**3.4. Image-based phenotyping**

Quantitative characterization of ventricular function has become important for the assessment of murine cardiac performance in heart disease [Wiesmann 2000, Wiesmann\_Circ\_Res 2001, Dawson 2004]. As the manipulation of the mammalian genome becomes routine, it is now possible to generate animal models to study cardiovascular function and dysfunction [Wiesmann\_Card\_Magn\_Reson 2001, Epstein 2002, Epstein 2007, Bucholz 2008, Stuckey 2012]. Critical to successful phenotypic screening of mouse models of the cardiovascular system using MRI are highly efficient four-dimensional (4D) acquisition ex-vivo and [Zamyadi 2010] in-vivo protocols [Bucholz 2010]. Such protocols ought to span the scales of the embryo to the adult, fully-developed mouse, and ought to lead to the reduction of the computational image processing complexity for accurate quantification of motion, global and regional cardiac function, strain, elasticity, and others.

#### *3.4.1. Quantification of global mechanical function and hemodynamic indices*

Traditionally, quantification of global hemodynamic indices has been associated with calculation of hemodynamic indices of function. MRI and other diagnostic techniques, (including microCT and ultrasound) are not the only existing techniques to accomplish functional characterization. While ultrasonic techniques are associated with inherent limitations for such quantification [Dawson 2004], microCT has been gaining interest and popularity [Badea 2005]. Miniature catheterization techniques introduced in the late 1990s [Georgakopoulos 1998] have revolutionized cardiac-based phenotypic in transgenic mice. However, their major drawbacks include their invasive nature and other technical limitations [Porterfield 2009, Constantinides\_IEEE 2011, Constantinides 2012]. As newly emerging techniques and applications develop to investigate in-vivo sarcomeric force generation with contrast agents [Constantinides\_ABME 2011], the current scientific focus is still on image-based calculations of global indices of function.

362 Practical Applications in Biomedical Engineering

**3.4. Image-based phenotyping** 

In summary, numerous practical benefits are associated with mouse cardiac MR imaging, including the non-invasive nature of the technique, the inherent capability to map cardiac morphology and function, for both LV and RV chambers, and their motional patterns. High spatial and temporal resolution imaging can thus be achieved, through execution of highthroughput protocols, yielding direct, accurate estimates of global and regional indices of cardiac function, avoiding any assumptions whatsoever or model-based derivation

Despite the extensive use of cartesian imaging with adequate SNR performance and spatial resolution (using FLASH, SSFP, or FISP), 3D acquisition studies maybe more efficiently completed using radial or spiral imaging sequences, especially for dynamic cardiac imaging (with pharmacologic interventions or contrast agent infusions). Nevertheless important and critical drawbacks are associated with such sequences, including the necessity to maintain data density as sampling extends to outer k-space regions, the convoluted and complex reconstructions (often associated with data re-gridding, kernel de-convolution, filtering, and inverse fourier transformation), and inherently lower SNR performance than rectilinear imaging. Thus, the choice between cartesian and radial imaging reduces to a tradeoff

between SNR, spatial resolution, and efficiency of data sampling for 3D coverage.

motion, global and regional cardiac function, strain, elasticity, and others.

*3.4.1. Quantification of global mechanical function and hemodynamic indices* 

Traditionally, quantification of global hemodynamic indices has been associated with calculation of hemodynamic indices of function. MRI and other diagnostic techniques, (including microCT and ultrasound) are not the only existing techniques to accomplish functional characterization. While ultrasonic techniques are associated with inherent limitations for such quantification [Dawson 2004], microCT has been gaining interest and popularity [Badea 2005]. Miniature catheterization techniques introduced in the late 1990s [Georgakopoulos 1998] have revolutionized cardiac-based phenotypic in transgenic mice. However, their major drawbacks include their invasive nature and other technical limitations [Porterfield 2009, Constantinides\_IEEE 2011, Constantinides 2012]. As newly

Quantitative characterization of ventricular function has become important for the assessment of murine cardiac performance in heart disease [Wiesmann 2000, Wiesmann\_Circ\_Res 2001, Dawson 2004]. As the manipulation of the mammalian genome becomes routine, it is now possible to generate animal models to study cardiovascular function and dysfunction [Wiesmann\_Card\_Magn\_Reson 2001, Epstein 2002, Epstein 2007, Bucholz 2008, Stuckey 2012]. Critical to successful phenotypic screening of mouse models of the cardiovascular system using MRI are highly efficient four-dimensional (4D) acquisition ex-vivo and [Zamyadi 2010] in-vivo protocols [Bucholz 2010]. Such protocols ought to span the scales of the embryo to the adult, fully-developed mouse, and ought to lead to the reduction of the computational image processing complexity for accurate quantification of

approaches endorsed by other imaging techniques such as ultrasound.

