TWI681407B - Computer-aided recognition system, its method and its computer program product thereof - Google Patents

Abstract

A computer-aided prediction system for predicting a diabetes mellitus patient’s onset possibility of CRC is provided. The system includes a deep neural network (DNN) model used to analyze a plurality of baseline characteristics from the diabetes mellitus patient by a deep neural path. The DNN model includes a plurality of neurons and an output layer. At least a part of the neurons corresponds to the baseline characteristics. The output layer outputs an output result related to the onset possibility of CRC.

Description

Computer-aided prediction system, method and computer program product

The invention belongs to the field of computer-aided prediction technology, in particular to the field of computer-aided prediction technology for predicting the possibility of rectal cancer.

Diabetes is one of the most common diseases, and the number of patients worldwide is increasing year by year. For diabetics, good medical care will improve survival. However, recent studies have found that people with diabetes have a higher risk of developing rectal cancer. Once the disease occurs, it will affect the quality of medical care for patients with diabetes and reduce the survival rate significantly. In this regard, if we can predict the cancer risk of diabetic patients early, we can give appropriate medical care, which can improve the survival rate of patients. At present, although there are some technologies that can predict the complications of diabetics, such as the Adapted Diabetes Complication Severity Index (aDCSI), its accuracy still does not meet expectations. It can be seen that there is still an urgent need for a technology that can accurately predict the possibility of diabetic patients suffering from rectal cancer

The present invention proposes a computer-aided prediction technology, which is based on a deep neural network and combined with pathological factor data of a large number of diabetic patients with or without rectal cancer. The basic model of the neural network is trained. When the training is completed, the deep neural network can accurately predict the possibility of diabetic patients suffering from rectal cancer.

According to one aspect of the present invention, a computer-aided prediction system is proposed to predict the possibility of diabetic patients suffering from rectal cancer. The system includes a deep neural network model to analyze the characteristics of multiple pathological factors in diabetic patients through a deep neuron path. The deep neural network model includes a plurality of neurons and output layers. At least some neurons correspond to pathological factors. The output layer outputs the output result related to the possibility of suffering based on this feature analysis. Among them, the DNN model determines the weight value corresponding to each neuron through a plurality of trainings, and then establishes a deep neuron path.

According to another aspect of the present invention, there is provided a computer-aided prediction method for predicting the possibility of diabetic patients suffering from rectal cancer. The method is implemented by a computer-aided prediction system, wherein the computer-aided prediction system includes a plurality of neurons And the DNN model of the output layer. The method includes the steps of: obtaining a plurality of pathological factor data of a diabetic patient; performing feature analysis on the pathological factor data through a deep neuron path through a DNN model; and outputting the possibility of cancer development based on the feature analysis through an output layer Sex-related output results; among them, the DNN model determines the weight value corresponding to each neuron through multiple trainings, and then establishes a deep neuron path.

According to yet another aspect of the present invention, it is to provide a computer program product, which is stored in a non-transitory computer readable medium for the operation of a computer-aided prediction system, wherein the computer-aided prediction system is used to predict the occurrence of diabetes The possibility of rectal cancer, and includes a DNN model with multiple neurons and output layers. The computer program product includes: an instruction to obtain a plurality of pathological factor data of diabetic patients; an instruction to enable the DNN model to perform a feature analysis on the pathological factor data through the deep neuron path; and an instruction to enable the output layer to output based on the feature analysis Loss related to cancer risk Outcome; Among them, the DNN model determines the weight value corresponding to each neuron through multiple trainings, thereby establishing a deep neuron path.

