Electronic Thesis and Dissertation Repository


Doctor of Philosophy


Biomedical Engineering


Terry Peters

2nd Supervisor

Shuo Li

Joint Supervisor


Automating global and segmental (regional) assessments of cardiac Left Ventricle (LV) function in Magnetic Resonance Images (MRI) has recently sparked an impressive research effort, which has resulted a number of techniques delivering promising performances. However, despite such an effort, the problem is still acknowledged to be challenging, with substantial room for improvements in regard to accuracy. Furthermore, most of the existing techniques are labour intensive, requiring delineations of the endo- and/or epi-cardial boundaries in all frames of a cardiac sequence.

On the one hand, global assessments of LV function focus on estimation of the Ejection Fraction (EF), which quantifies how much blood the heart is pumping within each beat. On the other hand, regional assessments focus on comprehensive analysis of the wall motions within each of the standardized segments of the myocardium, the muscle which contracts and sends the blood out of the LV.

In clinical practice, the EF is often estimated via manual segmentations of several images in a cardiac sequence. This is prohibitively time consuming, or via automatic segmentations, which is a challenging and computationally expensive task that may result in high estimation errors. Additionally, the diagnosis of the segmental dysfunction is based on visual LV assessments, which are subject to high inter-observer variability.

In this thesis, we propose accurate methods to estimate both global and regional LV function with minimal user inputs in real-time from statistics estimated in MRI. From a simple user input, we build image statistics for all the images in a subject dataset. We demonstrate that these statistics are correlated with regional as well as global LV function. Different machine learning techniques have been employed to find these correlations. The regional dysfunction is investigated in terms of a binary/multi-classification problem.

A comprehensive evaluation over 20 subjects demonstrated that the estimated EFs correlated very well with those obtained from independent manual segmentations. Furthermore, comparisons with estimating EF with recent segmentation algorithms show that the proposed method yielded a very competitive performance. For regional binary classification, we report a comprehensive experimental evaluation of the proposed algorithm over 928 cardiac segments obtained from 58 subjects. Compared against ground-truth evaluations by experienced radiologists, the proposed algorithm performed competitively, with an overall classification accuracy of 86.09% and a kappa measure of 0.73. We also report a comprehensive experimental evaluation of the proposed multi-classification algorithm over the same dataset. Compared against ground-truth labels assessed by experienced radiologists, the proposed algorithm yielded an overall 4-class accuracy of 74.14%.