KR20200027148A - System and method for pathological staging of prostate cancer based on Deep Belief Network and Dempster-Shafer theory - Google Patents

Abstract

The present invention relates to a system for predicting a pathological stage of a prostate cancer based on a deep belief network (DBN) and a Dempster-Shafer (DS) theory, to predict the pathology of the prostate cancer with an inference theory, and a method thereof. According to the present invention, the system comprises: a pretreatment unit pretreating each of a prostate specific antigen (PSA), a Gleason score, and a clinical T stage, which are used as input data, into a multi-input variable of the DBN; a DBN unit learning and predicting each of PSA, Gleason score, and clinical T stage data pretreated by the pretreatment unit; and a DS inference unit using the DS theory to linearly combine prediction values of the DBN unit with respect to the PSA, the Gleason score, and the clinical T stage to predict the pathological stage of the prostate cancer.

Description

System and method for pathological staging of prostate cancer based on Deep Belief Network and Dempster-Shafer theory based on Deep Belief Network and Dempster-Shafer theory

The present invention relates to predicting pathological stages of prostate cancer, and more particularly, to a system and method for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory.

Prostate cancer is one of the most common cancers in men, and in 2012, approximately 1.1 million cases and 300,000 deaths in men worldwide. It is also estimated that 40-50% of men could potentially have an exogenous disease. Carcinectomy and radiation therapy are typical treatments for prostate cancer. Choosing a treatment for prostate cancer requires extensive experience and analysis of treatment cases.

Pathology Stage is the process of predicting the likelihood of prostate cancer disease from a patient prior to treatment. Clinical stage assessments are based on data collected from clinical trials available prior to treatment or prior to surgical removal of the tumor. Cancer onset assessment occurs both before and after the tumor is removed and is a clinical and pathological stage.

Pathological staging is determined after tumor tissue removal and post-surgery. This can be more accurate than clinical preparation because it assesses the direct nature of the disease. Therefore, prediction of pathological stage (stage) using clinical data analysis is an important factor in the treatment of prostate cancer.

Pathological staging is important because it provides doctors with optimal treatment and management strategies. For example, radical prostatectomy, a surgical removal of the prostate gland, provides the best treatment method for prostate cancer to be localized and accurate prediction of the pathological stage to provide the most beneficial treatment method.

Currently, the Partin table is used to predict prognostic clinical outcomes for prostate cancer and is structured based on statistical methods such as logistic regression. The Partin table uses clinical test data, including prostate specific antigen (PSA), Gleason score, and Clinical T stage to predict pathological stages.

The Partin table has been validated between 2001 and 2011, but there is a problem with the applicability to current patients due to environmental changes. Therefore, a new classification method using machine learning is needed for accurate prediction of the pathological stage.

That is, the pathological prediction of prostate cancer can be made by predicting the patient's prostate metastasis state before surgery using biopsy information. Since biopsy variables related to pathological predictions have uncertainty about individual differences among patients, a method for classifying according to variables is needed.

The problem to be solved by the present invention is that the biopsy variable related to the above-described pathological prediction has an uncertainty for each individual patient, so there is a need to classify according to the variable, and to solve this need To predict the pathology of prostate cancer based on the Deep Belief Network and Dempster-Shafer theory, which predicts the pathology of prostate cancer using inference theory that solves artificial intelligence and uncertainty using probability. Is to provide a way.

The system for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention for achieving the above technical problem is a prostate-specific antigen (PSA) of a prostate cancer, a Gleason score ) And a clinical T stage (Clinical T stage) as input data, each of the input data pre-processing unit for pre-processing with multiple input variables of the Deep Belief Network (DBN); A DBN unit for learning and predicting prostate-specific antigen (PSA), Gleason score, and clinical T stage data pre-treated by the pre-treatment unit through Deep Belief Network (DBN), respectively; And the prostate-specific antigen (PSA), Gleason score (Gleason score), and clinical T stage (Clinical T stage) of the prostate cancer, using the DS (Dempster-Shafer) theory of DS to linearly combine the predicted values of the prostate cancer It includes a DS reasoning unit for predicting the pathological stage of.

The proximal region is a first conversion unit for converting a prostate-specific antigen (PSA) of prostate cancer into an input variable into a binary number; A second conversion unit that converts a Gleason score of prostate cancer into an input variable and converts it into a binary number; And a third conversion unit for converting a clinical T stage of prostate cancer into an input variable and converting it into a binary number.

The first conversion unit converts the data of the input variable prostate-specific antigen (PSA) into binary numbers, and the second conversion unit converts the data of the input variable Gleason score into flag-type binary numbers, and the third conversion. The department is characterized by converting clinical T stage data into flag type data.

 The DBN unit has a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and inputs each bit of the prostate-specific antigen (PSA) data converted to binary by the first conversion unit. A first DBN consisting of a Deep Belief Network having two output nodes for learning as a node and outputting a probability value of OCD (Organ Confine Disease) and a probability value of Non Organ Confine Disease (NODC); It has a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and learning is performed by using each bit of the Gleason score data converted to binary by the second conversion unit as an input node. A second DBN consisting of a Deep Belief Network having two output nodes that output OCD probability values and NODC probability values; And a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and each bit of the Gleason score data converted to binary by the second converter is used as an input node. And a third DBN consisting of a Deep Belief Network having two output nodes for learning and outputting probability values of OCD and probability values of NODC.

