Optimization system and method of AI algorithm for prediction coronary artery lesions based on FFR

1 . An optimization system of an artificial intelligence (AI) algorithm for predicting a lesion in a coronary artery based on a fractional flow reserve (FFR), comprising:

one or more memories; and

one or more processors operably coupled with the one or memories, the one or more processors configured to cause:

collecting preset factor data in order to predict an FFR numerical value;

analyzing a correlation between the factor data and eliminating specific factor data;

performing learning processing using the factor data from which the specific factor data have been eliminated processing unit using a plurality of pre-stored AI algorithms, analyze a learning result, and perform optimization processing of each AI algorithm based on an analysis result,

dividing all of the factor data from which the specific factor data have been eliminated into training data and test data according to a plurality of predetermined ratios to generate a plurality of training data sets; and

inputting each of the plurality of training data sets into each of the plurality of AI algorithms to perform learning processing such that a plurality of learning result models are generated for each AI algorithm.

2 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 1 , wherein the collecting preset factor data includes:

receiving biometric factor data for each patient from the outside;

receiving blood vessel shape factor data generated based on medical image data for each patient from the outside;

generating flow factor data for a cardiovascular region for each patient using the blood vessel shape factor data, and

receiving an FFR value measured or predicted for each patient from the outside.

3 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 2 , wherein the generating the flow factor data includes:

generating a plurality of virtual blood vessel models in advance, performing a computational fluid dynamics (CFD) simulation for the generated virtual blood vessel models, and constructing a database of CFD simulation performing result data for each virtual blood vessel model to store and manage the CFD simulation performing result data; and

deriving performing result data corresponding to the blood vessel shape factor data and generating the performing result data as the flow factor data.

4 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 2 , wherein the analyzing a correlation between the factor data and eliminating specific factor data includes:

constructing a database of the biometric factor data, the blood vessel shape factor data, and the flow factor data by the data collection unit for each patient;

analyzing a correlation between each factor data and the received FFR value by applying a pre-stored technique;

selecting specific factor data of which a correlation is a predetermined reference or less based on an analysis result by applying a pre-stored technique and eliminate all factor data of a corresponding patient including the specific factor data; and

correcting and reconstructing the database by the DB construction unit based on an elimination result.

5 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 4 , wherein the performing learning processing includes:

performing learning processing by inputting the database as the training data to a plurality of pre-stored heterogeneous AI algorithms; and

receiving an FFR prediction result using a learning result model for each AI algorithm and analyzing prediction result accuracy for each learning result model,

wherein the performing learning processing by inputting the database includes dividing all the factor data included in the database into the training data and the test data according to the plurality of predetermined ratios by applying a pre-stored technique to generate the plurality of training data sets, and then inputting each of the plurality of training data sets into each AI algorithm to perform learning processing, and

wherein the receiving an FFR prediction result includes analyzing accuracy of an FFR prediction result output from each learning result model using the test data.

6 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 5 , wherein the receiving an FFR prediction result includes analyzing the accuracy of the FFR prediction result output from each learning result model to derive a specific ratio having the highest FFR prediction accuracy for each AI algorithm.

7 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 6 , wherein the performing learning processing further includes:

controlling a weight for a hyper parameter that determines a property for each of the plurality of pre-stored heterogeneous AI algorithms by applying a pre-stored technique and inputting the training data divided according to the specific ratio having the highest FFR prediction accuracy first to each controlled AI algorithm to perform learning processing; and

receiving an FFR prediction result for each learning result model using the test data divided according to the specific ratio having the highest FFR prediction accuracy and analyzing prediction result accuracy for each learning result model,

wherein the controlling a weight for a hyper parameter includes controlling control-weights for corresponding hyper parameters for each AI algorithm plural times under different conditions, and repeatedly performing learning processing for each controlled AI algorithm.

8 . The optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 7 , wherein the receiving an FFR prediction result includes extracting a learning result model having the highest FFR prediction accuracy for each AI algorithm, and analyzing a weight control condition of a hyper parameter for the corresponding learning result model.

9 . An optimization method of an AI algorithm for predicting a lesion in a coronary artery based on an FFR that uses an optimization system of an AI algorithm for predicting a lesion in a coronary artery based on an FFR performed by a computer, comprising:

collecting preset factor data in order to predict an FFR numerical value;

analyzing a correlation between the factor data and eliminating specific factor data; and

inputting the factor data from which the specific factor data have been eliminated to a plurality of pre-stored AI algorithms to perform learning processing, analyzing a learning result, and performing optimization processing of each AI algorithm based on an analysis result,

wherein the inputting the factor data comprises:

dividing all of the factor data from which the specific factor data have been eliminated into training data and test data according to a plurality of predetermined ratios to generate a plurality of training data sets; and

inputting each of the plurality of training data sets into each of the plurality of AI algorithms to perform learning processing such that a plurality of learning result models are generated for each AI algorithm.

10 . The optimization method of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 9 , wherein the collecting preset factor data includes:

receiving biometric factor data, blood vessel shape factor data, and a measured or predicted FFR value for each patient; and

generating flow factor data for a cardiovascular region for each patient using the blood vessel shape factor data.

11 . The optimization method of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 10 , wherein the generating flow factor data includes:

generating a plurality of virtual blood vessel models in advance, performing a CFD simulation for the generated virtual blood vessel models, and constructing a database of CFD simulation performing result data for each virtual blood vessel model to store and manage the CFD simulation performing result data; and

deriving CFD simulation performing result data corresponding to the blood vessel shape factor data and generating the CFD simulation performing result data as the flow factor data.

12 . The optimization method of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 10 , wherein the analyzing a correlation between the factor data and eliminating specific factor data includes:

constructing a database of the factor data for each patient;

analyzing a correlation between each factor data and the FFR value by applying a pre-stored technique;

selecting specific factor data of which a correlation is a predetermined reference or less based on an analysis result by applying a pre-stored technique and eliminating all factor data of a corresponding patient including the specific factor data; and

correcting and reconstructing the database based on an elimination result.

13 . The optimization method of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 12 , wherein the inputting the factor data includes:

performing learning processing by inputting the database as training data to a plurality of pre-stored heterogeneous AI algorithms, and dividing all the factor data included in the database into the training data and the test data according to the plurality of predetermined ratios by applying a pre-stored technique to generate the plurality of training data sets and then inputting each of the plurality of training data sets to each AI algorithm; and

receiving an FFR prediction result for each learning result model using each test data and the FFR value and analyzing accuracy of each learning result model based on the FFR prediction result, and

a specific ratio having the highest FFR prediction accuracy is derived for each AI algorithm.

14 . The optimization method of an AI algorithm for predicting a lesion in a coronary artery based on an FFR of claim 13 , wherein the inputting the factor data further includes:

controlling weights for hyper parameters that determine a property for each of the plurality of pre-stored heterogeneous AI algorithms plural times under different conditions by applying a pre-stored technique and inputting the training data divided according to the specific ratio having the highest FFR prediction accuracy derived to each controlled AI algorithm to perform learning processing; and

receiving an FFR prediction result for each learning result model using the test data divided according to the specific ratio having the highest FFR prediction accuracy derived and the received FFR and analyzing accuracy of each learning result model based on the FFR prediction result, and

a learning result model having the highest FFR prediction accuracy for each AI algorithm is extracted, and a weight control condition of a hyper parameter for the corresponding learning result model is analyzed.