WO2023054768A1

WO2023054768A1 - Deep learning model system for detecting pressure ulcer disease and determining stage of pressure ulcer disease, generation method therefor, and method for diagnosing pressure ulcer by using same

Info

Publication number: WO2023054768A1
Application number: PCT/KR2021/013466
Authority: WO
Inventors: 전병흡; 김건주; 홍성표
Original assignee: 주식회사 피플앤드테크놀러지
Priority date: 2021-09-30
Filing date: 2021-09-30
Publication date: 2023-04-06
Also published as: KR20240032002A; KR20230046720A

Abstract

Provided is an image-based method for generating a pressure ulcer disease determination model, the method comprising the steps of: (a) receiving, as an input, a photograph of a suspected pressure ulcer patient; (b) detecting a suspected affected area from the input photograph, and then labeling the photograph according to stage of pressure ulcer progression; (c) inputting the suspected affected area of the photograph labeled according to the stage into a CNN-binary discriminator, and outputting a predicted probability as a logit value; (d) receiving, as an input, a 4D vector consisting of the logit value to construct a DNN that determines the stage of pressure ulcer progression; and (e) learning the DNN by repeatedly inputting the 4D vector.

Description

Deep learning model system for detecting pressure ulcer disease and determining the stage, method for generating, and method for diagnosing pressure ulcer using the same

The present invention relates to a deep learning model system for capturing pressure ulcer diseases and determining stages, a generation method, and a method for diagnosing pressure ulcers using the same. It is about a deep learning model system, generation method, and method for diagnosing pressure ulcers using the deep learning model system for detecting pressure ulcer disease and determining the stage to receive appropriate medical services.

Artificial intelligence (AI) mimics the human brain and neural network of neurons, one day allowing computers and robots to think and act like humans.

For example, we can very easily tell dogs and cats apart just by looking at pictures, but computers can't.

To this end, a “Machine Learning (ML)” technique was devised. This technique inputs a lot of data into a computer and classifies similar ones. It's for the computer to classify.

Depending on how to classify data, many machine learning algorithms such as decision trees, Bayesian networks, support vector machines (SVMs), and artificial neural networks have emerged. did.

Among them, Deep Learning (DL) derived from artificial neural network algorithms is a technique used to cluster or classify data using artificial neural networks.

In machine learning and cognitive science, artificial neural networks are statistical learning algorithms inspired by neural networks in biology (the central nervous system of animals).

An artificial neural network refers to an overall model that has problem-solving ability by changing synapse coupling strength through learning in which artificial neurons that form a network by synapse coupling.

The core of deep learning using artificial neural networks is prediction through classification.

Computers divide data just as humans classify objects by discovering patterns in countless data.

This classification method is based on supervised (supervisor/teacher) learning that is optimized for the problem by inputting a signal (correct answer) from the leader (supervisor/teacher) and unsupervised (supervisor/teacher) learning that does not require a teacher signal from the leader. there is.

It is usually represented as an interconnection of neuron systems that computes values from inputs and is adaptable, allowing machine learning such as pattern recognition to be performed.

Like other machine learning methods that learn from data, neural networks are used to solve a wide range of problems, such as computer vision or speech recognition, that are generally difficult to solve with rule-based programming.

Such deep learning is used as an indicator for determining the behavioral state of the human body, and is used in various fields such as sports, medicine, health care, education, and art.

On the other hand, in the case of specific skin diseases such as pressure ulcers, they have different aspects depending on the stage of progression and are not connected. In the case of simply using a classifier, etc., it is difficult to obtain the required accuracy in the medical and healthcare field.In particular, deep learning that more accurately captures and classifies the advanced stage of pressure ulcer disease for the affected area of a patient with suspected pressure ulcer than before It is necessary to develop the model and its learning method.

An object of the present invention is to provide an accurate pressure ulcer stage discrimination model using a step-by-step probability model for pressure sore affected areas and a diagnosis method using the same.

