Flash 3d

Lung perfusion SPECT/CT was performed with patients positioned supine in the same position as RT simulation using a breast board allowing for direct registration of the SPECT/CT scans and RT treatment planning scans. The lung SPECT/CT scans were performed at baseline prior to RT and at one year following completion of RT. SPECT/CT was performed after intravenous injection of 5 mCi of [^99mTc]macroaggregated albumin (MAA). SPECT/CT imaging consisted of acquisition of 64 projection views covering 360° with a dual-head, hybrid tomograph (Siemens Symbia T6, Siemens Medical, Hoffman Estates, IL, USA). CT imaging of the field-of-view employed 70 mAs, 130 kVp, pitch 0.8, 2.5 mm axial reconstruction. The radiotracer projection views were reconstructed tomographically employing the CT data for attenuation correction using an iterative algorithm (OSEM, with 8 iterations and 4 subsets; FLASH 3D, Siemens Medical, Hoffman Estates, IL, USA). Use of attenuation correction in SPECT reconstructions renders the data linearly proportional to the voxel-wise concentration of radiotracer. Co-registered MAA perfusion images and X-ray CT images were provided for further analyses. All pre-RT and one year post-RT lung perfusion SPECT/CT scans were read by the same expert nuclear medicine physician. These two scans were compared for each patient and qualitative perfusion changes were recorded.

Test data, including the patient data, were reconstructed using an in-house developed OS-EM algorithm with the scatter estimate included as an additive term in the following statistical model:
$$Y_i~\mathrm{P}\mathrm{o}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{o}\mathrm{n}\left\{\left(\sum _{j=1}^{n_p}{a}_{ij}{x}_j\right)+s_i\right\},$$
where $Y_i$ denotes the number of counts measured in the ith detector pixel, $a_{ij}$ denotes elements of the system matrix A that models effects of depth-dependent attenuation and collimator/detector blur for a photon leaving the jth voxel towards the ith detector pixel, $s_i$ denotes the estimated scatter component for the ith detector pixel and $x=(x_1,\dots ,x_J)$ denotes the vector of unknown ⁹⁰Y activity voxel values. The current implementation used the same CNN-generated scatter estimate in all iterations, without updating. OS-EM reconstruction according to above equation was performed 1) without SC (NO-SC) using clinic software (Siemens Flash 3D) 2) with our previously reported [3 (link)] MC scatter model (MC-SC) 3) with the scatter estimate from the trained DCNN (DCNN-SC) 4) with the true scatter estimate from SIMIND (TRUE-SC), which was available only in the case of the simulated test data set. All reconstructions included CT-based attenuation correction, 3D collimator-detector response compensation and no post-filtering. MC-SC used two updates of the SIMIND generated scatter estimate based on our prior findings on convergence [3 (link)].