Synopsis

Deep learning has become an ubiquitous tool for image analysis, including CMR. The method is capable to learn specific human tasks from a large amount of data, provided sufficient computational power. Image segmentation and recognition are the two most used applications, but there are more creative solutions. Deep learning can learn to reconstruct MRI, leading to a faster MR acquisition. It can generate realistic contrast-enhanced MRI without using the actual contrast agent. In this course, we are going to learn how deep learning can be applied to solve CMR image analysis to derive cardiac function and anatomy of the heart.

Syllabus

• Relationship between deep learning, machine learning and artificial intelligence.
• Deep learning is just a neural network with more layers.
• Convolutional neural network (CNN) for image analysis
• The key elements of learning: forward/backward propagation, gradient descent optimization and loss function.

• Automatic MR slice view recognition
• Cardiac landmark detection with feature extraction network
• U-Net architecture for segmenting myocardium (and other chambers)
• Generating high-resolution CMR with deep learning
• 3D shape heart model as prior knowledge in deep learning

• The danger of overfitting and how to spot it.
• From lab to clinic: domain adaptation, transfer learning, external validations, benchmarking.

Acknowledgements

References