data recognition equipment of the medical image with data processing capacity, such as terminal equipment and a server. The terminal device may be a smart phone, a computer, a Personal Digital Assistant (PDA), a tablet computer, or the like; the server may be specifically an independent server or a cluster server.
The data processing equipment can have the capability of implementing a computer vision technology, wherein the computer vision is a science for researching how to enable a machine to see, and further means that a camera and a computer are used for replacing machine vision such as human eyes for identifying and measuring a target, and further image processing is carried out, so that the computer processing becomes an image which is more suitable for human eyes to observe or is transmitted to an instrument for detection. As a scientific discipline, computer vision research-related theories and techniques attempt to build artificial intelligence systems that can capture information from images or multidimensional data. Computer vision technologies generally include image processing, image Recognition, image semantic understanding, image retrieval, Character Recognition (OCR), video processing, video semantic understanding, video content/behavior Recognition, three-dimensional object reconstruction, 3D technologies, virtual reality, augmented reality, synchronous positioning, map construction, and other technologies, and also include common biometric technologies such as face Recognition and fingerprint Recognition.
In the embodiment of the application, the data processing device can identify and detect the noise data in the medical image with the label through the computer vision technology.
The data processing device may be ML capable. ML is a multi-field interdisciplinary, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks.
The noise data identification method for the medical image mainly relates to application of various artificial neural networks.
The following describes a method for recognizing noise data of medical images according to an embodiment of the present application, with reference to an application scenario and fig. 1. In the application scenario shown in fig. 1, a server 101, a scanner 102 and a display screen 103 are included.
During the application process, the scanner 102 may be used to scan a pathological section to obtain a medical image, and upload the medical image to the server 101. The server 101 may store the acquired medical image in an internal memory. Medical care personnel can label the focus area included in the medical image by calling the medical image in the memory of the server 101, so that the medical image with the label is obtained as the data of the label to be identified.
In the server 101, a first recognition model is deployed for recognizing an object to be recognized in tag data to be recognized. The first recognition model may set a network structure according to an application scenario and a recognition task. In the application scenario shown in fig. 1, the first recognition model may be a deep convolutional neural network for recognizing a lesion region in the medical image with a label.
The first recognition model is obtained after the initial network model is initially trained, namely N rounds of training are carried out on the initial network model, N is smaller than the total number M of rounds required for completing the complete training of the initial network model, and model parameters of the first recognition model are not converged. M identifies the total number of rounds required to fully train the initial network model, i.e., the number of trains that result in the initial network model being in a converged state. N identifies the number of rounds used to initially train the initial network model to obtain the first recognition model.
The purpose of performing initial training on the initial network model is to enable the initial network model to learn common characteristics of most similar data in a training sample set, avoid overfitting noise data in the training sample set, obtain a first recognition model with good anti-noise capability, and use the first recognition model for recognizing the noise data in the label data. In the application scenario shown in fig. 1, the training sample set may be a plurality of medical images with labels, and the server 101 may perform initial training on the initial network model by using the training sample set to obtain a first recognition model, and use the first recognition model for determining noise data, that is, the medical image with the wrong label.
During application, the number of rounds N required for initial training of the initial network model may be quantified as 1/10 of the total number of model training rounds M. Regardless of the domain or scenario involved in a particular noise determination task, N may generally take on a value of about 5-10. When the data volume of the tag to be identified is large and the initial network model structure is complex, N may be adjusted to 20. In practical application, the value of N can be determined and adjusted according to a specific application scenario and noise. In the application scenario shown in fig. 1, the server 101 may set N to 10 through the processor, and train 10 rounds on the initial network model using the medical images with labels in the training sample set to obtain a first recognition model.
It should be noted that, in the process of initially training the initial network model to obtain the first recognition model, in order to obtain better anti-noise performance of the first recognition model, in the training process, the model may be trained with a larger learning rate, so that the model learns broader contents, that is, common characteristics of most similar data in the training sample set, and reduces the attention degree to fine features (e.g., noise features) in the training sample set, thereby avoiding overfitting of noise data, that is, features related to noise are not learned, and thus achieving the anti-noise performance of the model. Generally, the learning rate can take on a value of 0.01 to 0.1. In practical application, the learning rate can be set and adjusted according to the complexity of the model, the data quantity of the tag to be identified and the noise. In the application scenario shown in fig. 1, the server 101 may set the model learning rate to 0.1 through the processor, and then train the initial network model for 10 rounds to obtain the first recognition model, and use the first recognition model to determine the noise data, i.e. the medical image with the wrong identification label.
Since the proportion of non-noise data, i.e., correct tag data, in the training sample is large, the first recognition model obtained after the initial training learns the common features of most similar data (correct tag data), and noise data is not fitted, so that the method has good anti-noise capability.
After obtaining the first identification model, the server 101 may identify the tag data to be identified by using the first identification model, and determine an identification result of the tag data to be identified. Wherein, the label data to be identified is a medical image with a label. The label is used for identifying whether a focus area or a focus type of the focus area is included in the label data to be identified. The identification result is used for identifying whether the first identification model includes the prediction result of the lesion area in the label data to be identified. In the application scenario shown in fig. 1, the tag data to be recognized may be a stomach section image with a tag, where the tag of the stomach section image identifies whether the stomach section image includes a gastric cancer cell region, and the recognition of the stomach section image by the first recognition model may result in a recognition result, where the recognition result identifies whether the stomach section image includes a gastric cancer cell region predicted by the first recognition model.