In order to solve the above problems, the present invention comprises the steps of (a) receiving a picture of a suspected pressure sore patient; (b) capturing a suspected affected area from an input photograph, and then labeling the photograph according to stages of pressure ulcer progression; (c) inputting the suspected affected area of the photograph labeled in each step into a CNN-binary discriminator and outputting the predicted probability as a logit value; (d) constructing a DNN that receives the 4-dimensional vector consisting of the logit values and determines the stage of pressure ulcer progression; and (e) learning the DNN by repeatedly inputting the 4-dimensional vector; Provides a method for generating an image-based pressure sore disease discrimination model comprising a.

In one embodiment of the present invention, in the step (b), the suspected affected area is captured through a deep learning object capturer, and in the step (c), the CNN-binary discriminator has four, and the CNN-binary The discriminator learns four decision boundaries: negative - 1st level or higher, 1st lower level - 2nd higher level, 2nd lower level - 3rd level or higher, and 3rd lower level - 4th level respectively.

In one embodiment of the present invention, the CNN-binary discriminator learns using binary cross-entropy for the discrimination result for the step, and the DNN applies an excitation softmax function to a 5-dimensional vector. to determine the pressure sore stage, and the depth and width of the DNN hidden layer are repeatedly learned to determine optimal values.

The present invention also provides a deep learning system for detecting pressure ulcer diseases and determining stages, generated by the above-described image-based pressure ulcer disease discrimination model generation method.

The present invention also includes the steps of inputting a patient picture to the deep learning system described above; Receiving information about the stage of progression of pressure sores determined by the deep learning system from the photo; and transmitting the received information to a user terminal.

According to the present invention, it is possible to implement a system capable of immediately determining information on the stage of pressure ulcers without the help of an expert by processing photos taken of suspected pressure ulcers through the deep learning artificial intelligence algorithm according to the present invention.

In other words, if an existing known deep learning object catcher or classifier is directly learned and applied using step-by-step learning data, in the case of skin diseases such as pressure sores, where the step-by-step characteristics show various aspects, the learning of clear decision boundaries is not guaranteed, so the required accuracy cannot be guaranteed, and in this case, it is necessary to simultaneously learn and apply various characteristic patterns to secure the required accuracy. However, in the present invention, the decision boundary of various classifiers is independently learned by applying a probabilistic model that separates the features of each step, and the vector value synthesized from the characteristics that determines the feature of each step is input into a separate DNN to perform an ensemble technique. The final output is obtained through this process to maximize the accuracy of discrimination. Therefore, through this, it is possible to diagnose the stage of pressure ulcer progression without going through a direct medical examination, and transmit the result to a medical staff to receive prompt treatment for pressure ulcers.

1 is a step diagram of a method for generating a pressure sore disease discrimination model according to an embodiment of the present invention.

2 is a configuration diagram of a deep learning object catcher according to an embodiment of the present invention.

3 is a diagram illustrating CNN binary discriminator label values according to an embodiment of the present invention.

4 is a configuration diagram of a CNN binary discriminator learning structure according to an embodiment of the present invention.

5 is a configuration diagram of a DNN discriminator learning structure according to an embodiment of the present invention.

6 is a step diagram of a diagnosis service method according to an embodiment of the present invention.

Hereinafter, the preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Before explaining the present invention in detail, the terms or words used in this specification should not be construed unconditionally in a conventional or dictionary sense, and in order for the inventor of the present invention to explain his/her invention in the best way Concepts of various terms can be appropriately defined and used.

Furthermore, it should be noted that these terms or words should be interpreted as meanings and concepts consistent with the technical idea of the present invention.

That is, the terms used in this specification are only used to describe preferred embodiments of the present invention, and are not intended to specifically limit the contents of the present invention.

It should be noted that these terms are terms defined in consideration of various possibilities of the present invention.

Also, in this specification, a singular expression may include a plurality of expressions unless the context clearly indicates otherwise.

In addition, it should be noted that similarly, even if expressed in a plurality, it may include a singular meaning.

Throughout this specification, when a component is described as "including" another component, it does not exclude any other component, but further includes any other component, unless otherwise stated. It can mean you can do it.