The DS inference unit linearly combines the outputs of the first DBN, the second DBN, and the third DBN using Dempster-Shafer theory to infer probability values for Organ Confine Disease (OCD) and probability values for Non Organ Confine Disease (NOCD). It is characterized by predicting OCD (Organ Confine Disease) or NOCD (Non Organ Confine Disease).

 The DS reasoning unit is the m1, m2, and m3 when the basic probability assignment functions for the three input variables, prostate specific antigen (PSA), Gleason score, and clinical T stage are m1, m2, and m3. , m2 and m3 empty set (1- (OCD probability + NOCD probability)) to calculate the empty set; A first merging unit that merges m1 and m2 using the calculated OCD, NOCD, and empty set probabilities of m1 and m2; A second merger for calculating the probability of OCD and the probability of NOCD by merging m1 and m2 with m3; And

As a result of merging the second merger, if the probability of OCD is greater than the probability of NOCD, it is determined as OCD, and when the probability of NODC is large, it includes a pathological staging unit that determines as NOCD.

A method of predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention for achieving the above technical problem, prostate-specific antigen of prostate cancer (PSA), Gleason score (Gleason score) ) And clinical T stage (Clinical T stage) as an input variable for each of the input variables to learn through a separate Deep Belief Network (DBN); And prostate cancer by linearly combining the predicted values of the DBN for each of the prostate-specific antigen (PSA), Gleason score, and clinical T stage of the prostate cancer using Desster-Shafer (DS) theory. Predicting the pathological stage of the.

The DBN includes two output nodes in which the organ confine disease (OCD) and the non organ confine disease (NOCD) are calculated with probability for each input variable.

 The learning through the DBN comprises converting the data of the three input variables prostate specific antigen (PSA), Gleason score and clinical T stage into binary numbers; Learning each bit of the prostate-specific antigen (PSA) data converted to binary number as an input node and learning through a DBN to infer PSA data as probability values of OCD and NOCD, respectively; Learning each bit of the Gleason score data converted to the binary number as an input node and learning through a DBN to infer the Gleason score data as probability values of OCD and NOCD; And learning each bit of the clinical T stage data converted to the binary number as an input node through DBN to infer the clinical T stage data to the probability values of OCD and NOCD, respectively. Includes.

The pathological stage prediction of prostate cancer using the Dempster-Shafer (DS) theory is the basic probability for the three input variables: prostate specific antigen (PSA), Gleason score, and clinical T stage. When the assignment function is m1, m2 and m3, calculating an empty set (1- (OCD probability + NOCD probability)) of m1, m2 and m3; Merging m1 and m2 using the calculated OCD, NOCD, and empty set probabilities of m1 and m2; Calculating the probability of OCD and the probability of NOCD by merging m1 and m2 with m3; And if the probability of the OCD is greater than the probability of the NOCD as a result of the calculation, inferring to the OCD, and if the probability of the NODC is large, the reasoning to the NOCD.

According to the method and system for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention, the predictive model for prostate cancer pathology staging is based on clinical tests and the spread of cancer Can be used to predict. It is possible to more accurately diagnose cancer at the pathological stage after surgery and determine the degree of metastasis of prostate cancer.

According to the present invention, a pathological stage of prostate cancer can be predicted through a DBN-DS-based multicenter approach.

In addition, the present invention can provide a predictive model that improves accuracy through deep learning and information fusion based on the relationship between data measured using clinical trials.

In addition, DBN-DS according to the present invention removes the uncertainty about the predicted output from the DBN by learning PSA, Gleason score and Clinical T stage variable using 3 DBNs, and uses DS for correct judgment. Combine information.

1 is a block diagram showing an embodiment of the configuration of a system for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention.
Figure 2 shows an example of a deep trust neural network (Deep Belief Network) used in the pathological stage prediction system of prostate cancer according to the present invention.
3 is a block diagram showing the detailed configuration of the DS reasoning unit.
4 is a flowchart illustrating a method for predicting pathological stages of prostate cancer based on a Deep Belief Network and Dempster-Shafer theory according to the present invention.
5 is a learning flow diagram for learning a deep learning model using a deep trust neural network (Deep Belief Network) in a system and method for predicting pathological stages of prostate cancer according to the present invention.
6 is a flowchart illustrating a method for inferring a pathology stage according to data input in a system and method for predicting pathological stages of prostate cancer according to the present invention.
7 shows an example of implementation of a system for predicting pathological stages of prostate cancer based on a Deep Belief Network and a Dempster-Shafer theory according to the present invention.
8 shows ROC curves of various classification methods using a validation set.
9 shows the ROC curve of the DBN-DS based classification method according to the present invention using one validation set.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Configurations shown in the embodiments and drawings described herein are only one preferred embodiment of the present invention, and do not represent all of the technical spirit of the present invention, various equivalents that may be substituted for them at the time of the present application It should be understood that there may be variations and variations.

1 is a block diagram showing an embodiment of the configuration of a system for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention. The system for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention comprises a DBN unit 120 and a DS reasoning unit 130, and a pre-processing unit 110 ) May be further included.

The input variable used in the present invention uses prostate-specific antigen (PSA), Gleason score and clinical T stage of prostate cancer. PSA level, Gleason score, and Clinical T state are important variables for pathology prediction, and because the distribution of data can be different for each patient, there is uncertainty, so a combination of evidence for each variable is required.

