Fully automatic left ventricular myocardial strain estimation in 2D short-axis tagged magnetic resonance imaging

Abstract

Cardiovascular diseases are among the leading causes of death and frequently result in local myocardial dysfunction. Among the numerous imaging modalities available to detect these dysfunctional regions, cardiac deformation imaging through tagged magnetic resonance imaging (t-MRI) has been an attractive approach. Nevertheless, fully automatic analysis of these data sets is still challenging. In this work, we present a fully automatic framework to estimate left ventricular myocardial deformation from t-MRI. This strategy performs automatic myocardial segmentation based on B-spline explicit active surfaces, which are initialized using an annular model. A non-rigid image-registration technique is then used to assess myocardial deformation. Three experiments were set up to validate the proposed framework using a clinical database of 75 patients. First, automatic segmentation accuracy was evaluated by comparing against manual delineations at one specific cardiac phase. The proposed solution showed an average perpendicular distance error of 2.35 ± 1.21 mm and 2.27 ± 1.02 mm for the endo- and epicardium, respectively. Second, starting from either manual or automatic segmentation, myocardial tracking was performed and the resulting strain curves were compared. It is shown that the automatic segmentation adds negligible differences during the strain-estimation stage, corroborating its accuracy. Finally, segmental strain was compared with scar tissue extent determined by delay-enhanced MRI. The results proved that both strain components were able to distinguish between normal and infarct regions. Overall, the proposed framework was shown to be accurate, robust, and attractive for clinical practice, as it overcomes several limitations of a manual analysis.

An Overview on Image Registration Techniques for Cardiac Diagnosis and Treatment

“…Selection of spatial transformation algorithms also depends on the image dimensionality, whether involving the registration from 2D to 2D [ 10 , 11 , 136 , 171 , 183 ], 3D to 3D [ 22 , 49 , 74 , 78 , 139 , 162 , 179 , 188 , 189 ], or 2D to 3D [ 79 , 92 , 93 , 143 , 190 , 191 ]. For 2D to 2D registration, where the acquisition tightly controls the geometry of the images, the images can simply be registered via a rotation and two orthogonal translations.…”

MITT: Medical Image Tracking Toolbox

“…Sandro Queirós, Pedro Morais, Daniel Barbosa, Jaime C. Fonseca, João L. Vilaça, and Jan D'hooge M Overall, while image tracking techniques have a tremendous applicability, most existing solutions consist of independent implementations with very specific target applications. Taking the cardiology field as an example [17,[32][33][34][35][36][37][38], image tracking solutions are typically aimed at a certain modality or sequence (e.g., ultrasound, computed tomography, cine or tagged magnetic resonance sequences) with a particular image dimensionality (bi-or tridimensional sequences) and to assess a single chamber (left or right ventricle or atrium, the myocardium or a cardiac valve), thus lacking the versatility to deal with distinct end-goals without the need for methodological tailoring and/or exhaustive tuning of numerous and complex parameters.…”

Discriminative dictionary learning for local LV wall motion classification in cardiac MRI

“…Tagged magnetic resonance (TMR) is considered as gold standard for LV motion tracking as motion fields extracted from this modality already constitute an indicator of local function and are used to derive other indicators, such as strains (Garcia-Barnes et al, 2010;Morais et al, 2017). However, TMR sequences are not always available or of sufficient quality to use them in clinical routine while cine-MRI is the first sequence used.…”