Implementation of a deep-learning based tool for automatic Cardiac MR planning: DeepCardioPlanner

Pedro Louro Costa Osório

Key information

Authors:

Pedro Louro Costa Osório (Pedro Louro Costa Osório)

Supervisors:

Teresa Matias Correia; Rita Homem de Gouveia Costanzo Nunes (Rita Homem de Gouveia Costanzo Nunes)

Published in

07/10/2023

Abstract

A Ressonância Magnética Cardíaca (RMC) é uma técnica poderosa para exames cardíacos abrangentes, mas está limitada a centros especializados devido à necessidade de operadores altamente treinados. Foram propostos métodos de aprendizagem profunda (DL) para automatizar o planeamento cardíaco, e um estudo recente apresenta o DeepCardioPlanner, um conjunto de quatro modelos de redes neurais convolucionais profundas (CNN). Treinados usando uma abordagem de aprendizagem multitarefa, esses modelos preveem a posição e a orientação de cada plano de visão cardíaca a partir de uma varredura 3D adquirida rapidamente. O estudo discute os desafios enfrentados durante o desenvolvimento da ferramenta, incluindo particularidades do conjunto de dados, arquitecturas de rede, estratégias de ponderação de perdas e afinação de hiperparâmetros. O teste do DeepCardioPlanner em dados de CMR de pacientes revela métricas de erro comparáveis às da literatura e corresponde ao intervalo de variabilidade entre operadores. Ao automatizar o planeamento de vistas cardíacas, o DeepCardioPlanner mostra potencial para reduzir a dependência de operadores altamente treinados e melhorar a eficiência nos exames de CMR. Cardiac Magnetic Resonance (CMR) is a powerful technique used for comprehensive cardiac examinations. However, its adoption is often limited to specialized centers due to the complexity involved in determining the four standard cardiac planes. Deep Learning (DL) methods have emerged as a potential solution for automating cardiac planning. In a recent study, researchers proposed DeepCardioPlanner, a set of four deep convolutional neural network (CNN) models trained through multi-task learning. These models can predict the position and orientation of each cardiac view plane from a rapidly acquired 3D scan. The development of DeepCardioPlanner involved overcoming various challenges, including handling dataset particularities, exploring different network architectures, optimizing loss weighting strategies, and tuning hyperparameters. To evaluate the performance of DeepCardioPlanner, the models were tested on real patient CMR data. The results showed that the error metrics obtained from the test set were comparable to those reported in the existing literature. Notably, they matched the range of values for inter-operator variability, demonstrating the effectiveness of DeepCardioPlanner in automating the planning of cardiac views. By automating the planning process, DeepCardioPlanner has the potential to reduce the reliance on highly trained operators, making CMR examinations more accessible and efficient. This advancement in DL-based automation could facilitate broader adoption of CMR and contribute to improved cardiovascular care.