The prostate specific antigen (PSA) is a proteolytic enzyme synthesized in epithelial cells of the prostate, and is rarely expressed in tissues other than the prostate, and thus is a useful tumor marker used for the selection of prostate cancer. However, PSA is specific for prostate tissue but not tumor, so it may increase in prostate hyperplasia, prostatitis, and prostate infarction. PSA can be useful for screening prostate cancer as well as for recurrence determination after surgery.

The Gleason score (Gleason score) is an index expressing the malignancy of prostate cancer, the pathologist observed the prostate tissue under a microscope and scored the Gleason score from 1 to 5, the score of the most sites and the score of the most sites Add up. Therefore, the Gleason score is between 2 and 10, and the Gleason score can accurately predict the prognosis of prostate cancer. The clinical T stage represents the clinical stage of prostate cancer.

The output parameters of the pathologic staging system of prostate cancer based on the Deep Belief Network and Dempster-Shafer theory according to the present invention include pathological T stages (pT2a, pT2b, pT3a, pT3b, pT3c) and N stages ( pN1) can be used, and finally, the output variable is Organ Confine Disease (OCD; pT2 +) and Non Organ Confine Disease (NOCD, pT3) based on the criteria of the American Joint Committee on Cancer (AJCC) guidelines. + Or N +)

The pre-processing unit 110 uses the prostate-specific antigen (PSA), the Gleason score, and the clinical T stage of the prostate cancer as input data, and inputs each of the input data into a deep belief network (DBN) multiplexed. It is pre-processed as a variable, and includes a first conversion unit 112, a second conversion unit 114, and a third conversion unit 116.

The first converter 112 converts the prostate-specific antigen (PSA) of the prostate cancer into an input variable and converts the data of the input variable prostate-specific antigen (PSA) into a binary number, for example. The second conversion unit 114 converts the Gleason score of the prostate cancer into an input variable and converts the data of the input variable Gleason score into a binary number in the form of a flag. The third conversion unit 116 may convert a clinical T stage of prostate cancer into an input variable and convert the data of the input variable Gleason score into a binary number in a flag format. .

The DBN unit 120 learns through the Deep Belief Network (DBN) for each of the prostate specific antigen (PSA), Gleason score, and Clinical T stage data preprocessed by the pretreatment unit 110. And predict the stage. DBN (Deep Belief Network) is a deep learning technology, which is effective for classification prediction. DBN supports both unsupervised learning and supervised learning, so you can effectively learn about uncertain data relationships. Deep belief network (DBN) is a type of deep neural network or a generative graphical model consisting of several layers of potential variables, where nodes within each layer are not connected, and layers and layers The nodes between are connected. DBN consists of several layers of Restrict Bolzmann Machine (RBM). In the DBN learning method, the visible layer and the hidden layer are composed of one RBM layer, and may be composed of several layers of RBM. When learning through the first layer, data to the next layer is transmitted and re-learned. In this way, it is performed sequentially up to the last layer. In addition, the classification method using DBN is composed at the top layer of DBN and uses Back propagation. DBN shows better results than artificial neural networks (ANN) using randomly selected connection strengths.

Figure 2 shows an example of a deep trust neural network (Deep Belief Network) used in the pathological stage prediction system of prostate cancer according to the present invention. Referring to FIG. 2, the DBN used in the present invention may be configured with multiple classifiers for multiple input variables and output variables, and one classifier for each variable. Therefore, it is necessary to convert one variable to multiple input values, and since PSA is continuous data, it is converted to binary to construct an input node, and since Gleason Score and Clinical T stage are categorical data, data is configured in flag format to input. Nodes can be configured.

Referring to FIG. 1, the DBN unit 120 includes a first DBN 122, a second DBN 124, and a third DBN 126. The first DBN 122 has a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and data of a prostate-specific antigen (PSA) converted to binary by the first conversion unit 112 It consists of a Deep Belief Network, which has two output nodes that learn each bit as an input node and output the probability value of OCD (Organ Confine Disease) and the probability value of NODC (Non Organ Confine Disease).

The second DBN 124 has a deep learning structure in which a plurality of layers are stacked by using a limited Boltzmann machine (RBM) as a building block, and the Gleason score data converted into binary by the second conversion unit 114 It consists of a Deep Belief Network with two output nodes that learn each bit as an input node and output the probability value of the OCD and the probability value of the NODC.

The third DBN 126 has a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and each bit of the Gleason score data converted to binary by the second conversion unit. It consists of a Deep Belief Network having two output nodes for learning as an input node and outputting a probability value of OCD and a probability value of NODC.

The DS inference unit 130 calculates the predicted value of the DBN unit 120 for each of the prostate-specific antigen (PSA), Gleason score, and clinical T stage of the prostate cancer DS (Dempster-Shafer) Predict the pathology stage of prostate cancer by linear combination using theory, and linearly combine the outputs of first DBN 122, second DBN 124, and third DBN 126 using Dempster-Shafer theory By inferring the probability value for OCD (Organ Confine Disease) and the probability value for Non Organ Confine Disease (NOCD), the probability value is high and inferring OCD (Organ Confine Disease) or NOCD (Non Organ Confine Disease).

In the present invention, multiple classifiers are used for the input variables, and the prediction of classifiers is linearly combined using a DS theory (Dempster-Shafer Theory). Dempster-Shafer Theory is a mathematical theory dealing with uncertain and inaccurate problems presented by Arthur Dempster and Glenn Shafer. DS provides an effective method for establishing evidence intervals by using belief values and likelihood values for data sets. DS expresses the degree of confidence in intervals and sets a set of hypotheses that are mutually exclusive, such as probability. The set of objects is called the environment and is denoted by θ.

