KR20210084214A - A recurrence prediction system based on deep learning for prostate cancer using time series data of examination - Google Patents

Abstract

The present invention relates to a prostate cancer recurrence prediction system based on deep learning using time series examination data, which is able to collect examination data of regular examinations before and after the diagnosis and the surgery for prostate cancer, especially the prostate specific antigen (PSA) data of a patient, apply the collected time series data of the PSA to a deep learning neural network, and predict the prostate cancer recurrence of the patient, comprising: a neural network module having a circulation neural network; a pre-processing unit which normalizes a large number of a series of PSA levels in the order of time series; a neural network learning unit which makes the PSA learning data pre-processed by the neural network pre-processing unit, and makes the neural network learn by using the pre-processed learning data; and a determination unit which receives the series of PSA data of the patient, makes the neural network pre-processing unit pre-process the PSA data of the patient, applies the pre-processed data to the neural network, and outputs the presence or absence of the recurrence of prostate cancer in accordance with the results of application. The present invention is able to normalize the examined PSA data on a regular basis in a time series, pre-process the PSA data, make the deep learning neural network more effectively learn, and more precisely diagnose the recurrence of prostate cancer.

Description

{ A recurrence prediction system based on deep learning for prostate cancer using time series data of examination }

The present invention collects checkup data regularly checked before and after diagnosis of prostate cancer and before and after surgery, in particular, a patient's prostate-specific antigen (PSA) data, and applies the collected PSA time series data to a neural network, It relates to a deep learning-based prostate cancer recurrence prediction system using time-series screening data that predicts recurrence.

In general, since the symptoms and prognosis experienced by cancer patients vary, it is important to treat each patient with an active treatment suitable for each patient's condition. Therefore, evaluation of the therapeutic effect of cancer treatment is an important task in the field of cancer treatment. In addition, although cancer appears to have been temporarily removed through surgery or other treatment processes, there are cases where cancer that has not been completely removed by surgery remains. very important.

In particular, prostate cancer (C61) is a malignant neoplasm that occurs in the prostate. It is diagnosed by directly palpating the prostate area with hands or using a Prostate Specific Antigen (PSA) test. Prostate cancer is suspected. If necessary, the characteristics of cancer cells are diagnosed through ultrasound, MRI, or CT. In other words, from the time when prostate cancer is suspected and visit the hospital to the time when 5-year survival after surgery is confirmed, patients receive PSA tests at intervals of 1 month and 3 months as short as 1 month. become wider

According to international clinical guidelines, if the PSA level is 3 or higher, the possibility of prostate cancer is diagnosed. However, even in patients with a PSA level of less than 10, the probability of confirming prostate cancer is less than 20%. In other words, it is difficult to diagnose or confirm the recurrence of prostate cancer with only one PSA test.

In order to solve this problem, techniques for diagnosing prostate cancer more accurately and effectively preventing recurrence after surgery have been proposed. As an example, a technique for diagnosing cancer occurrence with the learned AI using clinical values and cancer progression stage values has been proposed [Patent Document 1]. Also, a technique for predicting the pathological stage of prostate cancer by deep learning various variables affecting the occurrence and progression of prostate cancer has been proposed [Patent Document 2].

However, the prior art uses a simple examination result of a patient and does not consider a change in the PSA level that is regularly checked. Therefore, a change in the level of PSA, which is measured regularly, can have a positive effect on the occurrence and recurrence of prostate cancer.

However, since the test time points are different for each patient and the PSA level is greatly affected by the time of drug, radiotherapy, and surgery, it is necessary to correct the PSA level according to the time of the examination at regular intervals. In addition, in reality, all tests and examinations conducted in each hospital inevitably have irregular time points, so the result value is also greatly affected by the corresponding time interval.

Korea Patent Publication No. 10-1018665 (2011.03.04. Announcement) Korean Patent Publication No. 10-2019-0110834 (published on October 1, 2019)

SUMMARY OF THE INVENTION An object of the present invention is to solve the problems described above, by collecting screening data regularly checked before and after diagnosis of prostate cancer and before and after surgery, in particular, prostate-specific antigen test (PSA) data of a patient, and collecting the collected data. To provide a deep learning-based prostate cancer recurrence prediction system using time series screening data that predicts prostate cancer recurrence of a patient by applying time series data of PSA to a recurrent neural network.