Further, the server 101 may determine, by using the processor, whether the tag to be recognized is the noise data according to the confidence level represented by the recognition result corresponding to the tag data to be recognized, that is, whether the tag of the medical image is labeled with an error. Here, the confidence level may be understood as a confidence level of the recognition result corresponding to the data to be labeled. The confidence degree represents the possibility that the tag data to be recognized belongs to the noise data from the perspective of the possibility that the tag data to be recognized is recognized as a certain class by the first recognition model.
If the confidence degree embodied by the identification result is lower, the first identification model indicates that the identification result of the to-be-identified label data is determined to be relatively inconclusive. That is, since the first recognition model itself has noise immunity, if the first recognition model cannot accurately determine the category of the tag data to be recognized, it is proved that the features of the tag data to be recognized are not obvious enough relative to the common features of the correct tag data, or the tag data to be recognized has a high possibility of carrying the features related to noise (because the first recognition model has not learned the features related to noise yet). Therefore, if the first recognition model cannot accurately determine the category of the tag data to be recognized, the tag data to be recognized has a high possibility of being noise data.
If the confidence degree embodied by the identification result is higher, the first identification model is shown to determine that the identification result of the to-be-identified label data is more confident. That is, since the first recognition model itself has noise immunity, if the first recognition model can accurately determine the category of the tag data to be recognized, it is proved that the features of the tag data to be recognized are more obvious than the common features of the correct tag data, or the tag data to be recognized has a lower possibility of carrying the features related to noise (because the first recognition model has not learned the features related to noise yet). Therefore, if the identification result is different from the label corresponding to the label data to be identified, it indicates that the label data to be identified has a high possibility of being noise data.
In the application scenario shown in fig. 1, if the confidence level of the recognition result obtained after the first recognition model recognizes the stomach slice image is low, which indicates that the recognition result of the first recognition model on the stomach slice image is relatively inconclusive, the stomach slice image may be determined as noise data, that is, it is considered that the label labeling of the stomach slice image is incorrect. If the confidence degree embodied by the identification result obtained by the first identification model after identifying the stomach section image is higher, which indicates that the identification result of the first identification model on the stomach section image is more confident, the identification result of the stomach section image is compared with the label, and if the comparison result indicates that the identification result is inconsistent with the label, the stomach section image can be determined as noise data, namely, the label of the stomach section image is considered to be wrongly labeled.
Due to the anti-noise capability of the first recognition model, the confidence degree embodied by the recognition result obtained by the first recognition model can effectively express the possibility that the label data to be recognized belongs to the noise data, so that the medical image with the wrong label can be effectively screened out from the medical images with the labels, and the accuracy of determining the noise data is improved.
The server 101 may store the screened medical image in the internal memory after screening the medical image with the wrong label by executing the method for identifying the noise data of the medical image provided in the above embodiment. Medical personnel can call and view the medical images with the wrong label marking through the display screen 103 and re-mark the medical images.
Referring to fig. 2, fig. 2 is a schematic flowchart of a noise data identification method for medical images according to an embodiment of the present disclosure. For convenience of description, the following embodiments describe the noise data identification method of the medical image with a server as an execution subject. As shown in fig. 2, the method comprises the steps of:
s201, determining the identification result of the tag data to be identified according to the first identification model.
In the training scene of supervised learning of the deep learning model, the training samples are label data with labeling information. In the field of medical imaging, even medical experts cannot give accurate labeling information due to the complexity of lesion identification. Therefore, for medical images whose label data is labeled, there is no guarantee that the label is hundred percent correct. Therefore, the medical image with the label can be used as the label data to be identified, and the identification result of the label data to be identified is determined by using the first identification model, so that whether the label data to be identified is noise data or not can be determined, namely, the medical image with the wrong label can be identified. The data of the label to be identified is a medical image with a label, and the label is used for identifying whether the medical image comprises a focus area or a focus type of the focus area. The medical image may be any type of pathological image, such as a lung CT image, a stomach slice image, and so forth. The recognition result identifies a prediction result of the first recognition model for the tag data to be recognized. In the application scenario shown in fig. 1, in the task of identifying the stomach slice image by the server 101 using the first identification model, the identification result includes: gastric cancer cell regions are included in the gastric section images, and gastric cancer cell regions are not included in the gastric section images.
The first recognition model is obtained by performing N rounds of training on an initial network model, wherein N is less than the total number M of rounds required for completing the complete training of the initial network model, and model parameters of the first recognition model are not converged. The initial network model is a pre-established parameter initialization model, namely, the model parameters are initial values and have not been trained and adjusted. The initial Network model may be constructed by various artificial Neural networks, such as a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), and the like, and may be constructed according to a specific application scenario and a noise data recognition task of a medical image. M identifies the total number of training cycles to train the initial network model such that the model converges. N identifies the number of training cycles to initially train the initial network model so that the model has noise immunity.
In the process of performing initial training on the initial network model to obtain the first recognition model, the initial network model may be initially trained by using a training sample set, where training samples in the training sample set belong to tag data used in the same learning task. It can be understood that most samples in the training sample set have similar or identical features, so that in the initial training process, the initial network model can easily learn the common characteristics of most samples, and overfitting to noise data is avoided, so that the first recognition model with anti-noise performance is obtained.