θ can have multiple elements such as θ = {θ ₁ , θ ₂ , θ ₃ , ..., θ _n }, and the number of subsets is 2 ^k . When θ has only one element, it is called an identification frame. What consists of 2 ^k subsets is called Setset (Power Set) and is denoted by Θ. What consists of 2 ^k subsets is called Setset (Power Set) and is denoted by Θ. The degree to which Θ is supported by some evidence is called the basic probability assignment function m, and is shown in Equation 1. m is mapped to a probability value of 0 for the empty set, and m is 1 for all subsets of Θ (Equation 2).

Based on the evidence given in DS, Belief (H), a belief value for an arbitrary hypothesis H (Hypnosis), is equal to Equation 3.

The degree of confidence is determined by the reliability of the given evidence and the overall environmental impact, and the ratio of the degree can be expressed as e (Equation 4).

r is a value between 0 and 1 (0≤r≤1), true if r = 0, false if r = 1. DS calculates the value of new faith through the process of fusion between different evidences. Therefore, the fusion between the evidences can be expressed as in Equation 5, and if X = Y = D, the belief value of the fusion between the two evidences is 0.

DS does not express the confidence measure for H as a Bel (H) value, but is represented by [Bel (H), Pls (H)] and strong intervals. This interval is called the Evidential Interval. Plausibility (Pls) refers to the extent that the hypothesis is not denied based on evidence (empty intervals except for the true and false intervals), and the possibility of maximum trust. Bel has a range from 0 to 1 (range of true and false), Pls can be defined as Equation 6, and has a value of [0,1]. Also, likelihood of fusion of faith values, likelihood values can represent the process of fusion from multiple evidences.

In the present invention, three output data predicted from a multi classifier are fused and calculated.

3 is a block diagram showing the detailed configuration of the DS reasoning unit 130. Referring to FIG. 3, the DS inference unit 130 may include an empty set calculation unit 310, a first merge unit 320, a second merge unit 330, and a stage reasoning unit 340.

The empty set calculation unit 310 when the basic probability assignment functions for the three input variables prostate specific antigen (PSA), Gleason score, and clinical T stage are m1, m2, and m3 , Calculate the empty set of m1, m2 and m3 (1- (OCD probability + NOCD probability)). The first merging unit 320 merges m1 and m2 using the calculated OCD, NOCD, and empty set probabilities of m1 and m2. The second merging unit 330 calculates the probability of OCD and the probability of NOCD by merging m1 and m2 and m3. The staging unit 340 determines the OCD when the probability of the OCD is greater than the probability of the NOCD as the result of the merge of the second merger 330, and infers the NOCD when the probability of the NODC is large.

The DS reasoning unit 130 will be described in detail. The first DBN (Initial PSA, 122) is m1, the second DBN (Gleason score, 124) is m2, the third DBN (Clinical T stage, 126) is m3, and everything is composed of m4. For the output data, the empty set of m1, m2 and m3 is as shown in Equation 7.

As above, m1, m2, and m3 are obtained, respectively, and then combined with m4. The combination of m4 is as shown in Equation 8.

Next, the evidential interval section of the OCD and NOCD can be summarized as in Equation 9.

As shown above, evidential interval section is configured for OCD and NOCD, and the highest probability value among OCD and NOCD is determined as the final output value.

4 is a flowchart illustrating a method for predicting pathological stages of prostate cancer based on a Deep Belief Network and Dempster-Shafer theory according to the present invention. According to the method of predicting the pathological stage of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention, the prostate-specific antigen (PSA), Gleason score and clinical T stage of prostate cancer Using (Clinical T stage) as an input variable, each of the input variables is learned through a separate Deep Belief Network (DBN) and outputs OCD and NOCD probability values. The prostate-specific antigen (PSA), Gleason score, and clinical T stage of the prostate cancer were then linearly combined using the DS (Dempster-Shafer) theory to predict the DBN prediction values for each of the prostate cancers. Predict the pathological stage of cancer. The DBN includes two output nodes in which the organ confine disease (OCD) and the non organ confine disease (NOCD) are calculated with probability for each input variable. The output through the Dempster-Shafer (DS) predicts Organ Confine Disease (OCD) and Non Organ Confine Disease (NOCD) by linearly combining the output nodes of the DBN for each input variable. The input variable prostate-specific antigen (PSA) is converted into a binary number and composed of the input node of the DBN, and the input variable Gleason score and clinical T stage are converted into flag data. It can be configured as a DBN input node.

Referring to FIG. 4, the method for predicting pathological stages of prostate cancer based on the Deep Belief Network and the Dempster-Shafer theory according to the present invention is first of all three input variables prostate specific antigen (PSA), Gleason score ( When data for Gleason score) and clinical T stage are input, they are converted into binary numbers (step S410).

Each bit of the prostate-specific antigen (PSA) data converted to the binary number is used as an input node to learn through DBN to infer PSA data as probability values of OCD and NOCD. (Step S420) Also, the Gleason score converted to the binary number (Gleason score) Each bit of the data is used as an input node to learn through the DBN to infer the Gleason score (Gleason score) data to the probability values of OCD and NOCD, respectively. (Step S420) Also, the clinical T stage converted to the binary number ( Clinical T stage) Using each bit of data as an input node, learning through the DBN to infer clinical T stage data to the probability values of OCD and NOCD (step S420).