In addition, an object of the present invention is to collect PSA (prostate-specific antigen test) data of patients who are regularly screened, normalize the collected PSA data to a time series, and apply the normalized PSA data to a cyclic neural network, time series examination data To provide a deep learning-based prostate cancer recurrence prediction system.

In order to achieve the above object, the present invention relates to a deep learning-based prostate cancer recurrence prediction system using time series examination data, comprising: a neural network module having a recurrent neural network; a preprocessor that normalizes a series of multiple PSA (Prostate Specific Antige) values in time series order; a neural network learning unit that preprocesses PSA learning data by the neural network preprocessor, and trains the neural network using the preprocessed learning data; And, by receiving a series of PSA data of the patient, preprocessing the PSA data of the patient by the neural network preprocessor, applying the preprocessed data to the neural network, and outputting whether prostate cancer recurrence occurs according to the application result It is characterized in that it includes a judgment unit.

In addition, the present invention provides a deep learning-based prostate cancer recurrence prediction system using time-series examination data, wherein the neural network is composed of long short-term memory networks (LSTM), and the neural network outputs PSA values of the next period do it with

In addition, the present invention is characterized in that in the deep learning-based prostate cancer recurrence prediction system using time series examination data, the input of the neural network further includes the patient's biometric index in addition to the time series data of the PSA value.

In addition, in the present invention, in a deep learning-based prostate cancer recurrence prediction system using time-series examination data, the pre-processing unit sets a representative value of each period by binding a series of PSA values to a predetermined period, and time-series representative values of each period It is characterized in that it is configured and normalized.

In addition, in the present invention, in a deep learning-based prostate cancer recurrence prediction system using time-series examination data, when a plurality of PSA values exist within each period by cycle, the preprocessor sets the median value of the numerical values within the period as a representative value. characterized in that

In addition, in the present invention, in a deep learning-based prostate cancer recurrence prediction system using time series examination data, when the PSA level is missing within the period, the weighted moving average method is applied to the PSA level before and after the period. , it is characterized in that the representative value of the corresponding period is estimated.

In addition, in the present invention, in a deep learning-based prostate cancer recurrence prediction system using time-series examination data, the preprocessor estimates the missing representative value using only the representative values of at most two periods before and after the missing representative value, but the weight is 0.5 is assigned to the representative value of the previous period and 0.5 to the representative value of the subsequent period, and in the case of two, 0.3 is assigned to the closest representative value and 0.2 to the next closest representative value, and representative of one period in the preceding or subsequent period It is characterized in that 0.5 is given if there is only a value.

In addition, the present invention is characterized in that in the deep learning-based prostate cancer recurrence prediction system using time series examination data, the time series data is further included in the PSA value and treatment information is configured.

As described above, according to the deep learning-based prostate cancer recurrence prediction system using time-series examination data according to the present invention, by preprocessing the regularly screened PSA data by normalizing the time-series, it is possible to more effectively train the circulatory neural network, and The effect of more accurately diagnosing cancer recurrence is obtained.

1 is a block diagram of an entire system for implementing the present invention.
2 is a block diagram of the configuration of a deep learning-based prostate cancer recurrence prediction system using time series examination data according to an embodiment of the present invention.
3 is a diagram illustrating the structure of a recurrent neural network (RNN) and an LSTM neural network according to an embodiment of the present invention.
4 is a detailed structural diagram of an LSTM neural network according to an embodiment of the present invention.
5 is an exemplary graph for PSA time series data according to an embodiment of the present invention.

Hereinafter, specific contents for carrying out the present invention will be described with reference to the drawings.

In addition, in demonstrating this invention, the same part is attached|subjected with the same code|symbol, and the repetition description is abbreviate|omitted.

First, the configuration of the entire system for implementing the present invention will be described with reference to FIG. 1 .

As shown in Fig. 1(a), a deep learning-based prostate cancer recurrence prediction system (hereinafter, a recurrence prediction system) using time-series examination data according to the present invention receives PSA examination data and predicts the recurrence of prostate cancer, a computer terminal (10) It can be implemented by the above program system.

That is, the recurrence prediction system 30 may be implemented as a program system on the computer terminal 10, such as a PC, a smartphone, or a tablet PC. In particular, the relapse prediction system is composed of a program system or a mobile application (or an application, an app), it can be installed and executed in the computer terminal (10). The recurrence prediction system 30 provides a service for predicting recurrence of prostate cancer by analyzing PSA examination data using hardware or software resources of the computer terminal 10 .