In order to make the first recognition model have anti-noise performance, in the actual training process, no matter the complexity of the model or the noise determination task, a large learning rate (for example, the learning rate a is 0.1) may be set, and the initial network model is trained for N cycles to obtain the first recognition model for the noise data recognition task of the medical image.
The initial network model can reduce the attention degree of the initial network model to the fine features due to the large learning rate, and in the initial training process, the attention of the initial network model can be focused on the common characteristics of the training samples for learning, so that the first recognition model after the initial training does not learn the features of the noise, and has the anti-noise performance. Therefore, the influence of the noise immunity can be reflected in the identification result of the tag data to be identified determined by the first identification model.
And S202, determining whether the tag data to be identified is noise data or not according to the confidence degree embodied by the identification result.
Since the first recognition model has a certain anti-noise performance, the confidence level of the recognition result of the tag data to be recognized output by the first recognition model can be used for determining whether the tag data to be recognized is noise data, and the noise data is a medical image with a wrong tag label. Wherein the confidence level embodied by the recognition result identifies the trustworthiness of this recognition result. The confidence degree of the first recognition model for the recognition result is reflected from the perspective of the confidence degree of the recognition result. If the confidence degree embodied by the recognition result is higher, the first recognition model is more confident for the recognition result; and if the confidence degree embodied by the recognition result is lower, the first recognition model is less confident about the recognition result. In the application process, the confidence level embodied by the recognition result can be compared with a threshold value, and the confidence level of the first recognition model for the recognition result can be determined according to the comparison result.
In a possible implementation manner, if the confidence level embodied by the recognition result is lower than the first conditional threshold, the tag data to be recognized corresponding to the recognition result is determined to be noise data.
In practical applications, the first conditional threshold may be a fixed value set in advance. The first conditional threshold may also be set to a certain value, and then the first conditional threshold is used to perform noise screening on a certain amount of tag data to be identified, and then the size of the first conditional threshold is adjusted according to the identification accuracy of the noise data. For example, a first condition threshold a is preset, after 1000 pieces of tag data to be recognized are subjected to noise screening according to the confidence level represented by the recognition result and the first condition threshold a, 100 pieces of tag data to be recognized are screened out as noise data, and after the 100 pieces of noise data are corrected, if 10 pieces of tag data are correct in the 100 pieces of noise data, the recognition accuracy of the noise data is 90%. If it is desired to further increase the recognition accuracy of the noise data (up to 95%), the first conditional threshold a may be appropriately adjusted to b.
If the confidence level embodied by the recognition result is lower than the first condition threshold, the confidence level embodied by the recognition result is lower, namely the first recognition model is less confident for the recognition result. Because the first identification model has the anti-noise performance, if the first identification model cannot accurately identify the category of the tag data to be identified (that is, the identification result corresponding to the tag data to be identified by the first identification model is relatively unreliable), it indicates that the common characteristic of the tag data to be identified relative to the correct tag data is not obvious enough, or the possibility that the tag data to be identified carries the characteristic related to noise is high. Therefore, if the confidence level embodied by the identification result is lower than the first condition threshold value, the tag data to be identified can be determined as noise data.
In another possible implementation manner, if the confidence level embodied by the identification result is higher than a second condition threshold, determining whether the tag of the tag data to be identified is consistent with the identification result; and if the data are inconsistent, determining the tag data to be identified corresponding to the identification result as noise data.
In practical applications, the second conditional threshold may be set to a fixed value, or may be set first and then adjusted, which is similar to the first conditional threshold and will not be described herein again. It should be noted that the first conditional threshold and the second conditional threshold may be equal or may not be equal. Specifically, the first condition threshold and the second condition threshold may be set according to the noise determination task.
If the confidence level of the recognition result is higher than the second condition threshold, it indicates that the confidence level of the recognition result is higher, i.e. the first recognition model is more confident for the recognition result. Because the first identification model has the anti-noise performance, if the first identification model can accurately identify the type of the tag data to be identified (that is, the identification result corresponding to the tag data to be identified by the first identification model is relatively confident), it indicates that the common characteristic of the tag data to be identified relative to the correct tag data is relatively obvious, or the probability that the tag data to be identified carries the characteristic related to noise is relatively low. Therefore, if the confidence level of the recognition result is higher than the second condition threshold and the tag of the tag data to be recognized is inconsistent with the recognition result, the tag data to be recognized may be determined as the noise data.
Based on the above, the confidence level embodied by the recognition result is compared with the first condition threshold or the second condition threshold, so as to determine the confidence level of the first recognition model for the recognition result. And noise data is selected according to the comparison result, so that the labeling accuracy of the label data to be identified is improved.
There may be a variety of different ways of determining the confidence level embodied in the recognition result.
In one possible implementation, the confidence level embodied by the recognition result may be determined by identifying a probability distribution of the recognition result.
The probability distribution of the identification result can be understood as the probability value condition of the tag data to be identified as each category. In practical application, the category corresponding to the maximum probability value in the probability distribution can be determined as the recognition result corresponding to the tag data to be recognized, and the maximum probability value is determined as the confidence degree embodied by the recognition result.