When the basic probability assignment functions for the three input variables, prostate specific antigen (PSA), Gleason score, and clinical T stage are m1, m2 and m3, the m1, m2 and m3 Calculate the empty set of (1- (OCD probability + NOCD probability)) (step S430).

M1 and m2 are merged using the calculated OCD, NOCD, and empty set probabilities of m1 and m2 (step S440). The result of merging m1 and m2 and m3 are merged to calculate the probability of OCD and the probability of NOCD. (Step S450) If the probability of the OCD is greater than the probability of the NOCD (step S460), infer to the OCD (step S470). If the probability of the NODC is large, infer to the NOCD (step S480).

Figure 5 shows a learning flow chart for learning a deep learning model using a deep trust neural network (Deep Belief Network) in the system and method for predicting pathological stages of prostate cancer according to the present invention. Referring to FIG. 5, PSA, Gleason score, and Clinical T stage data are input (500), and binary conversion of PSA, Gleason score and Clinical T stage data is performed (510). The PSA, Gleason score, and Clinical T stage composed of binary data are trained (520) with a Deep Belief Network, respectively. DBN learning 520 includes PSA data learning 522, Gleason score data learning 524 and Clinical T stage data learning 526. After learning DBN, the learning model is completed (530) with PSA, Gleason score, and Clinical T stage.

6 is a flowchart illustrating a method for inferring a pathology stage according to data input in a system and method for predicting pathological stages of prostate cancer according to the present invention. Referring to FIG. 6, PSA, Gleason score, and Clinical T stage data are input (600), and binary conversion of PSA, Gleason score and Clinical T stage data is performed (610). The PSA, Gleason score, and Clinical T stage composed of binary data are inferred (620) by the Deep Belief Network learning model. DBN learning model inference 620 includes PSA data inference 622, Gleason score data inference 624, and clinical T stage data inference 626.

After inference with the DBN learning model, the inference is made using the Dempster-Shafer method using the results inferred from the PSA, Gleason score, and Clinical T stage (630). In the Dempster-Shafer method, inference (630) calculates the empty set () of m1, m2, and m3 (632), merges m1 and m2 (OCD, NOCD, empty set inference, 634), and the result of the merge of m1 and m2. M3 is merged (OCD, NOCD, co-set inference, 636), and finally, the result of OCD and NOCD of m4 is provided (638).

Examples for FIGS. 5 and 6 are as follows. The input data is PSA (m1), the actual data is 88.85, and when converted to binary, it becomes 1011000, and the probability (OCD / NOCD) from deep learning is 0.126 / 0.812.

For Gleason Score (m2), the actual data is 7a, when converted to binary, it is 00000010000, and the probability (OCD / NOCD) from deep learning is 0.602 / 0.397. For Clinical T stage (m3), the actual data is 1c, when converted to binary, it is 00100000, and the probability (OCD / NOCD) from deep learning is 0.707 / 0.292.

Table 1 shows the deep learning results for the input values.

Real data Binary Deep Learning Results OCD NOCD PSA 88.85 1011000 0.126 0.812 Gleason Score 7a (5) 00000010000 0.602 0.397 Clinical T stage 1c (3) 00100000 0.070 0.292

As an example for reference numeral 630 of FIG. 6, an example in which a Dempster-shafer algorithm is applied will be described. First, an empty set is obtained (reference numeral 632 in FIG. 6).

m1 (φ) = 1- (0.126 + 0.812) = 0.062

m2 (φ) = 1- (0.602 + 0.397) = 0.001

m3 (φ) = 1- (0.707 + 0.292) = 0.001

Calculate OCD and NOCD as follows.

m2(OCD)

m3(OCD)m4 (OCD) = m1 (OCD)

m2 (OCD)

m3 (OCD)

m2(NOCD)

m3(NOCD)m4 (NOCD) = m1 (NOCD)

m2 (NOCD)

m3 (NOCD)

The above process calculates m1 and m2, and then combines m3. The order is as follows.

m2(OCD)을 진행한다.(도 6의 참조번호 634)First, m1 (OCD)

Proceed with m2 (OCD) (reference numeral 634 in FIG. 6).

m2(OCD) = ((m1(OCD)*m2(OCD)) + (m1(OCD)*m2(φ)) + (m2(OCD)*m1(φ))) * (1/(1-((m1(OCD)*m2(NOCD)) + (m2(OCD)*m1(NOCD))))m1 (OCD)

m2 (OCD) = ((m1 (OCD) * m2 (OCD)) + (m1 (OCD) * m2 (φ)) + (m2 (OCD) * m1 (φ))) * (1 / (1- ( (m1 (OCD) * m2 (NOCD)) + (m2 (OCD) * m1 (NOCD))))

= ((0.126 * 0.602) + (0.126 * 0.001) + (0.602 * 0.062)) * (1 / (1-((0.126 * 0.397) + (0.602 * 0.812))))

= (0.0759 + 0.0001 + 0.0373) * (1 / (1-0.5388)

= 0.1133 * 2.1685

= 0.2457

m2(NOCD)을 진행합니다. 계산 방법은 m1(OCD)

m2(OCD)와 동일하다.(도 6의 참조번호 634)Second, m1 (NOCD)

Proceed with m2 (NOCD). The calculation method is m1 (OCD)

It is the same as m2 (OCD). (Reference number 634 in FIG. 6)

m2(NOCD) m1 (NOCD)

m2 (NOCD)

