Imprimir Resumo


Teste
Abstract: 131-1

131-1

Monte Carlo simulation with Machine Learning Models for Predicting Absorbed Doses in Hemodynamic Procedures

Authors:
Isabella P. Tobias (PPGEB/FEELT/UFU - Programa de Pós-graduação em Engenharia Biomédica, Faculdade de Engenharia Elétrica, Universidade Federal de Uberlândia) ; Monique F. Silva (PPGEB/FEELT/UFU - Programa de Pós-graduação em Engenharia Biomédica, Faculdade de Engenharia Elétrica, Universidade Federal de Uberlândia) ; Ana P. Perini (PPGEB/FEELT/UFU - Programa de Pós-graduação em Engenharia Biomédica, Faculdade de Engenharia Elétrica, Universidade Federal de Uberlândia, INFIS/UFU - Instituto de Física, Universidade Federal de Uberlândia) ; William Santos (PPGEB/FEELT/UFU - Programa de Pós-graduação em Engenharia Biomédica, Faculdade de Engenharia Elétrica, Universidade Federal de Uberlândia, UFS - Departamento de Física, Universidade de Sergipe) ; Lucio Pereira Neves (PPGEB/FEELT/UFU - Programa de Pós-graduação em Engenharia Biomédica, Faculdade de Engenharia Elétrica, Universidade Federal de Uberlândia, INFIS/UFU - Instituto de Física, Universidade Federal de Uberlândia)

Abstract:

In hemodynamic procedures, imaging techniques guided by ionizing radiation are widely used to diagnose and treat cardiovascular diseases. Although these procedures provide significant benefits to patients, occupational exposure to radiation is a major concern for the medical team involved [1]. Currently, radiation dose monitoring by medical team is performed using personal dosimeters, these devices do not provide detailed information on absorbed dose specificied by organs, which could allow to assess risks for each organ and implement targeted protective measures if necessary. The main objective of this study is to develop machine learning models using data from Monte Carlo simulations to quantify the absorbed dose during procedures in hemodynamic rooms. The results of these models will be analyzed with the ones obtained directly from Monte Carlo simulations and their the effectiveness compared against each other based on their respective metrics. The dataset utilized in this research originated from Monte Carlo simulations conducted through the MCNP 6.2 software. The scenarios of hemodynamic procedures were simulated with of a hemodynamic room, encompassing essential components, such as x-ray equipment, the patient table, coated walls of Barite concrete Type BA, door and leaded glass window. The professionals were simulated with essential personal protective equipment,  including eye protectors, thyroid shields and lead aprons. The patient was also included in the simulations. The models evaluated include Decision Tree, Random Forest, Linear Regression with Ridge, Elastic Net, and MLP Regressor. Performance metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Relative Error (MRE) were used to assess the models' accuracy. The Decision Tree and Random Forest models emerged as the most robust in terms of error metrics, with consistently low MSE and MAE, indicating a superior ability to accurately predict the absorbed dose to organs. Specifically, the MSE in the order of 10-9 for both models reflects a minimal dispersion between the predictions and the actual values. Comparatively, models such as Linear Regression with Ridge and Elastic Net, although they demonstrated acceptable performance, were shown to be less accurate compared to the decision tree-based approaches. The MLP Regressor, on the other hand, showed an intermediate performance, with an MRE that suggests a greater variability in the predictions compared to the real values [2, 3]. To verify the robustness of the models and ensure that they do not suffer from overfitting, the cross-validation technique was used. The cross-validation results for Decision Tree and Random Forest showed an average MAE of 3.68x10-3 and 3.57x10-3, respectively, with relatively low standard deviations 3.33x10-3 for Decision Tree and 3.24x10-3 for Random Forest. These values indicate that Decision Tree and Random Forest models maintain consistent performance across different divisions of the data, suggesting that overfitting was minimized.

The authors would like to thank the Brazilian agencies CNPq (Grants 312160/2023-2 (L.P.N), 312124/2021-0 (A.P.P), 309675/2021-9 (W.S.S) and 406303/2022-3) and FAPEMIG (Grants  APQ-04215-22, APQ-01254-23 and APQ-04348-23).

Keywords:
 Hemodynamic, Radiation dose monitoring, Monte Carlo simulations, Machine learning models, Personal protective equipment