1‧‧‧ Computer-aided prediction system

10‧‧‧Deep Neural Network (DNN) Model 10

12‧‧‧ input layer

13‧‧‧Basic neuron

14‧‧‧ hidden nerve layer

141‧‧‧The first hidden nerve layer

142‧‧‧The second hidden nerve layer

143‧‧‧The third hidden nerve layer

15‧‧‧ hidden neurons

16‧‧‧Output layer

17‧‧‧ output neuron

18‧‧‧ Deep Neural Path

20‧‧‧Data access interface

30‧‧‧Computer program products

40‧‧‧Test module

S31~S35‧‧‧Step

S41~S48‧‧‧Step

1(A) is a system architecture diagram of a computer-aided prediction system for rectal cancer according to an embodiment of the invention; FIG. 1(B) is an architecture diagram of a DNN model according to an embodiment of the invention; FIG. 2 is a DNN according to an embodiment of the invention Schematic diagram of the detailed architecture of the model (completed training); FIG. 3 is a flowchart of basic steps of a computer-aided prediction method according to an embodiment of the present invention; FIG. 4 is a flowchart of steps of a process of establishing a DNN model according to an embodiment of the present invention; 5 is a schematic diagram of experimental data according to an embodiment of the invention.

The following description will provide various embodiments of the present invention. Understandably, these embodiments are not intended to be limiting. The features of the embodiments of the present invention can be modified, replaced, combined, separated, and designed to be applied to other embodiments.

FIG. 1(A) is a system architecture diagram of a computer-aided prediction system 1 according to an embodiment of the invention. As shown in FIG. 1, the computer-aided prediction system 1 includes a deep neural network model 10 (Deep Neural Network, hereinafter referred to as DNN model 10) for predicting the probability of diabetic patients suffering from rectal cancer. In one embodiment, the computer-aided prediction system 1 can further include a data acquisition interface 20. The data obtaining interface 20 is used to obtain data from outside, that is, the user (for example, a physician) can use the data obtaining interface 20 to The data is input into the computer-aided prediction system 1. In addition, in an embodiment, the computer-aided prediction system 1 may further include a test module 40 for testing the prediction ability of the DNN model 10.

FIG. 1(B) is an architecture diagram of a DNN model 10 according to an embodiment of the present invention. Please also refer to FIG. 1(A). The DNN model 10 includes an input layer 12, a plurality of hidden neural layers 14, and an output layer 16. The input layer 12 has a plurality of basic neurons 13, wherein each basic neuron 13 corresponds to a pathological factor data and a weight value. The hidden neural layers 14 each have a plurality of hidden neurons 15, wherein the hidden neurons 15 are connected to the basic neurons 13, and each also corresponds to a weight value. The output layer 16 includes two output neurons 17 for generating two output results, wherein each of these output results corresponds to the cancer risk and non-cancer risk of diabetic patients. In one embodiment, the basic neuron 13, the hidden neuron 15 and the output neuron 17 can form a deep neuron path 18, and the DNN model 10 can conduct a plurality of pathological factor data of diabetic patients through the deep neuron path 18 According to a feature analysis, the output layer 16 can output the output results (probability of cancer and non-cancer) according to the feature analysis. In more detail, when the DNN model 10 obtains data of a plurality of pathological factors of diabetic patients, the pathological factors of diabetic patients can be input into the deep neuron path 18, and each basic element on the deep neuron path 18 can be used Neuron 13, hidden neuron 15 and output neuron 17 perform characteristic analysis on these pathological factors of diabetic patients. In one embodiment, "feature analysis" can be regarded as an operation performed by each neuron on a pathological factor, and "operation" may include a summation operation, an excitation operation, a weighted operation, or a combination of at least two of them , And not limited to this. With this, the DNN model 1 of the present invention can accurately predict the possibility of the diabetic patient suffering from cancer. Next, the details of each element will be explained.

The computer-aided prediction system 1 can be a data processing device, which can be implemented by any device with a microprocessor, such as a desktop computer, a notebook computer, a smart mobile device, a server, or a cloud host and the like. In one embodiment, the computer-aided prediction system 1 may be Communication function to transmit data through the network, where the network communication can be a wired network or a wireless network, so the computer-aided prediction system 1 can also obtain data through the network. In one embodiment, the computer-aided prediction system 1 can execute a computer program product 30 in the microprocessor to achieve its function, wherein the computer program product 30 can have a plurality of instructions, which can cause the processor to perform special operations, and The processor is enabled to implement functions such as the DNN model 10 or the test module 40. In one embodiment, the computer program product 30 may be stored in a non-transitory computer readable medium (such as memory), but is not limited thereto. In one embodiment, the computer program product 30 may also be pre-stored in a network server for users to download.

