Funding outcome

Cardiovascular magnetic resonance (CMR) is an excellent modality to examine heart diseases in children. However, regular CMR scans often need children to hold their breath, which can be difficult. When children are unable to hold their breath during the scan, the resulting images can be distorted due to motion, reducing their quality and making them harder to interpret accurately. Repeating scans to obtain good breath hold images can be frustrating for children. To solve this problem, we propose to develop a method to correct CMR images with breathing motion after they have been acquired, thereby allowing the child to breathe freely. First, we will train a model to learn the typical appearance of clear, motion-free CMR images. Next, we will build a 'degradation model' that learns the typical ways breathing and motion during scanning distort the images. Following this, the prior model, using the way distortions happen and what it learned about the clear images, learns to predict what the clear version of that distorted image should be. Therefore, when a distorted image is processed, the model essentially is determining, 'Given what I know about clear images and how they get distorted, what is the most likely clear version of this input distorted image?'. This approach aims to restore CMR images in a way that accounts for uncertainty while preserving critical anatomical details. It holds the promise of improving diagnostic accuracy of CMR for heart diseases in children, while enhancing patient comfort. Through this work, we aim to improve access to advanced diagnostic imaging and promote the responsible use of AI in healthcare.