**Figure 10.** (Top) Typical bright blood true-short axis cardiac imaging at an apical, middle, and basal levels of the C57BL/6 murine heart using a liposomal contrast agent; (Bot) Black-blood spin-echo images of the murine heart at basal, mid, and apical levels**.** 

**Figure 11.** (A) ECG-gated four chamber view of a healthy mouse heart at 78 x 78 μm2 in-plane resolution, and (B) corresponding self-gated (retrospective) view of a separate mouse using a navigator echo at the same spatial resolution, at 11.7 T using FLASH-MRI [Images courtesy of Bruker Biospin MRI, Ettlingen, Germany].

Correspondingly, estimated cardiac volumes can easily be computed using standard image processing tools and converted to absolute volume units using the voxel dimension and the myocardial tissue density. Hemodynamic indices such as end-diastolic (EDV) and end-

systolic volumes (ESV), stroke volume (SV), cardiac output (CO), and ejection fractions (EF) can now be routinely calculated according to standard cardiac mechanical functional relations [Constantinides\_SBI 2009], based on CINE-MRI (Figure 11). CINE based local contractile function, including estimation of wall thickness, motion and fractional shortening, indicators of systolic function and long-term prognostic biomarkers in dysfunction remain the gold-standard for assessment of motional patterns and cardiac function in small animals due to the excellent soft tissue contrast of MRI and its high spatial and temporal resolution (Figure 12).

Study of the Murine Cardiac Mechanical Function Using Magnetic Resonance Imaging:

elicited post-transgenic modifications and the genetic basis of pathology [Henkelman 2010]. Therefore mouse imaging, combined with probabilistic, statistical atlas constructions and morphometric methods [Zamayadi 2010], is envisaged to play an increasing role in imagebased phenotyping and gene expression mapping of genetically altered mice [Ng 2010] in the future. Construction of probabilistic and statistical atlases [Perperidis 2011] can potentially enable the study of murine, global cardiac structure and function with increased quantitative accuracy, identifying modal components of shape variability (from embryogenesis to adulthood), and disseminating components of global mechanical motion. Similar to existing cardiac atlases [Perperidis 1995, Hoogendoorn 2007], these can be population-based instead of single-subject. Prior efforts have focused on modeling cardiac anatomy in humans [Helm 2006] but only limited attempts have been made to construct accurate, high-spatial and hightemporal resolution computerized atlases for mice [Perperidis 2011]. Despite multiple prior efforts with construction of human brain atlases [Young 2009], Frangi et al. [Frangi 2002], Mitchell et al. [Mitchell 2002], and Lotjonen et al. [Lotjonen 2004] were the first to develop human ventricular statistical shape models. More recently Ordas et al. [Ordas 2007] developed a computational atlas of the entire heart using registration-based techniques. Perperidis et al. [Perperidis 1995, Perperidis 2005] and Hoogendoorn et al. [Hoogendoorn 2007] proposed 4D spatio-temporal human cardiac probability atlases from (MRI), while Beg et al. [Beg 2004] developed a large deformation diffeomorphic mapping (LDDM) for construction of cardiac statistical atlases. As an extension to Beg's work, Helm et al. [Helm 2006] employed LDMM to achieve inter-subject registration to a reference anatomical template to compare cardiac geometric variability using Principle Component Analysis (PCA) from diffusion tensor MRI in

normal and failing human hearts.

*Manual and semi-automated segmentation and registration approaches* 

ranking of 14 different non-linear deformation algorithms [Klein 2009].

diagrammatically summarized in Figure 14 [Perperidis 2011].

Critical to successful constructions of atlases and to the transformation of constructed surface models to the finite element models (for subsequent computational work of mechanical function), are efficient and accurate segmentation and registration techniques (Figures 13, 14). Using recently developed techniques in our group [Perperidis 2011], a segmented (template) reference MRI sequence is selected from the mouse database consisting of imaged anatomic structures (left and right ventricular myocardium, left and right ventricular blood pools, and papillary muscles) as representative sets of typical anatomic objects of interest. This template is subsequently used to register population images via global and local non-rigid transformations. Although such high-dimensional transformation differs from established one-to-one (invertible) diffeomorphic transformations [Helm 2006], its performance is comparable in accuracy, precision, simplicity, and computational intensity, as recently reported by a multi-center quantitative

Figure 13 show examples of manual, user-based spline segmentation and subsequent correction for left and right ventricular feature extraction, as discussed previously [Perperidis 2011]. Implementation of a global and local registration scheme is also

The Current Status, Challenges, and Future Perspectives 365

**Figure 12.** Single-slice, multiphase CINE functional cardiac imaging of the C57BL/6 mouse heart over 12 cardiac phases.