In the application scenario shown in fig. 1, in the task of identifying whether the stomach section image includes the gastric cancer cell region, the first identification model is used to identify the stomach section image, and the probability distribution is obtained as follows: the probability that the stomach section image includes the stomach cancer cell region is 0.9, the probability that the stomach section image does not include the stomach cancer cell region is 0.1, and the confidence coefficient embodied by the recognition result can be determined to be 0.9 according to the maximum probability value 0.9 in the probability distribution when the recognition result of the stomach section image includes the stomach cancer cell region.
In another possible implementation manner, the information entropy corresponding to the probability distribution of the identification result may be determined, and then the confidence level embodied by the identification result may be determined according to the information entropy.
The information entropy identifies uncertainty of the recognition result, that is, the difference degree of probability values corresponding to each category of probability distribution of the recognition result. The smaller the information entropy is, the smaller the uncertainty of the recognition result is, the larger the difference of the probability values corresponding to the categories in the probability distribution for identifying the recognition result is, and at this time, the first recognition model is more confident about the recognition result, that is, the confidence degree embodied by the recognition result is higher. The larger the information entropy is, the larger the uncertainty of the recognition result is, the smaller the difference of probability values corresponding to each category in the probability distribution for identifying the recognition result is, and at this time, the first recognition model is less confident about the recognition result, that is, the confidence degree embodied by the recognition result is lower.
The information entropy corresponding to the probability distribution of the identification recognition result can be calculated by the following mathematical expression:
wherein p (x) i ) The probability value of the tag data to be identified as a certain category is shown, and n represents the total number of the categories. HAnd (X) represents information entropy corresponding to probability distribution of the identification result.
For example, in the application scenario shown in fig. 1, if the probability that the stomach section image includes the gastric cancer cell region is 0.9, and the probability that the stomach section image does not include the gastric cancer cell region is 0.1, the information entropy h (x) corresponding to the probability distribution is:
H(X)＝-0.9*lg(0.9)-0.1*lg(0.1)≈0.14
if the corresponding probability distribution of the stomach section image is as follows: the probability that the stomach section image includes the gastric cancer cell region is 0.6, the probability that the stomach section image does not include the gastric cancer cell region is 0.4, and the information entropy h (x) corresponding to the probability distribution is:
H(X)＝-0.6*lg(0.6)-0.4*lg(0.4)≈0.29
due to the anti-noise capability of the first recognition model, the confidence degree embodied by the recognition result obtained by the first recognition model can effectively express the possibility that the to-be-recognized label data belongs to the noise data, so that the medical image with the wrong label can be effectively screened out from the medical images with the labels, and the determination accuracy of the noise data is improved.
In practical application, the first identification model can be used for repeatedly screening the data of the tag to be identified for many times, so that the tag accuracy of the tag data in the data of the tag to be identified can be improved.
Another method for recognizing noise data of medical images is described below with reference to fig. 3. In the method shown in fig. 3, the method including steps S301 to S302 are the same as steps S201 to S202 described above, and are not described again here. The method further comprises the following steps:
s303, acquiring the re-labeling data.
If the data of the tag to be identified is determined to be noise data, the data of the tag to be identified has a high possibility of carrying a tag with an error. Thus, the noisy data may be relabeled, i.e. relabeled data may be obtained.
It is to be understood that, the tag data in the training sample set used by the initial network model may be used as the tag data to be recognized, and the first recognition model is used to screen out the noise data in the training sample set. And re-labeling the screened noise data to obtain re-labeled data.
And S304, generating an adjusting sample set according to the re-labeling data and the residual data in the training sample set.
After the noise data is filtered out from the training sample set based on steps S301-S302, the remaining data in the training sample set is the label data except the noise data in the training sample set. Thus, the set of adjustment samples may be generated using the re-labeling data and the remaining data in the set of training samples.
S305, taking the adjusting sample set as the training sample set to perform N rounds of training on the initial network model again to obtain a second recognition model, and performing noise data recognition on the medical image based on the second recognition model.
Because the labels of the label data in the adjustment sample set are labeled manually, the label data in the adjustment sample set may still carry the wrong labels. In order to further reduce the amount of noise data in the adjustment sample set, the adjustment sample set may be used as a training sample set, the initial network model is retrained to obtain a second recognition model, and the second recognition model is used to recognize the noise data in the medical image.
In practical applications, the steps S301 to S305 may be repeated 1 to 3 times. By screening the noise data for multiple times and re-labeling the noise data, the accuracy of the labels corresponding to the label data in the training sample set is improved, and the proportion of the noise data in the training sample set is reduced.
The adjusted sample set is used as the training sample set to retrain the initial network model, and the proportion of correct label data in the adjusted sample set is increased, so that the common characteristics of the correct label data in the training sample set are more obvious, and the anti-noise performance of the obtained second recognition model is better. Furthermore, the second recognition model is used for screening the noise data in the training sample set, and the proportion of correct label data in the training sample set is further improved.
In the noise data identification method for medical images provided in the above embodiment, if it is necessary to identify whether there is noise data in a medical image with a tag, the medical image may be used as tag data to be identified, and may be identified by a specific first identification model. The first recognition model is obtained by performing N rounds of training on the initial network model, because N is less than the total number of rounds M required for completing the complete training on the initial network model, model parameters in the trained first recognition model are not converged, the initial training is just completed, the first recognition model learns common characteristics of most similar data in a training sample through training, and because non-noise data in a medical image serving as the training sample, namely the occupation ratio of a correct label is large, the common characteristics of the label data are mainly learned by the first recognition model, the noise data are not fitted, and the noise resisting capability is good. Due to the anti-noise capability of the first recognition model, the confidence degree embodied by the recognition result obtained by the first recognition model can effectively express the possibility that the label data to be recognized belongs to the noise data, so that the medical image with the wrong label can be effectively screened out from the medical images with the labels, and the accuracy of determining the noise data is improved.