= ((m1 (NOCD) * m2 (NOCD)) + (m1 (NOCD) * m2 (φ)) + (m2 (NOCD) * m1 (φ)))) * (1 / (1-((m1 (OCD) ) * m2 (NOCD)) + (m2 (OCD) * m1 (NOCD))))

= ((0.812 * 0.397) + (0.812 * 0.001) + (0.397 * 0.062)) * (1 / (1-((0.126 * 0.397) + (0.602 * 0.812))))

= (0.3224 + 0.0246 + 0.0008) * (1 / (1-0.5388)

= 0.3478 * 2.1685 = 0.7542

m2(φ)을 계산한다. 이것을 해야 m3을 합병할 수 있다.(도 6의 참조번호 634)Third, m1 (φ)

Calculate m2 (φ). This is necessary to merge m3 (reference numeral 634 in FIG. 6).

m2(φ)=(m1(φ)*m2(φ))*(1/(1-((m1(OCD)*m2(NOCD))+(m2(OCD)*m1(NOCD))))m1 (φ)

m2 (φ) = (m1 (φ) * m2 (φ)) * (1 / (1-((m1 (OCD) * m2 (NOCD)) + (m2 (OCD) * m1 (NOCD))))

= 0.0001 * (1 / (1-((0.126 * 0.397) + (0.602 * 0.812))))

= 0.0001 * (1 / (1-0.5388)

= 0.0001 (This is the result of calculating the empty set of m1 and m2.)

The combination result of m1 and m2 is OCD: 0.2457, NOCD: 0.7542, and φ: 0.0001.

Next, it is combined with m3 using the calculated OCD, NOCD, and empty set of m1 and m2. The final m4 is OCD. (Reference number 636 in FIG. 6)

m2(OCD)

m3(OCD)m4 (OCD) = m1 (OCD)

m2 (OCD)

m3 (OCD)

m2(OCD)*m3(OCD))+(m1(OCD)

m2(OCD)*m3(φ))+(m3(OCD)*m1(φ)m2(φ)))*(1/(1-((m1(OCD)

m2(OCD)*m3(NOCD))+((m1(NOCD)

m2(NOCD)*m3(OCD)))= ((m1 (OCD)

m2 (OCD) * m3 (OCD)) + (m1 (OCD)

m2 (OCD) * m3 (φ)) + (m3 (OCD) * m1 (φ) m2 (φ))) * (1 / (1-((m1 (OCD)

m2 (OCD) * m3 (NOCD)) + ((m1 (NOCD)

m2 (NOCD) * m3 (OCD)))

= ((0.2457 * 0.602) + (0.2457 * 0.001) + (0.602 * 0.0001) * (1 / (1- (0.0975 + 0.4540)))

= (0.1479 + 0.0002 + 0.0001) * (1 / (1-0.5516))

= 0.1482 * 2.2299

= 0.3305

Finally, m1 and m2 are combined with m3 using OCD, NOCD, and empty set. It is the final m4 NOCD. (Reference number 636 in FIG. 6)

m2(NOCD)

m3(NOCD)m4 (NOCD) = m1 (NOCD)

m2 (NOCD)

m3 (NOCD)

m2(NOCD)*m3(NOCD))+(m1(NOCD)

m2(NOCD)*m3(φ))+ (m3(NOCD)*m1(φ)m2(φ)))*(1/(1-((m1(OCD)

m2(OCD)*m3(NOCD))+ ((m1(NOCD)

m2(NOCD)*m3(OCD) ))= ((m1 (NOCD)

m2 (NOCD) * m3 (NOCD)) + (m1 (NOCD)

m2 (NOCD) * m3 (φ)) + (m3 (NOCD) * m1 (φ) m2 (φ))) * (1 / (1-((m1 (OCD)

m2 (OCD) * m3 (NOCD)) + ((m1 (NOCD)

m2 (NOCD) * m3 (OCD)))

= ((0.7542 * 0.397) + (0.7542 * 0.001) + (0.397 * 0.0001)) * (0.602 * 0.0001) * (1 / (1- (0.0975 + 0.4540)))

= (0.2994 + 0.0008 + 0.0001) * (1 / (1- (0.0975 + 0.4540)))

= 0.3002 * (1 / (1-0.5516))

= 0.3002 * 2.2299

= 0.6695

As a result, according to the input data, OCD has a probability of 0.3305, and NOCD has a probability of 0.6695. Because NOCD has a high probability, it is diagnosed as NOCD (reference numeral 638 in FIG. 6).

Meanwhile, an experimental example of a system and method for predicting pathological stages of prostate cancer based on Deep Belief Network and Dempster-Shafer theory according to the present invention will be described.

Table 2 is a summary of the initial PSA of 6,345 patients at the pathological stage of OCD or NOCD with clinically localized prostate carcinoma. Table 3 is the distribution of Gleason scores of 6,345 patients at the pathological stage of OCD or NOCD with clinically localized prostate carcinoma. Table 4 is the clinical T stage distribution of 6,345 patients at the pathological stage of OCD or NOCD with clinically localized prostate carcinoma.

As shown in Table 2, Table 3 and Table 4, as a characteristic of the Gleason score used in the experimental example, the Gleason score of OCD is widely distributed in 6. Below 5 points, OCD is more distributed than NOCD. NOCD is mostly distributed over 6 points. The difference between the OCD and NOCD groups is significant. As a characteristic of the clinical T stage, most patients are distributed in T2 +. T1a appeared only in OCD. In addition, T1 + belongs to OCD a lot, and T3 + belongs to NOCD a lot.

