Automatic high-resolution infarct detection using volumetric multiphase dual-energy CT.
J Cardiovasc Comput Tomogr. 2017 Jul - Aug;11(4):288-294. doi: 10.1016/j.jcct.2017.04.006. Epub 2017 Apr 18. Sandfort V1, Kwan AC2, Elumogo C1, Vigneault DM1, Symons R1, Pourmorteza A1, Rice K3, Davies-Venn C1, Ahlman MA1, Liu CY1, Zimmerman SL2, Bluemke DA4.
OBJECTIVES: Late contrast enhancement CT (LCE-CT) visualizes the presence of myocardial infarcts. Differentiation of the contrast-enhanced infarct from blood pool is challenging. We developed a novel method using data from first pass CT angiography (CTA) imaging to enable automatic infarct detection.
MATERIALS AND METHODS: A canine model of myocardial infarction was produced in 11 animals. Two months later, first pass CTA (90 kVp) and LCE-CT (dual energy 90 kVp/150 kVp tin filtered) were performed. Late gadolinium enhancement MRI was used as reference standard. The CTA and LCE-CT were co-registered using a fully automatic non-rigid method based on curved B-splines. The method allowed for limited elastic deformation and the considerable differences in attenuation between first-pass and delayed image. The blood pool was easily identified on the CTA image by high attenuation. Because CTA and LCE-CT were registered, the blood pool segmentation can be directly transferred to the LCE-CT - thereby solving the key problem of infarct/blood pool differentiation. The remaining segmentation of infarcted vs. noninfarcted myocardium was performed using a threshold. Automatic and MRI-guided expert segmentations of LCE-CT infarcts were compared to each other on volume and area basis (intraclass correlation coefficient, ICC) and on voxel basis (dice similarity coefficient, DSC between automatic and expert CT segmentation). CT infarct volumes were compared with the reference standard MRI.
RESULTS: The infarcts were mainly subendocardial (81%) and relatively small (median MRI infarct mass 7.4 g). The automatic segmentation showed excellent agreement with expert segmentation on volume and area measurements (ICC = 0.96 and 0.87, respectively). DSC showed moderately good agreement (DSC = 0.47). Compared to MRI there was modest agreement (ICC = 0.62) and excellent correlation (R = 0.9). Manual interaction was less than 1 min per exam.
CONCLUSION: We propose an automatic method for infarct segmentation on LCE-CT using multiphase CT information, which showed excellent agreement with expert readers and favorable correlation with MRI.