Comparison of multi modal SPECT reconstruction methods using a clinically relevant assessment of the image resolution at various noise levels

摘要

The resolution of an iterative SPECT reconstruction cannot be characterized meaningfully by the standard approach of measuring the FWHM of an image of a point (line) in air or in a medium, because an iterative SPECT reconstruction is typically not linear. Our objective is to develop a general method to analyze the noise and resolution characteristics. The basic idea is to derive the resolution from the edge response of extended object and to correlate the resolution to the noise response. We now have measure akin to a contrast-detail diagram. Siemens introduced xSPECT Quant (xQ) that enables quantitative SPECT for Tc99m, and xSPECT Bone (xB) which is a dedicated solution for bone SPECT, where CT data is used to delineate tissue boundaries to improve image quality. We reconstruct images using the OSCGM based xSPECT method at varying updates (12, 24, 48, 96, 192, 480 and 960 updates, with one subset), as well as angular sampling of 3-degree (120 views ) and 6-degree (60 views subsampled from 120 views), and compare it to Flash3D (F3D), our OSEM based product and xEMAS (xE), which has the xQ image formation yet it is OSEM based. xE is not commercially available. All methods include the respective 3d collimator response compensation, attenuation and scatter correction, and a Gaussian post-smoothing. A uniform cylinder phantom, a hot-cold-rod (HCR) phantom and a hot-cold-sphere (HCS) were reconstructed using the four methods. We calculated the mean and standard deviation (SD) of activity concentration of a big VOI in the central part of the uniform cylinder or in the uniform section of HCS phantom, and used the mean/SD ratio to define the noise level. A matched filter method was used to measure the resolution in the HCR and HCS phantom, based on edge analysis of the rod sections. We then compared the resolutions at the same noise level at varying updates. Comparing reconstructions of two angular sampling, reconstructions from 3-degree data has over 20% lower noise level- compared to 6-degree angular sampling data. Comparing different reconstruction methods, at the same noise level, xQ has about 10% smaller FWHM than F3D, and xB has significantly better resolution than both F3D and xQ, for both 3-degree and 6-degree data. We attribute these advantages of xSPECT to the improved accuracy of the image formation as seen when comparing Flash3D, xEMAS and xQ.