In one embodiment, the data acquisition interface 20 may be a physical port for acquiring external data. For example, when the computer-aided prediction system 1 is a computer, the data acquisition interface 20 may be a USB interface on the computer, various transmission line connectors, etc. , But not limited. In addition, the data acquisition interface 20 can also be integrated with a wireless communication chip, so that data can be received by wireless transmission.

The DNN model 10 of the present invention is an artificial intelligence model for data analysis, which uses a plurality of operation nodes as neurons of a neural network, and the operation of each neuron can be regarded as a feature analysis of pathological factors. After training with a large amount of data, the DNN model 10 can construct the weight value corresponding to each neuron. In one embodiment, before training, the basic model of the DNN model 10 (that is, the untrained basic architecture) can be established in advance, for example, the number of neurons, the number of hidden neural layers 14, and the number of neurons are preset. Link, etc., and the system 1 uses the instructions in the computer program product 30 to train the untrained DNN model 10 to determine the weight value of each basic neuron 13 and hidden neuron 15 to further establish a deep neuron path 18. In an embodiment, the basic model can undergo multiple trainings to generate multiple neuron paths, and the accuracy of each neuron path can be tested through the test module 40. It should be noted that, in order to distinguish the DNN model 10 before and after training, the untrained DNN model 10 is hereinafter referred to as the basic model, and after the training is completed, the DNN model 10 is referred to.

FIG. 2 is a detailed schematic diagram of the DNN model 10 (training completed) according to an embodiment of the present invention. Please refer to FIGS. 1(A) and 1(B) at the same time. In order to accurately predict the possibility of diabetic patients suffering from rectal cancer, the number of hidden neural layers 14, the number of basic neurons 13 and the number of hidden neurons 15 of the DNN model 10 (or basic model) of the present invention can be regarded as variable parameter. In the embodiment of FIG. 2, the input layer 12 may have 37 basic neurons 13, that is, the DNN model 10 uses 37 pathological factors as the basis for feature analysis. In addition, the DNN model 10 may have 3 hidden neural layers 14, and each hidden neural layer 14 includes 30 hidden neurons 15 each. As shown in FIG. 2, the input layer 12 is connected to the first hidden nerve layer 141, the first hidden nerve layer 142 is connected to the second hidden nerve layer 142, and the second hidden nerve layer 142 is connected to the third hidden nerve layer Layer 143, the third hidden neural layer 143 is connected to the output layer 16, therefore, when 37 pathological factor data of a patient are input to the DNN model 10, it will be analyzed at the input layer 12 first, and then enter the hidden nerves in order The layers 141 to 143 perform analysis, and then the output layer 16 generates an output result 17 according to the analysis result. The above "analysis" refers to the "operation" performed by each neuron on the received data. In one embodiment, when the data passes through a neuron, it can be regarded as the execution of one operation.

In an embodiment, each basic neuron 13 in the input layer 12 is connected to each hidden neuron in the first hidden neuron layer 141, that is, the calculation result of each basic neuron 13 is transmitted to Each hidden neuron 15 of the first hidden neural layer 141. Each hidden neuron 15 of the first hidden neural layer 141 is connected to each hidden neuron 15 of the second hidden neural layer 142, that is, the operation result of each hidden neuron 15 of the first hidden neural layer 141 Each hidden neuron will be transmitted to the second hidden neural layer 142. Each hidden neuron 15 in the second hidden neural layer 142 is connected to each hidden neuron 15 in the third hidden neural layer 143, that is, the operation of each hidden neuron 15 in the second hidden neural layer 142 The result is transmitted to each hidden neuron 15 in the third hidden neural layer 143. Each hidden neuron 15 in the third hidden neuron layer 143 is connected to each output neuron 17 in the output layer 16, that is, The operation result of each hidden neuron 15 of the third hidden neural layer 143 is transmitted to each output neuron 17. After the calculation of the output neuron 17, the output layer 16 can generate the cancer probability and the cancer probability of the patient.