In order to facilitate understanding of the method for recognizing noise data of medical images provided in the foregoing embodiments, the method for recognizing noise data of medical images is described below with reference to a specific application scenario. For convenience of description, a server is used as an execution subject in the following.
In the application scenario shown in fig. 1, the server 101 uses the processor to identify the labeled stomach section image through the first identification model, determine a probability value of whether the stomach section image includes a stomach cancer cell region, and use the probability value as the identification result. The first recognition model is obtained by training a pre-built deep convolutional neural network for 10 rounds by using a training sample set, and the learning rate can be set to be 0.1 in the training process; the training sample set comprises a stomach section image with a label, and serves as label data to be identified.
The server 101 may utilize the processor to calculate a corresponding information entropy according to the recognition result, and compare the information entropy with a preset threshold to determine a confidence of the recognition result. And if the information entropy is larger than the threshold value, and the identification result is inconsistent with the digital slice image to be identified, or the information entropy is smaller than the threshold value, determining the stomach slice image as noise data, namely, considering that the label labeling of the stomach slice image is wrong. And the medical staff can obtain the re-labeled stomach slice image after re-labeling the screened stomach slice image.
After the server 101 acquires the relabeled stomach slice image, an adjustment sample set is generated by the relabeled stomach slice image and the stomach slice image with the correct label in the training sample set, the adjustment sample set is used as the training sample set to perform 10 rounds of training on the initial network model again, and the retrained first recognition model is used for determining noise data in the adjustment sample set. After 3 times of noise data screening are repeated, a stomach section image with high label accuracy can be obtained.
For the above-described method for recognizing noise data of medical images, the embodiment of the present application further provides a corresponding device for recognizing noise data of medical images.
Referring to fig. 4, fig. 4 is a schematic structural diagram of a medical image noise data identification apparatus according to an embodiment of the present disclosure. As shown in fig. 4, the noise data identification device 400 for medical images includes a first determination unit 401 and a second determination unit 402:
the first determining unit 401 is configured to determine an identification result of the tag data to be identified according to the first identification model; the first recognition model is obtained by carrying out N rounds of training on an initial network model, N is less than the total number M of rounds required by complete training on the initial network model, and model parameters of the first recognition model are not converged; the label data to be identified is a medical image with a label;
the second determining unit 402 is configured to determine whether the tag data to be identified is noise data according to the confidence level embodied by the identification result; the noise data is a medical image with a label with an error.
Wherein the second determining unit 402 is configured to:
and if the confidence degree embodied by the recognition result is lower than a first condition threshold value, determining the tag data to be recognized corresponding to the recognition result as noise data.
Wherein the second determining unit 402 is configured to:
if the confidence degree embodied by the identification result is higher than a second condition threshold value, determining whether the label of the label data to be identified is consistent with the identification result;
and if the identification result is inconsistent, determining the to-be-identified label data corresponding to the identification result as noise data.
Wherein the second determining unit 402 is further configured to:
determining the confidence embodied by the recognition result by identifying the probability distribution of the recognition result.
Wherein the second determining unit 402 is configured to:
determining information entropy corresponding to the probability distribution for identifying the recognition result;
and determining the confidence degree embodied by the recognition result according to the information entropy.
And the label data to be identified is label data in a training sample set used for training the initial network model.
The device further comprises an acquisition unit, a generation unit and a training unit:
the acquiring unit is used for acquiring re-labeling data, and the re-labeling data is obtained by re-labeling the noise data with labels;
the generating unit is used for generating an adjusting sample set according to the re-labeling data and the residual data in the training sample set, wherein the residual data is label data except noise data in the training sample set;
the training unit is configured to perform N rounds of training on the initial network model again by using the adjusted sample set as the training sample set, use the trained model as the first recognition model, and trigger the first determining unit 401.
The noise data identification device for medical images provided in the above embodiments may identify, for a medical image with a tag, if it is necessary to identify whether there is noise data in the medical image, the medical image as tag data to be identified, and identify the medical image by using a specific first identification model. The first recognition model is obtained by performing N rounds of training on the initial network model, because N is less than the total number of rounds M required for completing the complete training on the initial network model, model parameters in the trained first recognition model are not converged, the initial training is just completed, the first recognition model learns common characteristics of most similar data in a training sample through training, and because non-noise data in a medical image serving as the training sample, namely the occupation ratio of a correct label is large, the common characteristics of the label data are mainly learned by the first recognition model, the noise data are not fitted, and the noise resisting capability is good. Due to the anti-noise capability of the first recognition model, the confidence degree embodied by the recognition result obtained by the first recognition model can effectively express the possibility that the label data to be recognized belongs to the noise data, so that the medical image with the wrong label can be effectively screened out from the medical images with the labels, and the accuracy of determining the noise data is improved.
The embodiment of the present application also provides a server and a terminal device for noise data identification of medical images, and the server and the terminal device for noise data identification of medical images provided by the embodiment of the present application will be described in terms of hardware implementation.