7 is an example of an embodiment of the construction of a system for predicting pathological stages of prostate cancer based on a Deep Belief Network and Dempster-Shafer theory according to the present invention, and is used in an experimental example. Referring to FIG. 7, the training set 710 is changed to a binary format through the conversion unit 720. The initial PSA value is converted to 9 binary numbers based on the highest value (440 ng / mL). The Gleason score consists of 9 flags from 3 to 10. The clinical T stage consists of 8 flags from T1a to T3b.

The binary data of each of the three variables is trained by DBN classifiers 730, 740, and 750. That is, since the first DBN 530 is input data of initial PSA binary data, it consists of 9 input nodes. The output nodes of the DBN classifiers 730, 740, and 750 are composed of two so that OCD and NOCD can be calculated with probability. One DBN consists of three RBM layers, and the number of nodes in each RBM is the same as the number of input nodes. In the experimental example of the present invention, Unsupervised learning is performed 100 times, and Supervised learning using back propagation is performed 1000 times. Finally, the probability of the output variable is calculated as DS 760, and the final number of m4 (OCD) and m4 (NOCD) is determined as the final output.

Table 5 shows the experimental results of the classification method between the training set and the validation set.

ive Bayesian, Logistic Regression, Back Propagation, SVM, RF, Simple DBN, Partin Table 비교하였다. 실험 측도는 Sensitivity, Specificity, Accuracy, ROC curve를 이용하였다.In Table 5, consisting of 70% training set and 30% validation set, for comparison, C4.5, Na

We compared ive Bayesian, Logistic Regression, Back Propagation, SVM, RF, Simple DBN, Partin Table. Sensitivity, Specificity, Accuracy, and ROC curves were used for the experimental measurements.

Table 6 shows detailed ROC analysis results of various classification methods using a validation set.

8 shows ROC curves of various classification methods using a validation set.

Table 7 shows the results of experiments for each DBN combination. Table 7 shows the results of the DBN-DS mixing matrix comparing the training set and the validation set. DBN # 1 is for PSA, DBN # 2 is for Gleason score, and DBN # 3 is for Clinical T stage.

9 shows an ROC curve of a DBN-DS-based classification method according to the present invention using one validation set.

In the predictive model for prostate cancer pathology staging described above, the input uses Initial PSA level, Gleason score and clinical T stage variables, and the output uses OCD and NOCD in pathological staging (pT). This approach was evaluated using an existing validated patient data set that included 6345 patient records from the KPCR database, which collected data from multiple tertiary care institutions.

The performance of DBN-DS according to the present invention was compared with that of NB, LR, BPN, SVM, RF, DBN and Partin tables. The results showed that the DBN-DS according to the present invention performed better than all other methods.

The present invention can be embodied as computer readable code on a computer readable recording medium (including all devices having information processing functions). The computer-readable recording medium includes any kind of recording device in which data readable by a computer system is stored. Examples of computer-readable recording devices include ROM, RAM, CD-ROM, magnetic tape, floppy disks, and optical data storage devices. Further, in this specification, “unit” may be a hardware component such as a processor or a circuit, and / or a software component executed by a hardware component such as a processor.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

110: pre-processing unit 112: first conversion unit
114: second conversion unit 116: third conversion unit
120: DBN unit 122: first DBN
124: second DBN 126: third DBN
130: DS reasoning unit 310: empty set calculation unit
320: first merger 330: second merger
340: Ordnance Reasoning Division

Claims (12)