Next, the details of the calculation process will be explained. As shown in FIG. 2, in one embodiment, when 37 pathological factor data enter the input layer 12, each pathological factor data will be individually weighted with the corresponding weight value (that is, multiplied by the weight value) Then, all the weighted data are sent to each hidden neuron 15 in the first hidden neural layer 141 again. In an embodiment, for each hidden neuron 15 of each hidden neural layer 141 to 143, it will first perform a sum operation on the received data, and then perform a summation on the result of the sum operation One type of excitation operation, and then the result of the first type of excitation operation and the weight value corresponding to the hidden neuron 15 are weighted, and the weighted operation result of each hidden neuron 15 will enter the next one The neural layer 14 or the output layer 16 is hidden. In an embodiment, for each output neuron 17 of the output layer 16, the obtained data will be first summed up, and then a second type excitation operation will be performed, and after the second type excitation operation The result will form a probability value.

In one embodiment, the first type excitation operation and the second type excitation operation may be different. In one embodiment, the first excitation operation is defined as using a linear rectification function (Rectified Linear Unit, ReLU) as the activation function (Activation function) to perform the operation. In one embodiment, the second excitation operation is defined as using the Softmax function as the excitation function to perform the operation. Since the output interval of the ReLU function is 0 to infinity, it is suitable as an exciter for neurons in the middle part of the neural-like network, and because the output interval of the Softmax function is 0 to 1, it is suitable as the output of the neural-like network For example, the exciter of the neuron at the end can make the output result into a probability. It should be noted that the present invention is not limited to this, that is, the present invention can also use other excitation functions to perform excitation operations.

In this way, after the DNN model 10 completes training, as long as 37 pathological factors of a patient are input into the DNN model 10, the DNN model 10 can predict the possibility of the patient suffering from rectal cancer. In an embodiment, the pathological factor data can be first normalized or standardized to form a value under the same standard. For example, each pathological factor can be normalized or standardized and converted into a score, and these scores can be Weighted calculations are carried out via neurons, and eventually the cancer and non-cancer probabilities are formed.

In addition, for the DNN model 10, the data source of neurons may affect the predictive ability of the DNN model 10. In one embodiment, the patient's 37 pathological factor data may include physiological data (Biographical), comorbidity data (Comorbidities), diabetes complication data (Diabetes Complications), therapeutic drug data (Medications) and index data (Scoring System ), but not limited to this.

In an embodiment, the "physiological data" may include age, gender, low urbanization (Lowest Urbanization), medium urbanization (Medium Urbanization), high urbanization (High Urbanization), highest urbanization (Highest Urbanization), white-collar workers Class (White Collar Occupation), Blue Collar Class (Blue Collar Occupation) and other professional classes (Other Occupation) and other information, but not limited to this. In one embodiment, the "comorbidity data" may include hypertension, hyperlipidemia, stroke, congestive heart failure, colorectal polyps, obesity, COPD, CAD, asthma, smoking, inflammatory bowel disease, intestinal susceptibility Irritation syndrome, CKD and alcohol-related diseases, but not limited to this. In one embodiment, the "diabetic complications data" may include retinopathy, nephropathy, neuropathy, cerebrovascular, cardiovascular and metabolic information, but is not limited thereto. In one embodiment, "therapeutic drug information" may include metformin (Metformin), statin (Statin), insulin (Insulin), sulfonylurea (Sulfonylureas), other antidiabetic drugs (Other antidiabetic drugs), TZD and PVD and other information, But it is not limited to this. In one embodiment, the "index data" may include aDCSI Index information, but it is not limited thereto.

In an embodiment, each pathological factor data can be quantified into a corresponding score, where the way of quantification can be different according to the nature of the data, for example, certain features can be corresponding according to "the presence or absence of features" Different scores (for example, different genders will correspond to different scores, whether the use of drugs will correspond to different scores, etc.), and some features themselves can be divided into multiple grades, and through the grades are corresponding to different scores (such as 25 years old It can correspond to a score, 30 years old can correspond to another score, etc.); the above content is just an example, the invention is not limited to this.

