A Multimodal Spatiotemporal Atlas for Cardiac Functional Assessment

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

The estimation of cardiac motion is an important aid in the quantification of the contractility and function of the left ventricular myocardium, as well in detect-ing cardiovascular disease. A statistical cardiac motion atlas provides a space in which the motions of a cohort of subjects can be directly compared. Statistical atlases have been proposed for characterising abnormal cardiac motion, as well as for detecting suspected disease as early as possible. Typically, such atlases are formed and applied using data from the same modality, e.g. cardiac magnetic resonance (MR) or 3D ultrasound (US). This thesis proposes a new pipeline to build a multi-modal cardiac atlas from both MR and US data. The hypothesis is that such an atlas will benefit from the synergies between the motion features derived from the two modalities.
The processing pipeline of the multimodal motion atlas formation initially involves normalisation of subjects’ cardiac geometry and motion both spatially and over time, and extraction of motion descriptors, i.e. displacements. This step was accomplished following a similar pipeline proposed by other authors for single modality atlas formation. The main novelty of this project lies in the use of a dimensionality reduction algorithm to simultaneously reduce the dimension-ality of both the MR and US derived motion data. Three di˙erent dimensional-ity reduction algorithms were investigated: Principal component analysis (PCA), Canonical correlation analysis (CCA), and Partial least squares regression (PLS). A leave-one-out cross validation was employed to quantify the accuracy of the three algorithms. Results show that Partial least squares regression resulted in lower errors, with a reconstruction error less than 2.5 mm for MR-derived motion data, and less than 3 mm for US-derived motion data.
The second part of the project aims to describe a diagnostic pipeline which uses as input only US data, but is at the same time informed by a training database of multimodal MR and US data. To this end, the previous multi-modal cardiac motion atlas is used together with multi-view machine learning algorithms to combine and extract the most meaningful cardiac descriptors for classification using only US data. More specifically, two algorithms are proposed: multi-view linear discriminant analysis (MLDA) and multi-view Laplacian sup-port vector machines (MvLapSVM). Furthermore, a novel regional multi-view approach is proposed to exploit the regional relationships between the two modal-ities. The proposed pipeline is evaluated on the classification task of discrimi-nating between normals and patients with dilated cardiomyopathy. Results show that the use of multi-view classifiers together with a cardiac motion atlas results in a statistically significant improvement in accuracy compared to classification without the multimodal atlas. The highest accuracy for the global approach was achieved with the MvLapSVM algorithm and was 93.78%. In the regional case the highest accuracy was 95.78% using MvLapSVM.
Finally, the framework is extended to integrate automatically estimated strain values, and the strain values are used to validate the proposed pipeline for at-las formation and identification of DCM patients. Results show similar patterns using displacement and strain values. However, strain values consistently have slightly higher errors than displacement values.
Overall, I expect that the work presented in this thesis will have a significant impact on the assessment of cardiac function by enabling the exploitation of complementary information from multiple imaging modalities.
Date of Award2018
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorAndrew King (Supervisor), Paul Aljabar (Supervisor) & Julia Schnabel (Supervisor)

Cite this

'