基于机器学习的MRI液氦水平预测研究OA

Objective:To compare recursive multi-step predictive effects of support vector regression(SVR),long short-term memory(LSTM)net,gated recurrent unit(GRU),and transformer models for the inventory level of liquid Helium of magnetic resonance imaging(MRI)equipment,and to assess the applicability of each model in time series data of single equipment and small sample.Methods:A total of 340 records of inventory level of liquid helium were obtained from one magnetic resonance imaging(MRI)equipment at West China Hospital,Sichuan University,between February 24 and May 29,2024.The records were chronologically divided into training,validation,and test sets at a ratio of 6:2:2,which included respectively 204 records,68 records and 68 records.Four models were trained by using the information of historical liquid helium levels and time-interval,and recursive multi-step prediction was performed on the validation and test sets.The performance of recursive prediction was evaluated by using the coefficient of determination(R2),root mean squared error(RMSE),mean absolute error(MAE),mean absolute percentage error(MAPE),and residual standard deviation.Results:In test set,SVR achieved an R2 of 0.976,and the RMSE,MAE,and residual standard deviation of that were respectively 0.56,0.47 and 0.56,and the MAPE of that was 0.89%,which appeared the best comprehensive performance.The each predictive indicator of GRU was similar to those of SVR.In the validation set,the comprehensive error of LSTM was the lowest.Conclusion:Under the conditions of single equipment and small sample data of this study,SVR can demonstrate the highest prediction accuracy in the test set,however,the fitting superiority of each model existed significant fluctuation at different data phases,such as the validation and test sets.The stability and external applicability of model still need further validation through using multiple equipment,multiple cycles of supplemental liquid,and prospective data.