Furthermore, when a component is described as "existing inside or connected to and installed" of another component, this component may be directly connected to or installed in contact with the other component.

In addition, it may be installed at a certain distance, and in the case of being installed at a certain distance, a third component or means for fixing or connecting the corresponding component to another component may exist. .

Meanwhile, it should be noted that the description of the third component or means may be omitted.

On the other hand, when it is described that a certain element is "directly connected" to another element, or is "directly connected", it should be understood that no third element or means exists.

Similarly, other expressions describing the relationship between components, such as "between" and "directly between", or "adjacent to" and "directly adjacent to" have the same meaning. should be interpreted as

In addition, in the present specification, terms such as "one side", "the other side", "one side", "the other side", "first", and "second" refer to one component with respect to another component. It is used to make it clearly distinguishable from the elements.

However, it should be noted that the meaning of a corresponding component is not limitedly used by such a term.

In addition, in this specification, terms related to positions such as “upper”, “lower”, “left”, “right”, etc., if used, are to be understood as indicating relative positions of corresponding components in the drawing.

In addition, unless an absolute location is specified for these locations, these location-related terms should not be understood as referring to an absolute location.

Moreover, in the specification of the present invention, terms such as "unit", "group", "module", and "device", if used, mean a unit capable of processing one or more functions or operations.

It should be noted that this may be implemented in hardware or software, or a combination of hardware and software.

In the drawings accompanying this specification, the size, position, coupling relationship, etc. of each component constituting the present invention is partially exaggerated, reduced, or omitted in order to sufficiently clearly convey the spirit of the present invention or for convenience of explanation. may be described, and therefore the proportions or scale may not be exact.

In addition, in the following description of the present invention, a detailed description of a configuration that is determined to unnecessarily obscure the subject matter of the present invention, for example, a known technology including the prior art, may be omitted.

In order to solve the above problem, the present invention captures through the suspected affected area from a picture of a suspected pressure sore patient taken by a caregiver, medical staff, etc., labels the picture step by step based on this, and independently distinguishes each step in a DNN. It learns by inputting a value-based vector. As a result, in the case of skin diseases such as pressure sores, when the characteristics of each step have various aspects, the learning of a clear decision boundary is not guaranteed, so the conventional problem that cannot guarantee the necessary accuracy is solved through step-by-step CNN-binary discriminator and DNN learning. Accordingly, the present invention provides a method for generating a deep learning model for detecting pressure sores and determining stages, and a deep learning system for realizing the same. The deep learning system according to the present invention includes a deep learning object catcher, a CNN-binary discriminator, a DNN system, etc. The image-based pressure sore disease discrimination model according to the present invention will be described in more detail through the following examples and drawings. .

Referring to FIG. 2 , first, a picture of a suspected pressure ulcer patient is input. A picture may be taken from a terminal such as a family member or caregiver as well as a medical staff, and transmitted to a server providing an image-based pressure sore disease diagnosis service according to the present invention.

Then, the suspected affected area is captured from the input picture. "Suspicious affected area" refers to a skin area that has characteristics different from other skin conditions and has characteristics of a pressure sore. Detect suspicious lesions. The object catcher of one embodiment may use SSD or YOLO.

Afterwards, the suspected affected area of the photo labeled in each step is input to the CNN-binary discriminator and the logit value of the prediction probability is output. To this end, an embodiment of the present invention separately prepares four deep learning binary discriminators step by step that receive feature map vectors extracted using CNN and classify the steps. , Prepare to learn four decision boundaries: 1st level or lower - 2nd level or higher, 2nd level or lower - 3rd level or higher, and 3rd level or lower - 4th level. 3 shows CNN binary discriminator label values for each step according to an embodiment of the present invention.

In one embodiment of the present invention, a picture of a pressure sore affected area labeled for each step is used to learn the CNN-binary discriminator. The output of the CNN-binary discriminator is given as a logit real value, and learning is performed using binary cross-entropy for the discrimination result. 4 is a configuration diagram of a CNN binary discriminator learning structure according to an embodiment of the present invention.