In addition, as another embodiment, as shown in FIG. 1( b ), the relapse prediction system 30 is a server-client system composed of a relapse prediction client 30a and a relapse prediction server 30b on the computer terminal 10 . can be configured and executed.

Meanwhile, the recurrence prediction client 30a and the recurrence prediction server 30b may be implemented according to a typical method of configuring a client and a server. That is, the functions of the entire system can be divided according to the performance of the client or the amount of communication between the server and the server. Hereinafter, it will be described as a recurrence prediction system, but it may be implemented in various forms of distribution depending on the configuration method of the server-client.

Next, a deep learning-based prostate cancer recurrence prediction system 30 using time series examination data according to a first embodiment of the present invention will be described with reference to FIG. 2 .

As shown in Figure 2, the disease prediction service system 30 according to the first embodiment of the present invention is a neural network module 31 having a deep learning neural network, a series of a plurality of PSA (Prostate Specific Antige) values in time series order The preprocessing unit 32 that normalizes the PSA learning data by the preprocessing unit 32, and the learning unit 33 that trains the neural network using the preprocessed learning data, and the patient's PSA data to the neural network It is composed of a determination unit 34 that determines whether prostate cancer recurrence by applying.

First, the neural network module 31 includes a long short-term memory networks (LSTM) neural network.

LSTM (Long Short Term Memory networks) neural network is a type of RNN (Recurrent Neural Network), and has a structure in which past data can influence the future.

As shown in FIG. 3A , a recurrent neural network (RNN) is an artificial neural network in which hidden nodes are connected by directional edges to form a directed circle. In particular, since the recurrent neural network is a network structure that can accept input data and output data regardless of the length of time series data, various and flexible structures can be created according to the type of data to be learned.

In addition, as a kind of neural network, RNN compensates for the one-way limitation of the existing FFNets method, and is useful for handling time-series data or sequence data containing time and sequence information. That is, using the RNN neural network, if the PSA test and examination data to be analyzed has only a certain level of tendency and periodicity, the future PSA value can be predicted using the corresponding characteristics.

In addition, as shown in FIG. 3B , a Long Short Term Memory Network (LSTM) neural network is a type of RNN, but it can overcome the vanishing gradient problem to some extent by supplementing the shortcomings of the RNN. That is, in the existing RNN neural network, the values between [-1, 1] are continuously multiplied by the chain rule, which is a long-term dependency problem, so the value becomes smaller as it goes forward and eventually disappears. That is, the gradient annihilation problem is a problem that occurs as the interval at which the data to be trained is measured increases. Therefore, the existing RNN method has a problem that parameters are not updated.

LSTM is a method proposed to solve this problem, and is a method in which a cell state is added to a hidden node of the RNN structure. In the conventional RNN, when calculating the hidden state (St) in the hidden layer, it is simply calculated as h _t =tanh(Ux _t +Wh _t-1 ), but in LSTM, there are a total of four calculation processes. Specifically, as shown in FIG. 4 , there are four neural network layers in the hidden layer of the LSTM. In addition, the core of the LSTM goes through a very minor linear operation on the horizontal line at the top of the neural (A), and the information is transmitted to the next stage without much change through the entire chain.

On the other hand, the input of the LSTM neural network is time series data of the PSA value, and the output is the predicted value of the next PSA value.

In particular, the time series data of PSA values are PSA values measured after surgery (or treatment), and represent a series of values measured at regular time intervals (hereinafter, measurement period).

In addition, the input of the neural network may further include biometric information (or biometric index) of the patient, such as age, body mass index (BMI) value, Gleason score, highly malignant prostate intraepithelial neoplasia, in addition to the PSA value.

In addition, the output of the neural network is the predicted value of the PSA value immediately after the last PSA value (the PSA value of the last period) in the patient's time series data (the PSA value of the immediately following period).

Next, the preprocessor 32 normalizes a series of a plurality of PSA (Prostate Specific Antige) values in time series order.