#### *Interstrain morphological and 4D motional variability – Statistical altases*

The advances in high-field, high-resolution cardiac imaging techniques, have also allowed the development of atlas-based approaches for the description of anatomical structures and their function [Ali 2005, Sharief 2008] in normal [Ruff 1998, Bucholz 1998, Badea 2005, Zamyadi 2010] and transgenic mice [Chien 2000, Epstein 2007]. Completion of the sequencing of the mouse genome has led to increased requirements for identifying the specific phenotype elicited post-transgenic modifications and the genetic basis of pathology [Henkelman 2010]. Therefore mouse imaging, combined with probabilistic, statistical atlas constructions and morphometric methods [Zamayadi 2010], is envisaged to play an increasing role in imagebased phenotyping and gene expression mapping of genetically altered mice [Ng 2010] in the future. Construction of probabilistic and statistical atlases [Perperidis 2011] can potentially enable the study of murine, global cardiac structure and function with increased quantitative accuracy, identifying modal components of shape variability (from embryogenesis to adulthood), and disseminating components of global mechanical motion. Similar to existing cardiac atlases [Perperidis 1995, Hoogendoorn 2007], these can be population-based instead of single-subject. Prior efforts have focused on modeling cardiac anatomy in humans [Helm 2006] but only limited attempts have been made to construct accurate, high-spatial and hightemporal resolution computerized atlases for mice [Perperidis 2011]. Despite multiple prior efforts with construction of human brain atlases [Young 2009], Frangi et al. [Frangi 2002], Mitchell et al. [Mitchell 2002], and Lotjonen et al. [Lotjonen 2004] were the first to develop human ventricular statistical shape models. More recently Ordas et al. [Ordas 2007] developed a computational atlas of the entire heart using registration-based techniques. Perperidis et al. [Perperidis 1995, Perperidis 2005] and Hoogendoorn et al. [Hoogendoorn 2007] proposed 4D spatio-temporal human cardiac probability atlases from (MRI), while Beg et al. [Beg 2004] developed a large deformation diffeomorphic mapping (LDDM) for construction of cardiac statistical atlases. As an extension to Beg's work, Helm et al. [Helm 2006] employed LDMM to achieve inter-subject registration to a reference anatomical template to compare cardiac geometric variability using Principle Component Analysis (PCA) from diffusion tensor MRI in normal and failing human hearts.

#### *Manual and semi-automated segmentation and registration approaches*

364 Practical Applications in Biomedical Engineering

and temporal resolution (Figure 12).

12 cardiac phases.

systolic volumes (ESV), stroke volume (SV), cardiac output (CO), and ejection fractions (EF) can now be routinely calculated according to standard cardiac mechanical functional relations [Constantinides\_SBI 2009], based on CINE-MRI (Figure 11). CINE based local contractile function, including estimation of wall thickness, motion and fractional shortening, indicators of systolic function and long-term prognostic biomarkers in dysfunction remain the gold-standard for assessment of motional patterns and cardiac function in small animals due to the excellent soft tissue contrast of MRI and its high spatial

**Figure 12.** Single-slice, multiphase CINE functional cardiac imaging of the C57BL/6 mouse heart over

The advances in high-field, high-resolution cardiac imaging techniques, have also allowed the development of atlas-based approaches for the description of anatomical structures and their function [Ali 2005, Sharief 2008] in normal [Ruff 1998, Bucholz 1998, Badea 2005, Zamyadi 2010] and transgenic mice [Chien 2000, Epstein 2007]. Completion of the sequencing of the mouse genome has led to increased requirements for identifying the specific phenotype

*Interstrain morphological and 4D motional variability – Statistical altases* 

Critical to successful constructions of atlases and to the transformation of constructed surface models to the finite element models (for subsequent computational work of mechanical function), are efficient and accurate segmentation and registration techniques (Figures 13, 14). Using recently developed techniques in our group [Perperidis 2011], a segmented (template) reference MRI sequence is selected from the mouse database consisting of imaged anatomic structures (left and right ventricular myocardium, left and right ventricular blood pools, and papillary muscles) as representative sets of typical anatomic objects of interest. This template is subsequently used to register population images via global and local non-rigid transformations. Although such high-dimensional transformation differs from established one-to-one (invertible) diffeomorphic transformations [Helm 2006], its performance is comparable in accuracy, precision, simplicity, and computational intensity, as recently reported by a multi-center quantitative ranking of 14 different non-linear deformation algorithms [Klein 2009].

Figure 13 show examples of manual, user-based spline segmentation and subsequent correction for left and right ventricular feature extraction, as discussed previously [Perperidis 2011]. Implementation of a global and local registration scheme is also diagrammatically summarized in Figure 14 [Perperidis 2011].