Referring to fig. 5, fig. 5 is a schematic diagram of a server 1400 according to an embodiment of the present application, where the server 1400 may have a relatively large difference due to different configurations or performances, and may include one or more Central Processing Units (CPUs) 1422 (e.g., one or more processors) and a memory 1432, one or more storage media 1430 (e.g., one or more mass storage devices) for storing applications 1442 or data 1444. Memory 1432 and storage media 1430, among other things, may be transient or persistent storage. The program stored on storage medium 1430 may include one or more modules (not shown), each of which may include a sequence of instructions operating on a server. Still further, a central processor 1422 may be disposed in communication with storage medium 1430 for executing a sequence of instruction operations in storage medium 1430 on server 1400.
The server 1400 may also include one or more power supplies 1426, one or more wired or wireless network interfaces 1450, one or more input-output interfaces 1458, and/or one or more operating systems 1441 such as Windows Server, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.
The steps performed by the server in the above embodiments may be based on the server structure shown in fig. 5.
The CPU 1422 is configured to perform the following steps:
determining an identification result of the tag data to be identified according to the first identification model; the first recognition model is obtained by performing N rounds of training on an initial network model, N is less than the total number M of rounds required for completing the complete training of the initial network model, and the model parameters of the first recognition model are not converged; the label data to be identified is a medical image with a label;
determining whether the tag data to be recognized is noise data or not according to the confidence embodied by the recognition result; the noise data is a medical image with a label with an error.
Optionally, the CPU 1422 may further perform the method steps of any specific implementation manner of the noise data identification method for medical images in the embodiment of the present application.
With respect to the above-described noise data identification method for medical images, the present application embodiment also provides a terminal device for noise data identification for medical images, so that the above-described noise data identification method for medical images is practically implemented and applied.
Referring to fig. 6, fig. 6 is a schematic structural diagram of a terminal device according to an embodiment of the present application. For convenience of explanation, only the portions related to the embodiments of the present application are shown, and specific technical details are not disclosed, please refer to the method portion of the embodiments of the present application. The terminal device can be any terminal device including a tablet computer, a Personal Digital Assistant (English full name: Personal Digital Assistant, English abbreviation: PDA) and the like:
fig. 6 is a block diagram illustrating a partial structure related to a terminal provided in an embodiment of the present application. Referring to fig. 6, the terminal includes: radio Frequency (RF) circuit 1510, memory 1520, input unit 1530, display unit 1540, sensor 1550, audio circuit 1560, wireless fidelity (WiFi) module 1570, processor 1580, and power 1590. Those skilled in the art will appreciate that the tablet configuration shown in fig. 6 is not intended to be limiting of tablets and may include more or fewer components than those shown, or some components may be combined, or a different arrangement of components.
The following describes each component of the tablet pc in detail with reference to fig. 6:
the memory 1520 may be used to store software programs and modules, and the processor 1580 implements various functional applications of the terminal and data processing by operating the software programs and modules stored in the memory 1520. The memory 1520 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function (such as a sound playing function, an image playing function, etc.), and the like; the storage data area may store data (such as audio data, a phonebook, etc.) created according to the use of the cellular phone, and the like. Further, the memory 1520 may include high-speed random access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device.
The processor 1580 is a control center of the terminal, connects various parts of the entire tablet pc using various interfaces and lines, and performs various functions of the tablet pc and processes data by operating or executing software programs and/or modules stored in the memory 1520 and calling data stored in the memory 1520. Optionally, the processor 1580 may include one or more processing units; preferably, the processor 1580 may integrate an application processor, which mainly handles operating systems, user interfaces, application programs, and the like, and a modem processor, which mainly handles wireless communications. It is to be appreciated that the modem processor may not be integrated into the processor 1580.
In the embodiment of the present application, the terminal includes a memory 1520 that can store the program code and transmit the program code to the processor.
The processor 1580 included in the terminal may execute the noise data identification method for medical images provided in the above embodiments according to the instructions in the program code.
The embodiment of the present application further provides a computer-readable storage medium for storing a computer program for executing the noise data identification method for medical images provided by the above embodiment.
Those of ordinary skill in the art will understand that: all or part of the steps for realizing the method embodiments can be completed by hardware related to program instructions, the program can be stored in a computer readable storage medium, and the program executes the steps comprising the method embodiments when executed; and the aforementioned storage medium may be at least one of the following media: various media that can store program codes, such as read-only memory (ROM), RAM, magnetic disk, or optical disk.
It should be noted that, in the present specification, each embodiment is described in a progressive manner, and the same and similar parts among the embodiments may be referred to each other, and each embodiment is described with emphasis on differences from other embodiments. In particular, for the apparatus and system embodiments, since they are substantially similar to the method embodiments, they are described in a relatively simple manner, and reference may be made to some of the descriptions of the method embodiments for related points. The above-described embodiments of the apparatus and system are merely illustrative, and the units illustrated as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e. may be located in one place, or may also be distributed over a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
The above description is only one specific embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present application should be covered within 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 (15)
1. A method for identifying noisy data of a medical image, the method comprising:
determining an identification result of the tag data to be identified according to the first identification model; the first recognition model is obtained by carrying out N rounds of training on an initial network model, N is less than the total number M of rounds required by complete training on the initial network model, and model parameters of the first recognition model are not converged; the data of the label to be recognized is a medical image with a label, N is the number of rounds used for initial training of the initial network model, and the initial network model is trained by using a large learning rate in the initial training process, so that the first recognition model learns the common characteristic of similar data in a training sample set, and the attention degree of the noise feature in the training sample set is reduced;
determining whether the tag data to be identified is noise data or not according to the confidence degree embodied by the identification result; and the noise data is the medical image with the label labeled wrongly.
