Automated segmentation of cardiac structures in magnetic resonance imaging via deep convolutional neural networks

该项目的总体目标是设计一种全自动和高效的算法，用于在心脏电影-MRI成像数据中找到几种心脏结构，例如左和右室的腔。计算这样的结构可以转化为心脏形态和运动的综合测量，这在心脏病诊断中具有很高的临床意义。重点将是基于卷积神经网络的最先进的深度学习方法。具体的子目标包括：(1)研究一种新的半监督学习损失函数，该函数嵌入了考虑心脏解剖的几何先验。这种先验的目的是为了说明心脏图像分析中专家注释的训练数据的有限大小；(2)设计、实现和测试考虑心脏结构的特定上下文的卷积神经网络体系结构；以及(3)通过将结果与心脏病学家的地面事实注释进行比较来评估算法。这一Engage研究项目将使Corstein的算法能够与医学图像分析深度学习领域的前沿发展保持一致，并开发最先进的测试实例，有可能增强其产品组合和在快速增长的市场中的竞争力。用尖端算法增强Corstem的产品组合将有助于加拿大作为领先的人工智能和医学成像技术生产商的地位。

The overall objective of this project is to design a fully automated and efficient algorithm for finding severalheart structures in cardiac CINE-MRI imaging data, e.g., the cavities of the left and right ventricles. Computingsuch structures translates into comprehensive measures of heart morphology and motion, which are of highclinical interest in cardiac disease diagnosis. The focus will be on state-of-the-art deep learning approachesbased on convolutional neural networks. Specific sub-objectives include: (1) investigating a novelsemi-supervised learning loss function, which embeds geometric priors accounting for the anatomy of theheart. The purpose of such priors is to account for the limited size of expert-annotated training data in thecontext of cardiac image analysis; (2) designing, implementing and testing a convolutional neural networkarchitecture that accounts for the specific context of cardiac structures; and (3) evaluating the algorithm bycomparing the results to ground-truth annotations by cardiologists. This ENGAGE research project will enableCorstem's algorithms to align with cutting-edge developments in deep learning for medical image analysis, andto develop stat-of-the-art test examples, with the potential of enhancing its product portfolio andcompetitiveness in a fast growing market. Enhancing Corstem's product portfolio with cutting-edge algorithmswill contribute to Canada's stature as a leading AI and medical imaging technology producer.