Study of the Murine Cardiac Mechanical Function Using Magnetic Resonance Imaging:

**Figure 14.** Graphical representation of land-marking as an initial step to image registration. The landmarks are propagated from a reference landmarked template heart model to each remaining mouse heart of each cohort. (B) Typical mid-short and long-axis images from a mouse heart; (C) RView corrected binary masks; (D) Example of five-point landmarked mesh for global registration; (E) 3D epicardial/endocardial reference surface model rendition used for local registration, and (F) construction of the three-dimensional (3D) mesh volumetric dataset using the marching cubes algorithm. [This work originally appeared in Perperidis et al. (Perperidis et al. 2011), published in

Despite its potential usefulness, the major limitations of such a technique is associated with the significant image post-processing, the need for accurate registration methods (often a challenging task for cardiac datasets), the inherent assumptions associated with the normality of distribution of independent datasets, and confounding factors on attempts to

analyze motion and variability (often as a direct modulation of anesthesia effects).

Computerized Medical Graphics; Reproduced with the permission of Elsevier].

The Current Status, Challenges, and Future Perspectives 367

**Figure 13.** Flow diagrammatical representation of the empirical optimization process for generation of quantitatively accurate finite element models of the murine myocardium. [Reproduced from Constantinides et al. [Constantinides\_SBI\_2010] with permission from IEEE]. Insert (top right): (top row) Seed-point segmentation and binary mask construction in short axis MRI; (middle and bottom rows) RView 3D binary mask correction. [This work originally appeared in Constantinides et al. [Constantinides 2010] and Perperidis et al. [Perperidis et al. 2011], and was published in Computerized Medical Graphics and the IEEE Proceedings of the International Society of Biomedical Imaging; Reproduced with the permission of Elsevier and IEEE].

Manual, image-based techniques seem to provide the best avenue for segmentation (despite their inherent intra- and inter-observer variability inaccuracies).

Certainly, atlas-based approaches are envisaged to be of tremendous benefit and value for cardiac phenotyping characterization in the upcoming years, aiding spatial mapping of gene expression using novel cellular and sub-cellular probes and markers, understanding embryogenesis and development, accurately mapping strain-, age-, and sex-based morphological dependencies, quantifying patterns of motional variability, and efficiently screening cardiac functional changes based on semi-automatic template segmentation techniques for efficient estimation of hemodynamic indices of function, and high throughput phenotypic screening.

**Figure 13.** Flow diagrammatical representation of the empirical optimization process for generation of

Manual, image-based techniques seem to provide the best avenue for segmentation (despite

Certainly, atlas-based approaches are envisaged to be of tremendous benefit and value for cardiac phenotyping characterization in the upcoming years, aiding spatial mapping of gene expression using novel cellular and sub-cellular probes and markers, understanding embryogenesis and development, accurately mapping strain-, age-, and sex-based morphological dependencies, quantifying patterns of motional variability, and efficiently screening cardiac functional changes based on semi-automatic template segmentation techniques for efficient estimation of hemodynamic indices of function, and high

quantitatively accurate finite element models of the murine myocardium. [Reproduced from Constantinides et al. [Constantinides\_SBI\_2010] with permission from IEEE]. Insert (top right): (top row) Seed-point segmentation and binary mask construction in short axis MRI; (middle and bottom rows) RView 3D binary mask correction. [This work originally appeared in Constantinides et al. [Constantinides 2010] and Perperidis et al. [Perperidis et al. 2011], and was published in Computerized Medical Graphics and the IEEE Proceedings of the International Society of Biomedical Imaging;

Reproduced with the permission of Elsevier and IEEE].

throughput phenotypic screening.

their inherent intra- and inter-observer variability inaccuracies).

**Figure 14.** Graphical representation of land-marking as an initial step to image registration. The landmarks are propagated from a reference landmarked template heart model to each remaining mouse heart of each cohort. (B) Typical mid-short and long-axis images from a mouse heart; (C) RView corrected binary masks; (D) Example of five-point landmarked mesh for global registration; (E) 3D epicardial/endocardial reference surface model rendition used for local registration, and (F) construction of the three-dimensional (3D) mesh volumetric dataset using the marching cubes algorithm. [This work originally appeared in Perperidis et al. (Perperidis et al. 2011), published in Computerized Medical Graphics; Reproduced with the permission of Elsevier].

Despite its potential usefulness, the major limitations of such a technique is associated with the significant image post-processing, the need for accurate registration methods (often a challenging task for cardiac datasets), the inherent assumptions associated with the normality of distribution of independent datasets, and confounding factors on attempts to analyze motion and variability (often as a direct modulation of anesthesia effects).
