Best Paper Award is “huge boost” for cardiac imaging researchers

Early and accurate diagnosis is essential for heart patients, with Magnetic Resonance Imaging (MRI) representing the gold standard for structural and functional assessment of the heart.

However, the images generated by the MRI machine need to be looked over by a clinician who can segment out the main heart structures, particularly the ventricles. This segmentation, typically done by hand, is a costly, time-consuming process with significant room for human error.

Automation has the potential to change this, but it depends on deep learning approaches which are limited by the available data. Because of technical, ethical, confidentiality and economic constraints, the availability of images for clinical datasets is much more limited – they are made up of hundreds or thousands of images, next to the several million found in computer vision datasets.

DPhil candidates Jorge Corral Acero and Hao Xu are aiming to change this with their work on a process called augmentation. Based in Professor Vicente Grau’s group at the Oxford e-Research Centre, they hope to get around this issue by generating synthetic images to increase the size of these datasets.

“We have developed a new methodology called SMOD, or Statistical Models of Deformations,” explains Jorge. “By learning the shape variability of the heart, through statistical models, SMOD can generate anatomically meaningful deformations to enlarge the training datasets. The new augmented images maintain the appearance of the originals, but adopt new, plausible shapes. In this way, we believe that a better generalization is achieved during the course of the training.”

Their results show SMOD performing better than standard augmentation methods, especially when the training data available is very limited or when the structures being segmented are complex and variable. This, the pair believe, suggests that their method has particular potential for diagnosing diseases such as hypertrophic cardiomyopathy (which displays high anatomical variability) and right ventricle segmentations (a more complex shape than the left ventricle’s).

Going forwards, they hope to see models of deformation being learned from secondary larger datasets and applied to the actual training dataset, further increasing the range of plausible shapes and potentially leading to a better generalization. The method could also be scaled up and applied to other applications, including heart motion tracking and segmentation of any other biological structures, from organs to cells, in different imaging modalities.

The pair travelled to Bordeaux to present their studies at the 10th International Conference on Functional Imaging and Modelling of the Heart (FIMH), which took place in June 2019. They were rewarded with the Best Paper Award in the Imaging Category, beating researchers from around the world.

“We see the prize as a huge boost of motivation,” he said. “While for Hao it is a nice push to finish and write up the thesis, for me it is the best scenario that I could have ever imagined: being awarded in my first conference!

“What is clear is that the award is result of hard teamwork. Thus, we would like to thank our supervisor, Vicente Grau, and all the authors. In particular, we need to acknowledge Ernesto Zacur, a colleague who recently left the Department and contributed especially to this work, both professionally and personally.”