2. The method according to claim 1, wherein the determining whether the tag data to be recognized corresponding to the recognition result is noise data according to the confidence level embodied by the recognition result comprises:
and if the confidence degree embodied by the identification result is lower than a first condition threshold value, determining the tag data to be identified corresponding to the identification result as noise data.
3. The method according to claim 1, wherein the determining whether the tag data to be recognized corresponding to the recognition result is noise data according to the confidence level embodied by the recognition result comprises:
if the confidence degree embodied by the identification result is higher than a second condition threshold value, determining whether the label of the label data to be identified is consistent with the identification result;
and if the identification result is inconsistent with the identification result, determining that the to-be-identified label data corresponding to the identification result is noise data.
4. The method according to any one of claims 1 to 3, wherein before determining whether the tag data to be identified corresponding to the identification result is noise data according to the confidence level embodied by the identification result, the method further comprises:
determining the confidence embodied by the recognition result by identifying the probability distribution of the recognition result.
5. The method of claim 4, wherein determining the confidence level corresponding to the recognition result by identifying a probability distribution of the recognition result comprises:
determining information entropy corresponding to the probability distribution for identifying the recognition result;
and determining the confidence degree embodied by the recognition result according to the information entropy.
6. The method of claim 1, wherein the label data to be identified is label data in a training sample set used for training the initial network model.
7. The method of claim 6, wherein after determining noise data from the recognition result, the method further comprises:
acquiring re-labeling data, wherein the re-labeling data is obtained by re-labeling the noise data with labels;
generating an adjustment sample set according to the re-labeling data and residual data in the training sample set, wherein the residual data are label data except noise data in the training sample set;
and taking the adjusting sample set as the training sample set to perform N rounds of training on the initial network model again to obtain a second recognition model, and performing noise data recognition on the medical image based on the second recognition model.
8. An apparatus for recognizing noise data of medical image, comprising a first determining unit and a second determining unit:
the first determining unit is used for determining the identification result of the tag data to be identified according to the first identification model; the first recognition model is obtained by carrying out N rounds of training on an initial network model, N is less than the total number M of rounds required by complete training on the initial network model, and model parameters of the first recognition model are not converged; the label data to be identified is a medical image with a label; the N is the number of rounds used for carrying out initial training on the initial network model, and the initial network model is trained by using a large learning rate in the initial training process, so that the first recognition model learns the common characteristic of similar data in a training sample set, and the attention degree of the noise features in the training sample set is reduced;
the second determining unit is configured to determine whether the tag data to be identified is noise data according to the confidence level embodied by the identification result; the noise data is a medical image with a label with an error.
9. The apparatus of claim 8, wherein the second determining unit is configured to:
and if the confidence degree embodied by the recognition result is lower than a first condition threshold value, determining the tag data to be recognized corresponding to the recognition result as noise data.
10. The apparatus of claim 8, wherein the second determining unit is configured to:
if the confidence degree embodied by the identification result is higher than a second condition threshold value, determining whether the label of the label data to be identified is consistent with the identification result;
and if the identification result is inconsistent with the identification result, determining that the to-be-identified label data corresponding to the identification result is noise data.
11. The apparatus according to any one of claims 8-10, wherein the second determining unit is further configured to:
and determining the confidence embodied by the recognition result by identifying the probability distribution of the recognition result.
12. The apparatus of claim 11, wherein the second determining unit is configured to:
determining information entropy corresponding to probability distribution for identifying the recognition result;
and determining the confidence degree embodied by the recognition result according to the information entropy.
13. The apparatus of claim 8, wherein the label data to be identified is label data in a training sample set used for training the initial network model.
14. A noisy data identification device for medical imaging, characterized in that said device comprises a processor and a memory:
the memory is used for storing program codes and transmitting the program codes to the processor;
the processor is configured to perform the method of any of claims 1-7 according to instructions in the program code.
