TY - JOUR T1 - Model-based 3D reconstruction of myocardial perfusion SPECT images JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 26 LP - 26 VL - 58 IS - supplement 1 AU - Yinfeng Dong AU - Eric Frey AU - Yong Du Y1 - 2017/05/01 UR - http://jnm.snmjournals.org/content/58/supplement_1/26.abstract N2 - 26Objectives: Partial volume (PV) effects resulting from finite spatial resolution significantly degrade the accuracy of activity quantification and left ventricle (LV) segmentation in the analysis of myocardial perfusion SPECT (MPS) images. High-resolution anatomical images, such as contrast CT or MRI, can be used for segmentation and PV compensation, but they are typically not available in routine clinical practice. The objective of this study was to develop and evaluate a model-based 3D reconstruction of MPS images to improve quantification and LV segmentation accuracy without the use of anatomical images.Methods: The proposed model-based reconstruction method describes the LV using 3D geometry and activity distribution models instead of voxels. The geometry model parameterizes the endo- and epicardial surfaces using a set of rays originating from the long axis of the LV. The rays are in planes perpendicular to the axis and at equiangular increments in the basal and mid-ventricular regions. In the apex, the rays radiate from a point with equal polar and azimuthal angular increments. The surfaces are obtained by interpolating the intersection points of each surface and the rays using cubic-spline functions. The activity distribution model divides the myocardium into segments similar to those used in standard quantitative MPS analysis and assumes uniform activity concentrations in each segment and the blood pool and body background. The method estimates the model parameters of voxel intensities, greatly reducing the number of unknowns. The parameters are estimated that give the best match between the image generated from the geometry and activity models and the measured data. The metric for the goodness of fit was a maximum likelihood criterion based on the modeling of noise statistics in the reconstructed image. The effects of finite resolution were modeled with a spatially invariant Gaussian function. A penalty encouraging uniform wall thickness was incorporated into the objective function to regularize the reconstruction problem and to increase robustness to perfusion defects and noise. The parameters were optimized by seeking the maximum of the objective function. The proposed method was evaluated using simulated MPS images based on the XCAT phantom. The myocardium contained a perfusion defect with 50° angular and 20mm axial extent, and 50% severity. The relative activity concentrations were: blood pool=15, background=5, normal myocardium=95, perfusion defect=45. The noise level matched that in clinical MPS images. For simplicity, attenuation and scatter were not modeled in the projection data.Results: The resulting parameters provided an accurately segmented LV myocardium and PV compensated measurements of the true activities. The overlap between the resulting segmented LV myocardium and the truth was about 91%. The proposed model-based reconstruction method reduced the RMSE of estimated LV myocardial activity concentrations from 40.37 to 11.92 in comparison to FBP reconstruction.Conclusion: We have developed a model-based 3D reconstruction method that incorporates prior knowledge about the targeted object via a model and models the image formation process. The method can simultaneously produce accurate LV segmentation and PV compensated images without using anatomical images. Research Support: AHA 16GRNT3109007 ER -