Figure 5 illustrates that the PSA values of one patient are listed in time series order. In the example of FIG. 5 , after surgery on 11.01, PSA was examined on each date such as 11.10, 02.08, 03.02, 04.10, 06.27, 08.04, and the PSA values at that time were 10.6, 6.3, 13.5, 11.2, 13.2, respectively. It is measured as 15.3. For convenience of explanation, the numerical value of FIG. 5 is set as a numerical value obtained by converting the original numerical value 1 into 100. That is, 10.6 in FIG. 5 represents 0.106, and patient PSA data are generally recorded to the third decimal place.

Since the test time is different for each patient, and the PSA level is greatly affected by the time of drug treatment, the time of radiotherapy, and the time of surgery, it is necessary to correct the PSA level according to the time of the examination.

First, if the PSA test results exceed the outlier criteria (data distribution and outlier test results), all of them are replaced with the outlier criteria. For example, if the outlier criterion is 15.0, and there is a PSA value of 15.4 or 16.7, the value is corrected to 15.0 the outlier criterion.

In addition, a representative value is set by grouping a series of PSA values of a patient at a certain period. In the example of FIG. 5 , if the cycle is set to 2 months, each period is divided into November-December, January-February, March-April, May-June, and July-August.

On the other hand, if there are multiple PSA values within the corresponding period, the median of these numerical values is set as a representative value. In the example of FIG. 5 , representative values of each period are obtained as 10.6, 6.3, 12.2 (the median of 11.2 and 13.2), 13.2, 15.3, and the like.

In addition, when there is no PSA value within the corresponding period (when the PSA value is missing), the preprocessor 32 estimates the PSA value using the moving average method. That is, when there is a PSA missing value, the corresponding missing value is estimated by applying the weighted moving average method to the data (numerical values) in the period before and after the missing value.

The weighted moving average method is an improved averaging method from the point of view that the latest data has a greater influence on the derivation of the desired result than the older data from the current time point. That is, the weighted moving average method was developed to solve the problem of the simple moving average method, which assumes that the latest data and the old data have the same influence. Therefore, the weighted moving average method is a method of calculating the sum of the coefficients so that the most recent data is given a high weighting factor and the older data is given a low weighting factor.

Preferably, the missing data is estimated using only two data (PSA values) before and after at most in the missing data. In this case, 0.5 weight is given to the previous data and 0.5 weight is given immediately after it. In case of two, 0.3 is given to the closest data and 0.2 to the next closest data. If there is only one value before or after the missing value, 0.5 is assigned.

In the example of FIG. 5 , if the cycle is set to one month, the PSA values (or PSA data) for May and July are missing. In the case of May, there are 13.5 (March) and 11.2 (April) data for the immediately preceding month, and only 13.2 (June) data for the immediately preceding month. Therefore, the missing data for May is calculated as follows.

13.5×0.2 + 11.2×0.3 + 13.2×0.5 = 12.66

That is, the weights of March, April, and June data are given as 0.2, 0.3, and 0.5, respectively.

Next, the learning unit 33 pre-processes the PSA learning data by the pre-processing unit 32 and trains the neural network of the neural network module 31 using the pre-processed learning data.

That is, the learning unit 33 collects prostate cancer surgery patient data, selects patient data who have undergone the PSA test N or more times after surgery, groups the data in an N-month cycle, and extracts representative values.

According to the experiment, the data of patients who received at least 4 post-operative PSA tests at three large hospitals were grouped into three-month intervals. Before pretreatment, 6,977 patients and 74,825 PSA data were collected as raw data. Among the raw data, 63,012 PSA data were extracted from 4,510 patients with 8 or more PSA data per patient. Then, if the PSA data are grouped into three-month units and the missing values are filled using the weighted moving average method specified above, the total number of PSA data is 67,504. At this time, 4,492 pieces of provisional data entered as missing PSA data were generated. That is, the average data per patient was 14.97. This data includes even temporary data entered as missing data.

After estimating the value by applying the weighted moving average method to the provisional data entered as the above missing data, the completed data without missing data was trained using the LSTM technique.

Next, the determination unit 34 applies the patient's PSA data to the neural network to determine whether prostate cancer recurs.

That is, the determination unit 34 receives a series of patient PSA data measured after the time of treatment or surgery, that is, time series data of the patient's PSA value (when input is received), preprocessing is performed through the preprocessing unit 32 in advance. make it

Next, the determination unit 34 inputs the preprocessed time series data of the PSA value of the patient to the LSTM neural network, and obtains an output value of the LSTM neural network.