Next, the basic operation mode of the computer-aided prediction system 1 will be explained. FIG. 3 is a flowchart of basic steps of a computer-aided prediction method according to an embodiment of the present invention. The method is executed by the computer-aided prediction system 1 of FIG. 1(A), in which the DNN model 10 is in a trained state. Refer to FIG. 1(A) to FIG. 3. As shown in FIG. 3, first, step S31 is executed, and the data acquisition interface 20 acquires pathological factor data of a diabetic patient. After that, step S32 is executed, the input layer 12 of the DNN model 10 obtains the pathological factor data, and transmits the weighted pathological factor data to the hidden neural layer 14. After that, step S33 is executed. Each hidden neuron of each hidden neural layer 14 performs an operation on the received data, and the last hidden neural layer 14 transmits the result of the operation to the output layer 16. After that, step S34 is executed, and the output layer 16 performs calculation on the received data, and further outputs the cancer probability and non-cancer probability of the patient. After that, step S35 is executed, and the system 1 predicts the possibility of cancer of the patient based on the output result of the output layer 16.

Regarding step S31, in one embodiment, the pathological factor data may be the aforementioned 37 pathological factors, and has been formed into a score by normalization or standard deviation operation.

Regarding step S32, in an embodiment, the weight value corresponding to each pathological factor (the weight value of the basic neuron 13) has been determined during the training process of the DNN model 10 (basic model), in other words In short, one of the training objectives of the DNN model 10 is to determine the weight value corresponding to each pathological factor. In an embodiment, the score of each pathological factor is weighted in the basic neuron 13 and then transmitted to each hidden neuron 15 in the first hidden neuron layer 14.

Regarding step S33, in an embodiment, each hidden neuron 15 of each hidden neural layer 14 performs summation operation, excitation operation (first type excitation operation) and weighted operation on the received data, where each The weight value corresponding to a hidden neuron 15 is also determined during the training process of the DNN model 10 (basic model), that is, one of the training purposes of the DNN model 10 is to determine the weight value corresponding to each hidden neuron Why. In an embodiment, the weighted operation result of each hidden neuron 15 in the last hidden neuron layer 14 will be transmitted to each output neuron 17 in the output layer 16.

Regarding step S34, in an embodiment, each output neuron of the output layer 16 performs summation and excitation operations on the received data (second type excitation operation). In one embodiment, the sum of the two output results of the output layer 16 is 1 (that is, the summed result corresponds to a predicted probability of 100%).

Regarding step S35, in one embodiment, the system 1 compares the two output results of the output layer 16, and uses a higher probability as the prediction result. For example, when the output result corresponding to the probability of not suffering from cancer is 0.75 (that is, 75%), and when the output result corresponding to the cancer probability is 0.25 (that is, 25%), the system 1 predicts that the patient has a low possibility of cancer, but the invention is not limited to this.

It can be seen that after the establishment of the DNN model 10, as long as the patient's pathological factor data is input into the computer-aided prediction system 1, the DNN model 10 can predict the possibility of the patient's cancer, thereby allowing the patient to prevent it earlier , The survival probability can be greatly improved.

In addition, in order for the DNN model 10 to perform steps S31 to S35, the DNN model 10 must first establish the weight value of each neuron through training. The establishment process of the DNN model 10 will be described in detail below.

FIG. 4 is a flow chart of the steps of the process of establishing the DNN model 10 according to an embodiment of the present invention. These steps can be implemented by the processor of the computer-aided prediction system 1 executing instructions in the computer program product 30. Please also refer to FIG. 1 To Figure 4.

First, step S41 is executed, and the basic model of the DNN model 10 is set. After that, step S42 is executed, and the input layer 12 obtains a minimum batch of data from all training data. After that, step S43 is executed, and the basic model uses the acquired data for training to determine the weight value of each neuron in the basic model, thereby establishing a candidate neuron path. After that, step S44 is executed, and the input layer 12 of the basic model obtains a minimum number of additional plural training data, and executes step S43 again. After that, step S45 is executed, and step S44 is repeatedly executed until a preset condition is reached. After that, step S46 is executed, and steps S42 to S45 are repeated a plurality of times (itoration procedure). After that, step S47 is executed, and the prediction model 40 evaluates the prediction ability of each candidate neuron path. After that, step S48 is executed, and the system 1 sets the candidate neuron path with the best prediction ability as the deep neuron path 18 actually used by the DNN model 10. The above steps can be implemented at least by the processor of the system 1 executing instructions of the computer program product 30 or instructions of other computer program products.

