Avaliação de dose absorvida em procedimentos de radiologia intervencionista usando simulação de Monte Carlo para construção de modelos de aprendizado de máquina

• Ano de defesa: 2024
• Autor(a) principal: Tobias, Isabella Pereira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Biomédica
• Palavras-chave em Português: Radiologia Intervencionista; Simulação Monte Carlo; Aprendizado de Máquina; Interventional Radiology; Monte Carlo simulation; Machine Learning; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA::ENGENHARIA MEDICA; Engenharia Biomédica; Radiologia intervencionista; Radiação - Dosimetria; Radiação - Medidas de segurança; ODS::ODS 3. Saúde e bem-estar - Assegurar uma vida saudável e promover o bem-estar para todos, em todas as idades.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/44901; http://doi.org/10.14393/ufu.te.2024.797

Medical teams spend long periods of time exposed to ionizing radiation arising from x-ray beams used during interventional radiology (IR) procedures. Long exposure during the workday can lead to health problems, especially if the recommended personal protective equipment is not being used correctly or under adequate conditions. In Brazil, the number of IR procedures has grown and it is possible to observe a greater use of ionizing radiation for medical treatment. Thus, there is a need to develop procedures that assist and increase the protection of the medical team that works directly in these interventions. This work investigates the application of machine learning models in the prediction of absorbed radiation doses using Monte Carlo simulations based on hemodynamic rooms through the MCNP6.2 code, reflecting real operating conditions. The general objective of this study is to develop a method for applying machine learning models to estimate the absorbed dose values in the organs of the medical team during IR procedures. The results showed that the analyzed models presented varied performances. Performance metrics, such as Mean Squared Error (MSE) and Mean Absolute Error (MAE), indicated that decision tree-based models were the most accurate among the tested models. Validation of model predictions against simulated MCNP data suggests that machine learning tools are effective in analyzing and predicting doses. This study not only contributes to the understanding of the use of predictive models in radiology, but also highlights the clinical relevance of simulations in hemodynamic settings, aiming to increase the safety of patients and healthcare professionals. Future work could expand this research by integrating additional data and exploring practical techniques to further improve prediction accuracy.