4D Multimodal Atlas of the Swine Heart from PET-CT Images

Objectives Multimodal imaging systems, such as hybrid PET-CT, capture the dynamics of cardiac anatomy and function. Yet, the quantitative characterization and subsequent comparison of anatomy and function of the heart is a challenging task due to the variations in both anatomy and function across different subjects. The goal of this work is to develop a 4D (3D space with time) probabilistic multimodal atlas of the swine heart for study of methods of automated analysis that can potentially be translated to clinical application.

Methods We acquired fifteen healthy pigs with weights ranging from 30-92 Kg, using a hybrid PET-CT scanner yielding three data sets per study, including CT (0.97mm×0.97mm×3mm), CTA (0.33mm×0.33mm×1mm), and PET (2.14mm×2.14 mm×3mm). The 15 volume sets were used to compute mean and variance of anatomic, vascular, and physiological maps of flow in a standard orientation and consistent physical size—i.e., the multimodal atlas. PET volumes were acquired dynamically with the blood flow tracer [18F]Flurpiridaz and parametric perfusion maps were computed with a kinetic model. The approach is based on a hierarchical method, constructing the atlas transformations from the CTA volumes to the PET volumes. First, we created a common spatial space from CTA volumes using a diffeomorphic groupwise registration method to define a basic structural layer in which the other modalities can be mapped. Second, we transported all other modalities including CT, dynamic PET, and gated PET volumes of all subjects into this common space via the deformations learnt above. Third, we constructed a gated 4D PET atlas from deformed gated PET volumes using a frame-by-frame atlas construction scheme. In addition, we determined the best matching phase of gated PET atlas volumes, corresponding to the phase of CTA, followed by creating a “motion-frozen” volume (MFV) from gated PET atlas volumes for each subject using consecutive diffeomorphic registrations between the best matching phase and the remaining phases. For the atlas construction, widely used similarity measures including mutual information, sum-of-squared-differences (SSD), and cross correlation (CC) were used to evaluate the CTA atlas. For the quantitative comparison, two sharpness measures, including the intensity variance measure (M1) and the energy of image gradient measure (M2) were used. In addition, to measure the accuracy of the registration during the atlas building, two expert observers manually selected anatomical landmarks and measured target registration error (TRE).

Results Our mean multimodal atlas was successfully constructed as illustrated in Fig. 1, where the CTA atlas, CT atlas, and MFV volumes, and the fusion of all the modalities are depicted. We evaluated image quality using the two sharpness measures as listed in Table 1. The data suggest that the CC similarity measure provided the best quality. The TRE obtained from the two anatomical landmarks was 2.47±1.56mm.

Conclusions This atlas provides myocardial images of a standard size and orientation combining anatomic and physiological data for pigs weighing from 30 to 92 Kg. To our knowledge, this is the first probabilistic multimodal atlas combining anatomic and functional images in a standard space. This method can potentially be extended to clinical applications involving automated image interpretation. $$graphic_8B12ADD6-084C-4099-8656-A8E7E8107FA6$$

• Download figure
• Open in new tab
• Download powerpoint