Next, the determination unit 34 determines whether prostate cancer reoccurs based on the obtained output value of the neural network. That is, the output value of the neural network is the predicted value of the patient's next PSA level. If the predicted value (or output value) falls within a predetermined reference range, it is determined that the prostate cancer has recurred. Preferably, if the predicted PSA level exceeds the recurrence criterion (predetermined criterion) of 0.2, it is determined that the prostate cancer has recurred, and if it does not exceed it, it is determined that it does not.

Next, a deep learning-based prostate cancer recurrence prediction system 30 using time-series examination data according to a second embodiment of the present invention will be described.

The second embodiment of the present invention is the same as the first embodiment described above, except that time series data is configured by further including treatment information (or treatment data) in the PSA value.

Preferably, the time series data of the treatment information consists of medication information, radiation treatment cost, and the like within a corresponding period.

In addition, preferably, the dosing information or radiation therapy is set to a binary value according to whether the administration or treatment. For example, if it was administered or treated within the period, it is input as “1”, otherwise it is input as “0”.

In other words, if it is confirmed that the patient's PSA level rises abnormally after surgery, measures such as hormone therapy, radiation therapy, and chemotherapy are performed before re-operation, and if it does not improve, re-operation is performed. Therefore, in consideration of the dates of the above treatments that can affect the change in the patient's PSA level in the deep learning predictive model, a variable that specifies the information (coded as 1 or 0) on the PSA test day just before the PSA level changes significantly is added. do. In fact, adding these variables also has a positive effect on performance.

In the above, the invention made by the present inventors has been described in detail according to the above embodiments, but the present invention is not limited to the above embodiments, and various modifications can be made without departing from the gist of the present invention.

10: computer terminal 20: recurrence prediction client
30: relapse prediction system 31: neural network module
32: preprocessing unit 33: learning unit
34: judgment unit
40: database 80: network

Claims (8)

In a deep learning-based prostate cancer recurrence prediction system using time series examination data,
a neural network module having a recurrent neural network;
a preprocessor for normalizing a series of multiple PSA (Prostate Specific Antige) values in time series order;
a neural network learning unit that preprocesses PSA learning data by the neural network preprocessor, and trains the neural network using the preprocessed learning data; and;
A determination unit that receives a series of PSA data of a patient, normalizes the PSA data of the patient by the neural network preprocessing unit, applies the normalized data to the neural network, and outputs whether prostate cancer recurs according to the application result A deep learning-based prostate cancer recurrence prediction system using time-series screening data, characterized in that it includes.
According to claim 1,
The neural network is composed of long short-term memory networks (LSTM), and the neural network is a deep learning-based prostate cancer recurrence prediction system using time series examination data, characterized in that it outputs the PSA value of the next period.
According to claim 1,
The input of the neural network is a deep learning-based prostate cancer recurrence prediction system using time-series examination data, characterized in that it further includes the patient's bio-index in addition to the time-series data of the PSA value.
According to claim 1,
The pre-processing unit sets a representative value of each period by binding a series of PSA values to a predetermined period, and normalizes the representative values of each period by configuring a time series. Deep learning-based prostate cancer recurrence prediction using time series examination data system.
5. The method of claim 4,
The preprocessor is a deep learning-based prostate cancer recurrence prediction system using time-series examination data, characterized in that when a plurality of PSA values exist within each period according to the cycle, the median value of the values within the period is set as a representative value.
5. The method of claim 4,
When the PSA value is missing within the period, the preprocessor applies a weighted moving average method to the PSA value of the period before and after the period to estimate the representative value of the period. Prostate cancer recurrence prediction system.
7. The method of claim 6,
The preprocessor estimates the missing representative value by using only the representative values of the two periods before and after as much as possible from the missing representative values, but the weight is 0.5 to the representative value of the previous period and 0.5 to the representative value of the subsequent period, and two Deep learning-based prostate using time series examination data, characterized in that 0.3 is given to the closest representative value, 0.2 to the next closest representative value, and 0.5 is given if there is only a representative value of one period in the previous or subsequent period Cancer recurrence prediction system.
According to claim 1,
A deep learning-based prostate cancer recurrence prediction system using time series screening data, characterized in that the time series data is further included in the PSA value and treatment information is included.

Images