Abstract: Many cardiac pathologies involve changes in the ventricular shape and size. In current clinical practice, echocardiography remains the workhorse for examination of the heart. Echocardiography has many advantages: it has a low echonomical cost, can be operated at bedside and produces real-time imaging. While classically most of images are 2D, recent advances allow the acquisition of 3D volumes. 3D has the potential to provide a more accurate assessment than 2D, since the whole heart can be analysed instead of just slices, but its automatic processing remains challenging due to the low quality of the images, in terms of both resolution and noise levels.

The first step for many is semantical segmentation: identifying which voxels of the image correspond to the ventricle. Deep learning has the potential to overcome the previously stated limitations, as convolutional neural networks (NN) have shown a great success in segmentation tasks. The main objective of this project is to train a novel deep learning (convolutional or graph NN) to automatically segment the ventricles, taking into account the
data limitations, in both quality and size, present in real clinical datasets. After this objective is achieved, further research directions are available: create personalized 3D models from the segmentations masks and perform statistical shape analysis to study regional geometry of the ventricles.

The project will work with already acquired data provided by different clinical colaborators and applications: including paediatrics congenital heart disease (Sickkids Hospital in Toronto), study of the effect of ingravity to the cardiac ventricles (CHU Caen) and fetal growth restriction (Hospital Casa de la Maternitat – Hospital Clinic).

Requisites:

• Python programming
• Image processing
• Prior knowledge with tensorflow / keras or other Deep Learning libraries is a plus.
• Previous knowledge in medical image analysis is not required.

Supervisor: Gabriel Bernadino