Regarding step S41, this step is used to find the optimal variable parameter of the DNN model 10 (basic model), where the variable parameter may be, for example, the number of hidden nerve layers, what is the excitation function, etc., and is not limited thereto. This step can be realized by the system 1 receiving the command input by the user, and setting the basic model according to the command. In one embodiment, this step is to use a few training data to first create a plurality of simplified basic models with different parameters, and then use the K-fold cross validation method (K-fold cross validation) to find one of the best performance The simplified basic model is set as the basic model of the DNN model 10 (that is, the optimal parameter value can be found). Here, "a few training data" may be, for example, 1/100 of all training data, but it is not limited thereto. In one embodiment, K-fold cross Validation is to perform K verifications for each simplified basic model. Each verification includes a training process and a testing process, where the training process is to determine the weight value of each neuron of the simplified basic model, and the testing process is to test the simplified basic The predictive power of the model. For a simplified basic model, each verification is to divide the aforementioned few training materials into (K-1): 1 into a training group and a test group, where the training group is used for the training process and the test group is used for Testing process. When K times of verification are completed, the system 1 takes the average of the accuracy of each verification of the simplified basic model again, and uses the average as the accuracy of the simplified basic model. In an embodiment, K is 10, that is, each simplified basic model will be verified 10 times, and each verification is to divide the training data into a training group and a test group by 9:1, but the present invention does not Limited to this. In addition, if the DNN model 10 needs to use best hyperparameters and optimizers, the best hyperparameters and optimizers can also be set in step S41. In this way, step S41 can find a simplified basic model with the best parameters as a basic model for subsequent deep training.

Regarding step S42, the system 1 can first obtain all the training data, and extract the minimum batch of data from all the training data to input into the basic model, so that the basic model uses these minimum batches of data for deep training (the first time Deep training). In one embodiment, the amount of all training data is defined as at least one million records, and the minimum batch size is defined as at least 100 records. In one embodiment, the total training data is 1315899, and the minimum batch is 128 data, so the system 1 will randomly select 128 data from the 1315899 training data and enter it into the basic model, but the invention is not limited to this . It should be noted that each training data contains information on 37 pathological factors of a diabetic patient and whether the patient actually has cancer or not.

Regarding step S43, this step is for the basic model to use the data obtained in step S42 for training. Since the training data contains information on whether the diabetic patient actually suffered from cancer, the basic model This model can analyze the characteristics of possible pathological factors in the case of cancer and the characteristics of possible pathological factors in the case of no cancer, and then determine the weight value of each neuron. In one embodiment, the basic model performs gradient descent algorithm for training, and then determines the weight value of each neuron. In one embodiment, the gradient descent algorithm may be at least one of Stochastic gradient descent or Adam with Nesterov's accelerated gradient descent, and is not limited thereto. One of the purposes of using Stochastic gradient descent is to reduce the difference between the predicted value of the basic model and the actual result (loss), for example, to have a local minimum difference between the predicted result of the basic model and the real result, and adopt Adam with Nesterov's One of the purposes of accelerated gradient descent is to increase the probability of the absolute minimum difference between the predicted results of the basic model and the real results. In addition, since one of the key points of the present invention is to improve the accuracy of the prediction ability of the basic model through the characteristics of Stochastic gradient descent and Adam with Nesterov's accelerated gradient descent, the implementation process of Stochastic gradient descent and Adam with Nesterov's accelerated gradient descent It is not the focus, so the implementation process will not be detailed here. After step S43 is completed, the basic model can complete one training and a candidate neuron network can be established.

Regarding step S44, the system 1 will input another set of the smallest batch of data (that is, another 128 pieces of data) into the basic model, and the basic model uses the group of the smallest batch of data to re-train in step S43, and thereby generate another A candidate neural network. In an embodiment, the smallest batch of data selected by the system each time is randomly selected. Therefore, duplicate data may be selected for each group of smallest batch of data, but it is not limited.