15. A computer-readable storage medium, characterized in that the computer-readable storage medium is used to store a computer program for performing the method of any one of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010170514.7A CN111414946B (en) | 2020-03-12 | 2020-03-12 | Artificial intelligence-based medical image noise data identification method and related device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010170514.7A CN111414946B (en) | 2020-03-12 | 2020-03-12 | Artificial intelligence-based medical image noise data identification method and related device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111414946A CN111414946A (en) | 2020-07-14 |
| CN111414946B true CN111414946B (en) | 2022-09-23 |
Family
ID=71491108
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010170514.7A Active CN111414946B (en) | 2020-03-12 | 2020-03-12 | Artificial intelligence-based medical image noise data identification method and related device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111414946B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112801114B (en) * | 2021-01-20 | 2024-03-08 | 杭州依图医疗技术有限公司 | Method and device for determining projection position information of breast image |
| CN115147670A (en) * | 2021-03-15 | 2022-10-04 | 华为技术有限公司 | Object processing method and device |
| TWI777521B (en) * | 2021-04-26 | 2022-09-11 | 智齡科技股份有限公司 | Pressure injury identification and analysis system and method |
| JPWO2022259313A1 (en) * | 2021-06-07 | 2022-12-15 | ||
| CN113139628B (en) * | 2021-06-22 | 2021-09-17 | 腾讯科技（深圳）有限公司 | Sample image identification method, device and equipment and readable storage medium |
| CN113470031B (en) * | 2021-09-03 | 2021-12-03 | 北京字节跳动网络技术有限公司 | Polyp classification method, model training method and related device |
| CN114429581A (en) * | 2022-01-26 | 2022-05-03 | 北京百度网讯科技有限公司 | Data processing method and device, equipment, medium and product |
| CN116403074B (en) * | 2023-04-03 | 2024-05-14 | 上海锡鼎智能科技有限公司 | Semi-automatic image labeling method and device based on active labeling |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105095919A (en) * | 2015-09-08 | 2015-11-25 | 北京百度网讯科技有限公司 | Image recognition method and image recognition device |
| CN110427994A (en) * | 2019-07-24 | 2019-11-08 | 腾讯医疗健康(深圳)有限公司 | Digestive endoscope image processing method, device, storage medium, equipment and system |
| CN110458107A (en) * | 2019-08-13 | 2019-11-15 | 北京百度网讯科技有限公司 | Method and apparatus for image recognition |
| CN110473192A (en) * | 2019-04-10 | 2019-11-19 | 腾讯医疗健康(深圳)有限公司 | Digestive endoscope image recognition model training and recognition methods, apparatus and system |
| CN110689038A (en) * | 2019-06-25 | 2020-01-14 | 深圳市腾讯计算机系统有限公司 | Training method and device of neural network model and medical image processing system |
-
2020
- 2020-03-12 CN CN202010170514.7A patent/CN111414946B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105095919A (en) * | 2015-09-08 | 2015-11-25 | 北京百度网讯科技有限公司 | Image recognition method and image recognition device |
| CN110473192A (en) * | 2019-04-10 | 2019-11-19 | 腾讯医疗健康(深圳)有限公司 | Digestive endoscope image recognition model training and recognition methods, apparatus and system |
| CN110689038A (en) * | 2019-06-25 | 2020-01-14 | 深圳市腾讯计算机系统有限公司 | Training method and device of neural network model and medical image processing system |
| CN110427994A (en) * | 2019-07-24 | 2019-11-08 | 腾讯医疗健康(深圳)有限公司 | Digestive endoscope image processing method, device, storage medium, equipment and system |
| CN110458107A (en) * | 2019-08-13 | 2019-11-15 | 北京百度网讯科技有限公司 | Method and apparatus for image recognition |
Non-Patent Citations (1)
| Title |
|---|
| 《训练集噪声对于深度学习的影响》;seasermy;《CSDN》;20171009;https://blog.csdn.net/seasermy/article/details/78182328?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522165467049516781667858877%2522%252C%2522scm%2522%253A%252220140713.130102334..%2522%257D&request_id=165467049516781667858877&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~all~sobaiduend~default-1-78182328-null-null.142^v11^pc_search_result_control_group,157^v13^control&utm_term=%E8%AE%AD%E7%BB%83%E9%9B%86%E5%99%AA%E5%A3%B0&spm=1018.2226.3001.4187 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111414946A (en) | 2020-07-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111414946B (en) | Artificial intelligence-based medical image noise data identification method and related device | |
| CN109558832B (en) | Human body posture detection method, device, equipment and storage medium | |
| US20220036135A1 (en) | Method and apparatus for determining image to be labeled and model training method and apparatus | |
| CN110796199B (en) | Image processing method and device and electronic medical equipment | |
| CN112348117B (en) | Scene recognition method, device, computer equipment and storage medium | |
| CN112395979B (en) | Image-based health state identification method, device, equipment and storage medium | |
| WO2018196718A1 (en) | Image disambiguation method and device, storage medium, and electronic device | |
| JP2017062778A (en) | Method and device for classifying object of image, and corresponding computer program product and computer-readable medium | |
| CN111652181B (en) | Target tracking method and device and electronic equipment | |
| CN111401192B (en) | Model training method and related device based on artificial intelligence | |
| CN111783997B (en) | Data processing method, device and equipment | |
| CN110147806A (en) | Training method, device and the storage medium of image description model | |
| US20200089623A1 (en) | Data sharing system and data sharing method therefor | |
| CN115131604A (en) | Multi-label image classification method and device, electronic equipment and storage medium | |
| CN113128526B (en) | Image recognition method and device, electronic equipment and computer-readable storage medium | |
| CN111429414B (en) | Artificial intelligence-based focus image sample determination method and related device | |
| CN112990154B (en) | Data processing method, computer equipment and readable storage medium | |
| CN112995757B (en) | Video clipping method and device | |
| Wang et al. | Multistage model for robust face alignment using deep neural networks | |
| CN116569210A (en) | Normalizing OCT image data | |
| CN110008922A (en) | Image processing method, unit, medium for terminal device | |
| CN113239915B (en) | Classroom behavior identification method, device, equipment and storage medium | |
| CN113743448B (en) | Model training data acquisition method, model training method and device | |
| CN113076833B (en) | Training method of age identification model, face age identification method and related device | |
| CN118015386B (en) | Image recognition method and device, storage medium and electronic equipment |
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 | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 40025948 Country of ref document: HK |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |