CN111834004A - Unknown disease category identification method and device based on centralized space learning - Google Patents

Unknown disease category identification method and device based on centralized space learning Download PDF

Info

Publication number
CN111834004A
CN111834004A CN202010447732.0A CN202010447732A CN111834004A CN 111834004 A CN111834004 A CN 111834004A CN 202010447732 A CN202010447732 A CN 202010447732A CN 111834004 A CN111834004 A CN 111834004A
Authority
CN
China
Prior art keywords
unknown
class
space
known class
loss
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010447732.0A
Other languages
Chinese (zh)
Inventor
史业民
于重之
俞益州
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Shenrui Bolian Technology Co Ltd
Shenzhen Deepwise Bolian Technology Co Ltd
Original Assignee
Beijing Shenrui Bolian Technology Co Ltd
Shenzhen Deepwise Bolian Technology Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Shenrui Bolian Technology Co Ltd, Shenzhen Deepwise Bolian Technology Co Ltd filed Critical Beijing Shenrui Bolian Technology Co Ltd
Priority to CN202010447732.0A priority Critical patent/CN111834004A/en
Publication of CN111834004A publication Critical patent/CN111834004A/en
Pending legal-status Critical Current

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
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • 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
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides an unknown disease category identification method and device based on centralized space learning, wherein the method comprises the following steps: training an initial model through a known class sample, initializing a known class space, and mapping the known class sample to a hypersphere of a hidden space; training to generate an confrontation network; generating an unknown anchor point; generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the countermeasure network; taking the unknown image as an unknown class sample, combining the unknown class sample with the known class sample, and adjusting the known class space to obtain a trained model; acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each prototype point of the known class, determining that the new sample belongs to the class corresponding to the prototype point of the known class if the distance is less than a preset threshold, and determining that the new sample belongs to the unknown class if the distance is more than or equal to the preset threshold.

Description

Unknown disease category identification method and device based on centralized space learning
Technical Field
The invention relates to the field of computers, in particular to an unknown disease category identification method and device based on centralized space learning.
Background
With the rapid development of convolutional neural networks, CAD (computer aided diagnosis system) is rapidly developed, excellent performance is obtained, and wide clinical applications are developed. However, all existing CAD solutions assume that the predefined disease class includes all target classes and make predictions based on this set of predefined disease classes. This approach is associated with two risks: 1. for diseases not included in the training set (such as stubborn miscellaneous diseases or new disease categories), directly identifying the diseases as predefined categories can cause misdiagnosis and even delay treatment, and can cause unacceptable consequences; 2. due to the huge workload of the imaging department, doctors may input images with low quality and wrong part labeling into the CAD system, thereby misleading the system and causing the CAD system to give wrong diagnosis. In general, the CAD system based on the closed set (i.e., the predefined target class set) assumption has extremely low robustness when facing unknown classes or wrong data, and needs to research the CAD system with the unknown class identification capability to detect the unknown classes on the premise of ensuring the accuracy of the known class identification.
The existing solution mainly depends on a closed set recognition mode, namely, an input image is preprocessed and then input into a 2D/3D convolutional neural network, and then classification is realized through an activation function such as softmax. Thus, the detection capability of the unknown class is lacking.
Disclosure of Invention
The present invention aims to provide a method and apparatus for identification of unknown disease classes based on centralised spatial learning that overcomes or at least partially solves the above mentioned problems.
In order to achieve the purpose, the technical scheme of the invention is realized as follows:
one aspect of the present invention provides an unknown disease category identification method based on centralized space learning, including: training an initial model through a known class sample, initializing a known class space, and mapping the known class sample to a hypersphere of a hidden space, wherein the known sample is distributed on the hypersphere around an original point of a corresponding class; training and generating a confrontation network through known class samples and corresponding characteristics thereof; extracting the characteristics of a known image based on an initial model, and carrying out random disturbance to generate an unknown anchor point; generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the countermeasure network; taking the unknown image as an unknown class sample, combining the unknown class sample with the known class sample, and adjusting the known class space to obtain a trained model; acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each prototype point of the known class, determining that the new sample belongs to the class corresponding to the prototype point of the known class if the distance is less than a preset threshold, and determining that the new sample belongs to the unknown class if the distance is more than or equal to the preset threshold.
In the process of training the initial model through the known class samples, the separation loss is also utilized, and the known class space is learned based on the cross entropy loss and the prototype loss of the distance, so that the initial model is optimized.
Wherein training to generate the countermeasure network by the classified samples and their corresponding features comprises: training a generator and a discriminator to learn the mapping from the hidden space to the image domain; a countermeasure network is generated using a countermeasure loss, cyclic continuity loss optimization generator, and a discriminator.
Wherein, in the process of adjusting the known class space, the known class space is also adjusted by utilizing clustering loss, DCE (distributed component analysis) transient, prototype loss and separation loss.
In another aspect, the present invention provides an unknown disease category identification apparatus based on centralized space learning, including: the initialization module is used for training an initial model through known class samples, initializing the known class space and mapping the known class samples to a hypersphere of a hidden space, wherein the known samples are distributed on the hypersphere around prototype points of corresponding classes of the known samples; the generation module is used for training and generating a confrontation network through the known class samples and the corresponding characteristics thereof; extracting the characteristics of a known image based on an initial model, and carrying out random disturbance to generate an unknown anchor point; generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the countermeasure network; the adjusting module is used for adjusting the known class space by taking the unknown image as an unknown class sample and combining the unknown class sample with the known class sample to obtain a trained model; and the judging module is used for acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each known class prototype point, determining that the new sample belongs to the class corresponding to the known class prototype point if the distance is less than a preset threshold value, and determining that the new sample belongs to an unknown class if the distance is greater than or equal to the preset threshold value.
The initialization module further learns the known class space by using separation loss, distance-based cross entropy loss and prototype loss in the process of training the initial model through the known class samples, and optimizes the initial model.
The generation module trains and generates the confrontation network through the classified samples and the corresponding characteristics thereof in the following way: the generation module is specifically used for training the generator and the discriminator and learning the mapping from the hidden space to the image domain; a countermeasure network is generated using a countermeasure loss, cyclic continuity loss optimization generator, and a discriminator.
The adjusting module adjusts the known class space by using clustering loss, DCE (distributed component analysis) transient loss, prototype loss and separation loss in the process of adjusting the known class space.
Therefore, the unknown disease category identification method and device based on centralized space learning provided by the invention have the advantages that aiming at the problem that the existing solution assumes that all disease categories are known, the method for dividing the diseases into the known categories and the unknown categories and the method for centralized space learning are provided, and whether the samples belong to the known categories or not is automatically detected on the premise that the unknown category information is not introduced at all. According to the method, on the basis of closed set identification, a spherical hyperplane space is introduced, then three steps of known space initialization, unknown anchor point generation and centralized space adjustment are utilized, the known classes are adjusted to the edge of a hypersphere, each class is gathered around the prototype point of the class, and the unknown classes are gathered at the center of the hypersphere space, so that the known classes and the unknown classes are separated.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
FIG. 1 is a flowchart of an unknown disease category identification method based on centralized space learning according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of an unknown disease category identification apparatus based on centralized space learning according to an embodiment of the present invention.
Detailed Description
Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
The core of the invention is that: a centralized hidden space is provided, which is composed of a central point of a central area and prototype points around the central point. Samples of known classes are clustered around the prototype point of each class, while samples of unknown classes are clustered around the central point. Samples of known and unknown classes can be distinguished based on their distance from the nearest prototype point. Specifically, when the distance is smaller than the threshold, the sample belongs to the category of the corresponding prototype point; otherwise, the classification is considered to belong to an unknown class.
Therefore, the invention provides that all classes are divided into known classes and unknown classes, and then the boundary between the known class and the unknown class is learned through hypersphere, so that the detection of the unknown class is realized, and the unknown class is prevented from being identified as the known class. Whereas for known classes, disease identification can be achieved.
In particular, the present invention may comprise at least three steps: known category space initialization, unknown anchor point generation and centralized space fine adjustment are carried out, so that a trained model can be obtained.
Fig. 1 is a flowchart illustrating an unknown disease category identification method based on centralized space learning according to an embodiment of the present invention, and referring to fig. 1, the unknown disease category identification method based on centralized space learning according to an embodiment of the present invention includes:
s1, training an initial model through the known class samples, initializing the known class space, and mapping the known class samples to the hypersphere of the hidden space, wherein the known samples are distributed on the hypersphere around the prototype points of the corresponding classes.
The method comprises the steps of initializing a known class space, training an initial model through a known class sample, and initializing the known class space, so that the known sample is distributed on a hypersphere around a prototype point of a corresponding class (the prototype point can be used as a clustering center of a current class).
This initial model/space is used to: 1. the method is used as a pre-training model for subsequent centralized space fine tuning, so that the identification accuracy of the known classes is ensured; 2. the characteristic extraction module is used for learning the mapping from the characteristics to the image domain in the process of 'generating confrontation network training'; 3. and as a feature extraction module, generating a target anchor point in the hidden space anchor point generation process through random disturbance on features.
In particular, for convenience, the invention may be defined using the following notation. The known class set of N samples of the known K classes is
Figure BDA0002506543120000041
Hidden space of d dimension
Figure BDA0002506543120000042
By XiE.k denotes XiBelonging to class k. Is defined in
Figure BDA0002506543120000043
In (1), a set of K learnable prototypes
Figure BDA0002506543120000044
Each element in M corresponds to a known class. Defining a space based on the value of M
Figure BDA0002506543120000045
Has a center point c of
Figure BDA0002506543120000046
Each time M is updated, c will change accordingly.
Known class space initialization: in the initialization of the known space, a discriminant embedded model learns a mapping, a known class sample is mapped to a hypersphere in a hidden space, different known classes are separated to the maximum extent, and simultaneously prototypes corresponding to the classes are far away from the center as far as possible, so that a larger inter-class distance is obtained. Since the distance between the sample and the prototype is a standard for determining whether the sample belongs to a known class in the open set recognition, in this step, a discriminant embedding model is required to reduce the intra-class variance and increase the inter-class distance. After initialization, the space around c will be left blank, leaving space for subsequent unknown classes.
As an optional implementation manner of the embodiment of the present invention, in the process of training the initial model through the known class samples, the separation loss, the cross entropy loss based on distance and the prototype loss are also used to learn the known class space, and the initial model is optimized.
Specifically, given a trainable discriminant embedding model (forward propagation using (.) notation for simplicity) and a prototype M, the objective function in a given spatial initialization is:
Figure BDA0002506543120000047
wherein κ (,) represents the distance between them, ρ1And ρ2Are weights used to balance the quantities.
The present invention proposes to use separation loss, distance-based cross-entropy (DCE) loss and prototype loss to learn the initial known space, optimizing the above objective function.
Using DCE loss, known class samples are pushed away from prototypes that do not belong to the class, while being clustered near the corresponding class prototypes. Thereby, the inter-class distance can be increased at the same time and the intra-class variance can be slightly reduced. Known as XiE k, DCE loss is defined as:
Figure BDA0002506543120000048
Figure BDA0002506543120000051
wherein the content of the first and second substances,
Figure BDA0002506543120000052
is XiAnd mkIn hidden spaces2The square of the distance, γ, is a hyper-parameter that controls the classification strength.
Prototype loss is used to efficiently reduce intra-class variance, more densely clustering samples of the same class around the corresponding prototype. The loss is defined as:
Figure BDA0002506543120000053
the separation loss limits the prototype to a hypersphere and increases the radius of the hypersphere as much as possible. It is specifically defined as follows:
Figure BDA0002506543120000054
the separation penalty at each back-propagation update and M pushes the prototype closest to the sphere center c further away. Helping the algorithm learn that all known class samples are around the center of sphere c and that few known class samples fall around c. When a prototype falls near c, the segmentation loss continuously pushes it away from c. Until all prototypes are substantially the same distance to c, after which
Figure BDA0002506543120000055
Each prototype will be pushed in turn, gradually enlarging the radius.
In summary, it is known that in the spatial initialization, the total loss function is:
Figure BDA0002506543120000056
wherein alpha is1And beta1Is a hyperparameter used to balance the individual loss weights.
And S2, training and generating the confrontation network by knowing the class sample and the corresponding characteristics thereof.
The method comprises a step of generating a countermeasure network training step, wherein the countermeasure network is trained and generated through known class samples and corresponding characteristics thereof, and mapping from a characteristic domain to an image domain is realized, namely, an original image is generated from the characteristics by learning. This model is used for "anchor point image generation", generating a new image from a selected anchor point as an unknown class.
And S3, extracting the characteristics of the known image based on the initial model, and carrying out random disturbance to generate an unknown anchor point.
The method comprises the steps of generating an anchor point in a hidden space, extracting the characteristics of a known image based on an initial model, and carrying out random disturbance, so that the characteristics which do not exist originally are generated, namely the unknown anchor point. The output is used to: in "anchor point image generation", a new image is generated using a generation countermeasure network as an input feature.
And S4, generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the confrontation network.
The step is an anchor point image generation step, and based on the generated anchor points and the generated countermeasure network, an unknown image corresponding to the anchor points is generated and used as an unknown class in the centralized space fine adjustment to assist in adjusting the feature space.
Specifically, the above steps S2-S4 may be combined into one step, namely, the unknown anchor point generating step.
In the present invention, by generating a proxy image, the phenomenon that all samples tend to be embedded all at a position close to M in the hidden space is mitigated. Proxy images are images that are generated, do not belong to any known class, and are also embedded near M.
As an optional implementation manner of the embodiment of the present invention, training to generate the countermeasure network by using the already classified samples and their corresponding features includes: training a generator and a discriminator to learn the mapping from the hidden space to the image domain; a countermeasure network is generated using a countermeasure loss, cyclic continuity loss optimization generator, and a discriminator.
Specifically, in the present invention, a proxy image and corresponding unknown anchor point are generated using GAN. It is known that the distribution is X-p pre-trained in the known spatial initializationdata(X) known class sample
Figure BDA0002506543120000061
In this step, the CSL training generator
Figure BDA0002506543120000062
Sum discriminator
Figure BDA0002506543120000063
Learning from hidden space
Figure BDA0002506543120000064
To the image domain
Figure BDA0002506543120000065
To (3) is performed. At the same time, for
Figure BDA0002506543120000066
Is inputted (X)i) Output of
Figure BDA0002506543120000067
Should be close to (X)i). (for simplicity, use is made of
Figure BDA0002506543120000068
And
Figure BDA0002506543120000069
indicates its forward directionAnd (5) spreading. )
Thus, the present invention introduces, in addition to the opposing losses, a loss of cyclic continuity, which is defined as follows:
Figure BDA00025065431200000610
Figure BDA00025065431200000611
and
Figure BDA00025065431200000612
the challenge loss of (a) is:
Figure BDA00025065431200000613
Figure BDA00025065431200000614
therefore, the temperature of the molten metal is controlled,
Figure BDA00025065431200000615
and
Figure BDA00025065431200000616
the loss equation of (a) is:
Figure BDA00025065431200000617
Figure BDA00025065431200000618
where λ is a hyperparameter that balances the two losses, the parameters in sum M are fixed and not updated.
Thereafter, unknown anchor points are generated
Figure BDA00025065431200000619
And its corresponding proxy image
Figure BDA00025065431200000620
In the course of this process, the temperature of the molten steel is controlled,
Figure BDA00025065431200000621
and
Figure BDA00025065431200000622
the parameters in (1) are all fixed and cannot be updated. First, an initial anchor point f is generatedinit. For each generated finitAll have a probability PfSo that
Figure BDA00025065431200000623
finitAnd PfIs defined as:
finit＝(Xi)+#
Figure BDA00025065431200000624
wherein the content of the first and second substances,
Figure BDA00025065431200000625
is one satisfying | | | | non-calculation2∈(0,0.2]R is a threshold for screening anchor points, in practice r is set to be greater than 70% of the samples to l of the nearest prototype2Distance.
Since the distance of a sample to a prototype is inversely proportional to the probability of the sample belonging to the prototype, f is used in the unknown anchor point generationinitAnd the distance of M as a judgment
Figure BDA0002506543120000071
Whether it belongs to an unknown category and can be used as an index for proxy images. f. ofinitThe closer the distance M is, the greater the probability that it belongs to a known class, finitThere is less chance of being chosen as an unknown anchor point
Figure BDA0002506543120000072
In the unknown anchor point generation process, finitIs continuously generated until
Figure BDA0002506543120000073
The number of middle proxy images is the same as the number of training images.
And S5, combining the unknown image as an unknown class sample with the known class sample, and adjusting the known class space to obtain the trained model.
The step is a centralized space fine adjustment step, based on the generated unknown image, the unknown image is taken as an unknown class, and a characteristic space is adjusted by combining with a sample of a known class, so that the unknown class is adjusted to a space center, and the known class is regularly distributed on the edge of the hypersphere around a prototype point.
As an optional implementation manner of the embodiment of the present invention, in the process of adjusting the known class space, the known class space is also adjusted by using the clustering loss, the DCE transient, the prototype loss, and the separation loss.
Specifically, in the centralized spatial adjustment, through training, the implicit space is adjusted, the unknown anchor points are clustered around c, and the known and unknown samples are separated. The objective function is:
Figure BDA0002506543120000074
wherein ξ1,ξ2,ξ3Is a hyper-parameter that balances the individual losses.
We propose clustering loss to better push the unknown class samples to c, defined as:
Figure BDA0002506543120000075
to ensure that its classification performance on known class samples is preserved, while better distinguishing between known and unknown class samples, DCE loss, prototype loss and separation loss are also introduced into the centralized spatial adaptation. The total loss function is defined as:
Figure BDA0002506543120000076
wherein alpha is2,β2And θ is a hyper-parameter used to balance the individual losses.
And S6, acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each prototype point of the known class, determining that the new sample belongs to the class corresponding to the prototype point of the known class if the distance is less than a preset threshold, and determining that the new sample belongs to the unknown class if the distance is greater than or equal to the preset threshold.
The method comprises the steps of judging unknown types, extracting the characteristics of a new sample based on a trained model, then calculating the distance from the new sample to each prototype point of the known type, wherein if the distance is smaller than a threshold value, the sample belongs to the type corresponding to the prototype point, and if the distance is larger than or equal to the threshold value, the sample is the unknown type.
Therefore, by using the unknown disease category identification method based on centralized space learning provided by the embodiment of the invention, aiming at the problem that the existing solution assumes that all disease categories are known, the method for dividing the diseases into the known categories and the unknown categories and the method for centralized space learning are provided, and whether the samples belong to the known categories or not is automatically detected on the premise that the unknown category information is not introduced at all. On the basis of closed set identification, a spherical hyperplane space is introduced, then three steps of known space initialization, unknown anchor point generation and centralized space adjustment are utilized, the known classes are adjusted to the edge of a hypersphere, each class is gathered around a prototype point of each class, and the unknown classes are gathered at the center of the hypersphere space, so that the known classes and the unknown classes are separated.
In the invention, an open set auxiliary diagnosis concept is provided, and on the premise of ensuring disease diagnosis, diseases which do not belong to the known class of the training set are judged to prevent misdiagnosis caused by the unknown class; providing hypersphere hidden space distribution to adapt to the double-task requirements of disease classification and unknown class detection; the centralized space learning method is provided for realizing open set auxiliary diagnosis, and the centralized space construction is realized through three steps of known category space initialization, unknown anchor point generation and centralized space fine adjustment. The method supports the identification of known and unknown classes in a real environment, so that the robustness of an auxiliary diagnosis system under complex conditions is improved; the calculation efficiency is extremely high, few extra calculations are introduced into the reasoning part of the original auxiliary diagnosis system, and all other modules only influence the complexity of the training process; the auxiliary diagnosis of diseases and the identification of unknown classes can be realized simultaneously, and the diagnosis process is not changed.
In addition, as an optional embodiment of the present invention, only step S1 may be performed, an initial model is trained through a known class sample, a known class space is initialized, and the known class sample is mapped onto a hypersphere of a hidden space, where the known sample is distributed on the hypersphere around a prototype point of its corresponding class, that is, after a training classifier is initialized through the known class space, whether the known class is an unknown class can be determined according to a distance and a threshold.
Fig. 2 is a schematic structural diagram of an unknown disease category identification device based on centralized space learning according to an embodiment of the present invention, in which the above method is applied to the unknown disease category identification device based on centralized space learning, and the following is a brief description of the structure of the unknown disease category identification device based on centralized space learning, and for other things that are not the least, please refer to the related description in the above unknown disease category identification method based on centralized space learning, and referring to fig. 2, the unknown disease category identification device based on centralized space learning according to the present invention includes:
the initialization module is used for training an initial model through known class samples, initializing the known class space and mapping the known class samples to a hypersphere of a hidden space, wherein the known samples are distributed on the hypersphere around prototype points of corresponding classes of the known samples;
the generation module is used for training and generating a confrontation network through the known class samples and the corresponding characteristics thereof; extracting the characteristics of a known image based on an initial model, and carrying out random disturbance to generate an unknown anchor point; generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the countermeasure network;
the adjusting module is used for adjusting the known class space by taking the unknown image as an unknown class sample and combining the unknown class sample with the known class sample to obtain a trained model;
and the judging module is used for acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each known class prototype point, determining that the new sample belongs to the class corresponding to the known class prototype point if the distance is less than a preset threshold value, and determining that the new sample belongs to an unknown class if the distance is greater than or equal to the preset threshold value.
As an optional implementation manner of the embodiment of the present invention, in the process of training the initial model through the known class samples, the initialization module further learns the known class space by using the separation loss, the cross entropy loss based on the distance, and the prototype loss, and optimizes the initial model.
As an optional implementation manner of the embodiment of the present invention, the generation module trains and generates the countermeasure network by using the already classified samples and their corresponding features as follows: the generation module is specifically used for training the generator and the discriminator and learning the mapping from the hidden space to the image domain; a countermeasure network is generated using a countermeasure loss, cyclic continuity loss optimization generator, and a discriminator.
As an optional implementation manner of the embodiment of the present invention, in the process of adjusting the known class space, the adjusting module further adjusts the known class space by using the clustering loss, the DCE transient, the prototype loss, and the separation loss.
Therefore, by using the unknown disease category identification device based on centralized space learning provided by the embodiment of the invention, aiming at the problem that the existing solution assumes that all disease categories are known, the method for dividing the diseases into the known categories and the unknown categories and the centralized space learning are provided, and whether the samples belong to the known categories or not is automatically detected on the premise that the unknown category information is not introduced at all. On the basis of closed set identification, a spherical hyperplane space is introduced, then three steps of known space initialization, unknown anchor point generation and centralized space adjustment are utilized, the known classes are adjusted to the edge of a hypersphere, each class is gathered around a prototype point of each class, and the unknown classes are gathered at the center of the hypersphere space, so that the known classes and the unknown classes are separated.
In the invention, an open set auxiliary diagnosis concept is provided, and on the premise of ensuring disease diagnosis, diseases which do not belong to the known class of the training set are judged to prevent misdiagnosis caused by the unknown class; providing hypersphere hidden space distribution to adapt to the double-task requirements of disease classification and unknown class detection; the centralized space learning method is provided for realizing open set auxiliary diagnosis, and the centralized space construction is realized through three steps of known category space initialization, unknown anchor point generation and centralized space fine adjustment. The method supports the identification of known and unknown classes in a real environment, so that the robustness of an auxiliary diagnosis system under complex conditions is improved; the calculation efficiency is extremely high, few extra calculations are introduced into the reasoning part of the original auxiliary diagnosis system, and all other modules only influence the complexity of the training process; the auxiliary diagnosis of diseases and the identification of unknown classes can be realized simultaneously, and the diagnosis process is not changed.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.
Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.
The above are merely examples of the present application and are not intended to limit the present application. Various modifications and changes may occur to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims (8)

1. An unknown disease category identification method based on centralized space learning is characterized by comprising the following steps:
training an initial model through known class samples, initializing the known class space, and mapping the known class samples to a hypersphere of a hidden space, wherein the known samples are distributed on the hypersphere around prototype points of corresponding classes of the known samples;
training and generating a confrontation network through the known class samples and the corresponding characteristics thereof;
extracting the characteristics of the known image based on the initial model, and carrying out random disturbance to generate an unknown anchor point;
generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the countermeasure network;
taking the unknown image as an unknown class sample and combining the unknown class sample, and adjusting the known class space to obtain a trained model;
acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each prototype point of the known class, if the distance is smaller than a preset threshold value, determining that the new sample belongs to the class corresponding to the prototype point of the known class, and if the distance is larger than or equal to the preset threshold value, determining that the new sample belongs to the unknown class.
2. The method of claim 1, wherein in the training of the initial model by the known class samples, the initial model is optimized by learning the known class space further using separation loss, distance-based cross entropy loss, and prototype loss.
3. The method of claim 1, wherein training a generation of a countermeasure network by the already classified samples and their corresponding features comprises:
a training generator and a discriminator for learning the mapping from the hidden space to the image domain;
optimizing the generator and the arbiter to generate a challenge network using challenge loss, cyclic continuity loss.
4. The method of claim 1, wherein the adjusting the known class space further utilizes clustering loss, DCE transients, prototype loss, and separation loss.
5. An unknown disease category identification device based on centralized space learning, comprising:
the initialization module is used for training an initial model through known class samples, initializing the known class space and mapping the known class samples to a hypersphere of a hidden space, wherein the known samples are distributed on the hypersphere around prototype points of corresponding classes of the known samples;
the generation module is used for training and generating a confrontation network through the known class samples and the corresponding characteristics thereof; extracting the characteristics of the known image based on the initial model, and carrying out random disturbance to generate an unknown anchor point; generating an unknown image corresponding to the unknown anchor point based on the unknown anchor point and the countermeasure network;
the adjusting module is used for combining the unknown image serving as an unknown class sample with the known class sample, and adjusting the known class space to obtain a trained model;
and the judging module is used for acquiring a new sample, extracting the characteristics of the new sample based on the trained model, calculating the distance from the characteristics of the new sample to each prototype point of the known class, determining that the new sample belongs to the class corresponding to the prototype point of the known class if the distance is less than a preset threshold value, and determining that the new sample belongs to the unknown class if the distance is greater than or equal to the preset threshold value.
6. The apparatus of claim 5, wherein the initialization module further optimizes the initial model by learning the known class space using separation loss, distance-based cross entropy loss, and prototype loss during the training of the initial model by the known class samples.
7. The apparatus of claim 5, wherein the generation module trains the generation countermeasure network by the already classified samples and their corresponding features as follows: the generation module is specifically used for training a generator and a discriminator to learn the mapping from the hidden space to the image domain; optimizing the generator and the arbiter to generate a challenge network using challenge loss, cyclic continuity loss.
8. The apparatus of claim 5, wherein the adjustment module further adjusts the known class space using cluster loss, DCE transients, prototype loss, and separation loss in the adjusting the known class space.
CN202010447732.0A 2020-05-25 2020-05-25 Unknown disease category identification method and device based on centralized space learning Pending CN111834004A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010447732.0A CN111834004A (en) 2020-05-25 2020-05-25 Unknown disease category identification method and device based on centralized space learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010447732.0A CN111834004A (en) 2020-05-25 2020-05-25 Unknown disease category identification method and device based on centralized space learning

Publications (1)

Publication Number Publication Date
CN111834004A true CN111834004A (en) 2020-10-27

Family

ID=72913579

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010447732.0A Pending CN111834004A (en) 2020-05-25 2020-05-25 Unknown disease category identification method and device based on centralized space learning

Country Status (1)

Country Link
CN (1) CN111834004A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112508062A (en) * 2020-11-20 2021-03-16 普联国际有限公司 Open set data classification method, device, equipment and storage medium
CN112990300A (en) * 2021-03-11 2021-06-18 北京深睿博联科技有限责任公司 Foreground identification method, device, equipment and computer readable storage medium
CN114067429A (en) * 2021-11-02 2022-02-18 北京邮电大学 Action recognition processing method, device and equipment
CN114330597A (en) * 2022-01-14 2022-04-12 阿里巴巴达摩院(杭州)科技有限公司 User clustering method, data clustering method, device and electronic equipment

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108288506A (en) * 2018-01-23 2018-07-17 雨声智能科技（上海）有限公司 A kind of cancer pathology aided diagnosis method based on artificial intelligence technology
CN109658385A (en) * 2018-11-23 2019-04-19 上海鹰瞳医疗科技有限公司 Eye fundus image judgment method and equipment
CN110750665A (en) * 2019-10-12 2020-02-04 南京邮电大学 Open set domain adaptation method and system based on entropy minimization
CN110993094A (en) * 2019-11-19 2020-04-10 中国科学院深圳先进技术研究院 Intelligent auxiliary diagnosis method and terminal based on medical images

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108288506A (en) * 2018-01-23 2018-07-17 雨声智能科技（上海）有限公司 A kind of cancer pathology aided diagnosis method based on artificial intelligence technology
CN109658385A (en) * 2018-11-23 2019-04-19 上海鹰瞳医疗科技有限公司 Eye fundus image judgment method and equipment
CN110750665A (en) * 2019-10-12 2020-02-04 南京邮电大学 Open set domain adaptation method and system based on entropy minimization
CN110993094A (en) * 2019-11-19 2020-04-10 中国科学院深圳先进技术研究院 Intelligent auxiliary diagnosis method and terminal based on medical images

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112508062A (en) * 2020-11-20 2021-03-16 普联国际有限公司 Open set data classification method, device, equipment and storage medium
CN112990300A (en) * 2021-03-11 2021-06-18 北京深睿博联科技有限责任公司 Foreground identification method, device, equipment and computer readable storage medium
CN114067429A (en) * 2021-11-02 2022-02-18 北京邮电大学 Action recognition processing method, device and equipment
CN114067429B (en) * 2021-11-02 2023-08-29 北京邮电大学 Action recognition processing method, device and equipment
CN114330597A (en) * 2022-01-14 2022-04-12 阿里巴巴达摩院(杭州)科技有限公司 User clustering method, data clustering method, device and electronic equipment

Similar Documents

Publication Publication Date Title
CN111834004A (en) Unknown disease category identification method and device based on centralized space learning
CN107209873B (en) Hyper-parameter selection for deep convolutional networks
He et al. A two-stage genetic algorithm for automatic clustering
Zhou et al. Region growing for image segmentation using an extended PCNN model
CN111898617A (en) Target detection method and system based on attention mechanism and parallel void convolution network
Lim et al. Efficient-prototypicalnet with self knowledge distillation for few-shot learning
CN113569895A (en) Image processing model training method, processing method, device, equipment and medium
Sun et al. Optimized light-weight convolutional neural networks for histopathologic cancer detection
CN115063664A (en) Model learning method, training method and system for industrial vision detection
Hasan et al. Multistrategy self‐organizing map learning for classification problems
CN111832435A (en) Beauty prediction method and device based on migration and weak supervision and storage medium
He et al. ECS-SC: Long-tailed classification via data augmentation based on easily confused sample selection and combination
He et al. From macro to micro: rethinking multi-scale pedestrian detection
CN117523218A (en) Label generation, training of image classification model and image classification method and device
Yu et al. Efficient multi-spike learning with tempotron-like ltp and psd-like ltd
CN114693997A (en) Image description generation method, device, equipment and medium based on transfer learning
CN115775322A (en) Image classification method and device, computer equipment and storage medium
Shetty et al. Comparative analysis of different classification techniques
Lee et al. Improved method for learning data imbalance in gender classification model using DA-FSL
Arefinia et al. Using support vector machine (SVM) in modeling water resources systems
Xia et al. DEMST-KNN: A novel classification framework to solve imbalanced multi-class Problem
CN114283350B (en) Visual model training and video processing method, device, equipment and storage medium
US20240127075A1 (en) Synthetic dataset generator
Fu et al. Adaptive Open Set Recognition with Multi-modal Joint Metric Learning
CN116543302A (en) Parallel DCNN rapid scene matching recognition method based on Spark and feature clustering

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
CB03 Change of inventor or designer information
CB03 Change of inventor or designer information

Inventor after: Shi Yemin

Inventor after: Yu Zhongzhi

Inventor after: Yu Yizhou

Inventor before: Shi Yemin

Inventor before: Yu Zhongzhi

Inventor before: Yu Yizhou