Regarding step S45, the system 1 repeatedly executes step S41 to generate a plurality of neuron networks of the basic model until a predetermined condition is fulfilled. In one embodiment, "preset condition" means that all training data have been input into the basic model, and the basic model has been trained; In another embodiment, the "preset condition" may also be that a specified number of neuron networks have been established. The description of "preset conditions" is only an example, and the present invention is not limited thereto.

Regarding step S46, this step is used to perform an iteration procedure on the training of the basic model, that is, steps S42 to S45 are re-executed until the specified number of times is reached, thereby further improving the prediction ability of the neural network. In one embodiment, the "specified number of times" is set to at least 1000 times, but it is not limited.

Regarding step S47, this step is to evaluate the performance of each candidate neuron network through the test module 40. In an embodiment. The test of the test module 40 may include weighted average recall analysis to analyze the sensitivity of the candidate neuron networks. In an embodiment, the test of the test module 40 may include positive predictive value analysis to analyze the accuracy of the candidate neuron networks. In an embodiment, the test of the test module 40 may include F1 analysis to analyze the F1 values of the candidate neural network (ie, the harmonic average of sensitivity and accuracy). In one embodiment, at least two of weighted average recall, positive predictive value, and F1 analysis are performed together. In this way, the prediction performance of each candidate neuron network can be evaluated.

Regarding step S48, this step is to select the candidate neuron network with the best prediction performance as the deep neuron network 18 of the DNN model 10 actually used. When step S48 is completed, the training of the DNN model 10 is completed. In the future, the user (physician) only needs to input the pathological factor data of the patient to the DNN model 10, and the DNN model 10 can analyze the possibility of the patient suffering from cancer.

In addition, in an embodiment, during the training process, each hidden neural layer 14 and output layer 16 may be applied with a dropout (ie, a regularization technique to avoid overfitting). In an embodiment, the output layer 16 may use a categorical cross entropy function as a loss function. In addition, in an embodiment, the weight value of each neuron may be initialized using a normalized He initialization value (Normalized He initialization). The present invention is not limited to this.

5 is a schematic diagram of experimental data according to an embodiment of the present invention. The ROC curve is used to present the accuracy of the DNN model 10 of the present invention and the traditional aDCSI model in predicting the cancer risk of diabetic patients. The Y-axis is the true positive rate. (Marked with True positive rate), the X axis is the false positive rate (marked with False positive rate), where both are tested with the same data. As shown in FIG. 5, the area under the curve (AUC) of the ROC curve of the DNN model 10 is about 0.738, and the AUC of the aDCSI model is about 0.492. It can be seen that the DNN model 10 of the present invention has better performance than the traditional aDCSI model Predictive power.

In this way, the DNN model used in the present invention can be established. In other words, as long as the pathological factor data of the patient is input into the DNN model, the DNN model can automatically predict the possibility of the patient suffering from rectal cancer. Through deep learning training, the computer-aided prediction system of the present invention can accurately predict the cancer incidence of patients, and can assist patients in seeking the best medical care.

In addition, in an embodiment, the computer-aided prediction system, method and computer program product of the present invention can be described in the paper "Development of a Prediction Model for Colorectal Cancer among Patients with Type 2 Diabetes Mellitus Using a Deep Neural Network, Meng-Hsuen Hsieh, "Li-Min Sun, Cheng-Li Lin, Meng-Ju Hsieh, Kyle Sun, Chung-Y. Hsu, An-Kuo Chou, and Chia-Hung Kao" to achieve, but not limited to.

Although the present invention has been described through the above embodiments, it is understandable that many modifications and changes are possible in accordance with the spirit of the present invention and the patent application scope claimed by the present invention.