A pre-processing unit that pre-processes each of the input data as multiple input variables of Deep Belief Network (DBN) using prostate-specific antigen (PSA), Gleason score, and clinical T stage of prostate cancer as input data ; And
A DBN unit for learning and predicting prostate-specific antigen (PSA), Gleason score, and clinical T stage data pre-treated by the pre-treatment unit through Deep Belief Network (DBN), respectively; And
Prostate specific antigen (PSA), Gleason score (Gleason score) and the predicted value of the DBN part for each of the clinical T stages of the prostate cancer were linearly coupled using the theory of DS (Dempster-Shafer) to prostate cancer A pathological staging system for prostate cancer based on Deep Belief Network (DS) and Dempster-Shafer (DS) theories, including a DS inference unit for predicting pathological staging. The method according to claim 1, wherein the front teeth
A first conversion unit for converting a prostate-specific antigen (PSA) of prostate cancer into an input variable;
A second conversion unit that converts a Gleason score of prostate cancer into an input variable and converts it into a binary number; And
Pathology of prostate cancer based on Deep Belief Network (DS) and Dempster-Shafer (DS) theory, including a third transform unit that converts the clinical T stage of prostate cancer into a binary number using the clinical T stage as an input variable Staging system. The method of claim 2, wherein the first conversion unit
Convert the data of the input variable prostate specific antigen (PSA) to binary,
The second conversion unit
Convert the data of the input variable Gleason score (Gleason score) to a binary number in the form of a flag,
The third conversion unit
A system for predicting pathological stages of prostate cancer based on Deep Belief Network (DS) and Dempster-Shafer (DS) theory, characterized by converting clinical T stage data into flag type data. The method of claim 2, wherein the DBN unit
A limited Boltzmann machine (RBM) as a building block has a deep learning structure in which a plurality of layers are stacked, and each bit of data of the prostate-specific antigen (PSA) converted to binary by the first converter is used as an input node. A first DBN consisting of a Deep Belief Network having two output nodes for learning and outputting probability values of Organ Confine Disease (OCD) and probability values of Non Organ Confine Disease (NODC);
It has a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and learning is performed by using each bit of the Gleason score data converted to binary by the second conversion unit as an input node. A second DBN consisting of a Deep Belief Network having two output nodes that output OCD probability values and NODC probability values; And
It has a deep learning structure in which a plurality of layers are stacked using a limited Boltzmann machine (RBM) as a building block, and learning is performed by using each bit of the Gleason score data converted to binary by the second conversion unit as an input node. A deep trust neural network (Deep Belief Network) and DS (Dempster-Shafer) theory, including a third DBN consisting of a deep trust network (Deep Belief Network) having two output nodes that output the probability values of OCD and NODC -Based system for predicting pathology of prostate cancer. The method of claim 2, wherein the DS reasoning unit
The output of the first DBN, the second DBN, and the third DBN is linearly combined using Dempster-Shafer theory to infer the probability value for Organ Confine Disease (OCD) and the probability value for Non Organ Confine Disease (NOCD), and the OCD has a high probability value. (Organ Confine Disease) or NOCD (Non Organ Confine Disease) predicted pathological staging system of prostate cancer based on the Deep Belief Network (Deep Belief Network) and DS (Dempster-Shafer) theory, characterized by predicting. The method of claim 5, wherein the DS reasoning unit
When the basic probability assignment functions for the three input variables, prostate specific antigen (PSA), Gleason score, and clinical T stage are m1, m2 and m3, the m1, m2 and m3 An empty set calculation unit that calculates the empty set (1- (OCD probability + NOCD probability));
A first merge unit that merges m1 and m2 using the calculated OCD, NOCD, and empty set probabilities of m1 and m2
A second merger for calculating the probability of OCD and the probability of NOCD by merging m1 and m2 with m3; And
The deep merge neural network (Deep Belief Network), characterized in that it includes a pathological staging unit that determines the OCD when the probability of OCD is greater than the probability of NOCD as a result of merging the second merger, and determines that the probability of NODC is NOCD. Pathology staging system for prostate cancer based on DS (Dempster-Shafer) theory. Learning through a separate Deep Belief Network (DBN) for each of the input variables using prostate-specific antigen (PSA), Gleason score, and Clinical T stage of the prostate cancer as input variables; And
The prostate-specific antigen (PSA), Gleason score (Gleason score) and the predicted value of the DBN for each of the clinical T stages of the prostate cancer were linearly coupled using the DS (Dempster-Shafer) theory to determine the prostate cancer. A method of predicting pathological stages of prostate cancer based on Deep Belief Network (DS) and Dempster-Shafer (DS) theories, including the step of inferring pathological stages. The method of claim 7, wherein the DBN
Deep Belief Network (DS) and Dempster-Shafer (DS), characterized in that it includes two output nodes, each of which is computed with probability of Organ Confine Disease (OCD) and Non Organ Confine Disease (NOCD) for each input variable. ) A method for predicting pathological stages of prostate cancer based on theory. The method of claim 8, wherein the output through the DS (Dempster-Shafer)
Deep Belief Network (DS) and Dempster-Shafer (DS) characterized by predicting OCD (Organ Confine Disease) and NOCD (Non Organ Confine Disease) by linearly combining DBN output nodes for each input variable. A method for predicting pathological stages of prostate cancer based on theory. The method of claim 7,
The input variable prostate-specific antigen (PSA) is converted into a binary number and is composed of the input node of the DBN,
The input variable Gleason score (Gleason score) and clinical T stage (Clinical T stage) is converted to flag-type data, characterized in that composed of the input node of the DBN, Deep Trust Neural Network (Deep Belief Network) and DS ( Dempster-Shafer) method for predicting the pathological stage of prostate cancer based on the theory. The method of claim 7, wherein learning through the DBN
Converting data for the three input variables prostate specific antigen (PSA), Gleason score and clinical T stage into binary;
Learning each bit of the prostate-specific antigen (PSA) data converted to binary number as an input node and learning through a DBN to infer PSA data as probability values of OCD and NOCD, respectively;
Learning each bit of the Gleason score data converted to the binary number as an input node and learning through a DBN to infer the Gleason score data as probability values of OCD and NOCD; And
Including each bit of the clinical T stage (Clinical T stage) data converted to the binary number through the DBN to learn the clinical T stage (Clinical T stage) data to infer the probability value of each of the OCD and NOCD A method of predicting pathological stages of prostate cancer based on Deep Belief Network (DS) and Dempster-Shafer (DS) theory. 12. The method of claim 11, wherein the pathological stage prediction of prostate cancer using the DS (Dempster-Shafer) theory is
When the basic probability assignment functions for the three input variables, prostate specific antigen (PSA), Gleason score, and clinical T stage are m1, m2 and m3, the m1, m2 and m3 Calculating an empty set of (1- (OCD probability + NOCD probability));
Merging m1 and m2 using the calculated OCD, NOCD, and empty set probabilities of m1 and m2;
Calculating the probability of OCD and the probability of NOCD by merging m1 and m2 with m3; And
A deep trust network (Deep Belief Network) and a DS (Dempster-Shafer) theory comprising the step of inferring to OCD when the probability of the OCD is greater than the probability of the NOCD, and inferring to the NOCD when the probability of the NODC is large. -Based method for predicting the pathological stage of prostate cancer.

Images