Thereafter, a DNN is constructed that receives a 4-dimensional vector consisting of the logit values of the four binary discriminators and determines a step. The output of the DNN is a 5-dimensional vector, and the step is determined by applying the softmax function here. The optimal values for the depth and width of the DNN hidden layer are determined through trial and error by repeating learning.

In one embodiment of the present invention, the learning of the DNN is performed by applying the DNN through the CNN-binary discriminator learned by the above method with the step-by-step labeling of the pressure sore affected image as an input, and then using cross-entropy as a loss function It is done by

The pressure sore eradication model generated according to the present invention can label photos for each stage of pressure ulcer progression, and train the DNN to determine the stage of pressure ulcer progression with high accuracy and selectivity.

The present invention also provides a diagnosis service method for diagnosing pressure sores using the model and transmitting the results to medical staff.

Referring to FIG. 6 , the diagnosis method may include inputting a patient's picture into the image-based pressure sore disease discrimination model; Receiving information about the progression of pressure ulcers determined by the image-based pressure sore disease discrimination model from the photo; and transmitting the received information to a user terminal. This solves the problem that errors in the judgment assistance system in the health and medical fields act as a burden on nursing, nursing and medical personnel, and deep learning artificial intelligence designed by taking pictures of the affected area by nursing or nursing personnel of patients with suspected bedsores By processing through an algorithm, it is possible to implement an auxiliary system that can immediately determine the information on the pressure sore stage without the help of an expert.

In the above, several preferred embodiments of the present invention have been described with some examples, but the description of the various embodiments described in the "Specific Contents for Carrying Out the Invention" section is merely illustrative, and the present invention Those skilled in the art will understand from the above description that the present invention can be practiced with various modifications or equivalent implementations of the present invention can be performed.

In addition, since the present invention can be implemented in various other forms, the present invention is not limited by the above description, and the above description is intended to complete the disclosure of the present invention and is common in the technical field to which the present invention belongs. It is only provided to completely inform those skilled in the art of the scope of the present invention, and it should be noted that the present invention is only defined by each claim of the claims.

Industrial applicability is recognized as a deep learning model system for pressure sore disease detection and stage determination.

Claims

a) receiving a picture of a suspected pressure ulcer patient;

(b) capturing a suspected affected area from an input photograph, and then labeling the photograph according to stages of pressure ulcer progression;

(c) inputting the suspected affected area of the photograph labeled in each step into a CNN-binary discriminator and outputting the predicted probability as a logit value;

(d) constructing a DNN that receives the 4-dimensional vector consisting of the logit values and determines the stage of pressure ulcer progression; and

(e) learning the DNN by repeatedly inputting the 4-dimensional vector;

Image-based pressure sore disease discrimination model generation method comprising a.
According to claim 1,

In the step (b), the deep learning model generation method for trapping pressure sore disease and determining the stage, characterized in that for capturing the suspected affected area through a deep learning object capturer.
According to claim 1,

In the step (c), there are 4 CNN-binary discriminators, respectively, the CNN-binary discriminator is voice-level 1 or higher, level 1 or lower-level 2 or higher, level 2 or lower-level 3 or higher, level 3 or lower- A method for generating a deep learning model for capturing pressure sore diseases and determining stages, characterized by learning four decision boundaries in four stages.
According to claim 1,

The CNN-binary discriminator is characterized in that it learns using binary cross-entropy for the discrimination result for the stage, a method for generating a deep learning model for capturing pressure ulcer disease and determining the stage.
According to claim 1,

The DNN determines the pressure sore stage by applying the excitation softmax function to a 5-dimensional vector, and the depth and width of the DNN hidden layer determine optimal values by repeating learning. How to create a deep learning model for
A deep learning system for capturing pressure ulcer diseases and determining stages, generated by the method for generating an image-based pressure ulcer disease discrimination model according to any one of claims 1 to 5.
(a) inputting a patient picture into the deep learning system according to claim 6;

(b) receiving information about the progression of pressure sores determined by the deep learning system from the picture; and

(c) transmitting the received information to a user terminal.