1‧‧‧ Computer-aided prediction system

10‧‧‧Deep Neural Network (DNN) Model 10

20‧‧‧Data access interface

30‧‧‧Computer program products

40‧‧‧Test module

Claims (10)

A computer-aided prediction system for predicting the possibility of a diabetic patient suffering from rectal cancer, including: a deep neural network (DNN) model, through a deep neuron path to a feature of a plurality of pathological factor data of the diabetic patient Analysis, wherein the DNN model includes: a plurality of neurons, at least a part of which correspond to the pathological factor data; and an output layer, which outputs at least one output result related to the possibility of cancer according to the feature analysis; Among them, the DNN model determines a weight value corresponding to each neuron through a plurality of trainings, and then establishes the deep neuron path; wherein, the pathological factor data includes physiological data, comorbidity data, and diabetes complications Data, therapeutic drug data and index data. The computer-aided system according to claim 1, wherein the DNN model further includes an input layer and a plurality of hidden neural layers, wherein the input layer includes a plurality of basic neurons corresponding to the pathological factor data, each of the hidden neural layers It contains a plurality of hidden neurons, and each hidden neuron in one hidden neuron layer is connected to all the basic neurons. The computer-aided system according to claim 2, wherein the number of hidden neural layers is determined by performing a k-fold cross-validation method, where k is set to 10. The computer-aided system as described in claim 2, wherein the DNN model performs each training by randomly selecting a plurality of training data, and creates a plurality of candidate deep neural networks, where each training is defined as equivalent The training data performs at least one gradient descent operation to determine each basic god The weight value corresponding to the meridian and each hidden neuron, and each training data contains information on 37 pathological factors of a diabetic patient and whether the diabetic patient suffers from rectal cancer. The computer-aided system according to claim 4, wherein the deep neural network when the DNN model is actually used is one of the candidate deep neural networks with the best predictive ability, and the deep neural network The prediction ability is determined by at least one of weighted average recall, positive predicted value and F1 analysis. A computer-aided prediction method for predicting the possibility of a diabetic patient suffering from rectal cancer. The method is implemented by a computer-aided prediction system, wherein the computer-aided prediction system includes a DNN model, and the DNN model includes a plurality of nerves And an output layer, the method includes the steps of: obtaining a plurality of pathological factor data of the diabetic patient; performing a feature analysis on the pathological factor data through a deep neuron path through the DNN model; and through the output Layer, according to the feature analysis, output at least one output result related to the possibility of cancer; wherein, the DNN model determines a weight value of each basic neuron through a plurality of trainings, and then establishes the deep neuron path; Among them, the pathological factor data includes physiological data, comorbidity data, diabetes complication data, treatment drug data and index data. The computer-aided prediction method according to claim 6, wherein the DNN model further includes an input layer and a plurality of hidden neural layers, wherein the input layer includes a plurality of basic neurons corresponding to the pathological factor data, and the hidden neural layers Each contains a plurality of hidden neurons, and each hidden neuron in one hidden neural layer is connected to all basic neurons. The computer-aided prediction method as described in claim 7, wherein the DNN model performs each training by randomly selecting a plurality of training data, and creates a plurality of candidate deep neural networks, Each training is defined as performing at least one gradient descent operation on the peer training data to determine the weight value corresponding to each basic neuron and each hidden neuron, and each training data contains 37 diabetic patients Pathological factor information and whether the diabetic patient suffers from rectal cancer. The computer-aided system according to claim 8, wherein the deep neural network when the DNN model is actually used is one of the candidate deep neural networks that has the best predictive ability, and the deep neural network has The prediction ability is determined by at least one of weighted average recall analysis, positive predictive value analysis and F1 analysis. A computer program product, stored in a non-transitory computer-readable medium, for operating a computer-aided prediction system, wherein the computer-aided prediction system is used to predict the likelihood of a diabetic patient suffering from rectal cancer, And the computer-aided prediction system includes a DNN model with a plurality of basic neurons and an output layer. The computer program product includes: an instruction to obtain the plurality of pathological factor data of the diabetic patient; an instruction to make the DNN model through A deep neuron path performs a feature analysis on the pathological factor data; and an instruction to cause the output layer to output at least one output result related to the possibility of cancer according to the feature analysis; wherein, the DNN model uses complex numbers Training to determine a weight value for each basic neuron, and then establish the deep neuron path; wherein, the pathological factor data includes physiological data, comorbid disease data, diabetes complications data, treatment drug data and index data.

Images