Hallucination assessment for SPECT attenuation map generation from ¹⁷⁷Lu-Dotatate and ^99mTc MIP-1404

A 3d-UNET based approach is used to generate SPECT emission data only attenuation ("µ" )-map from ¹⁷⁷ Lu-Dotatate or ^99m Tc MIP-1404 acquisitions. Two 3d-UNETs (U 1 and U 2 ) are trained using clinical SPECT/CT ¹⁷⁷ Lu-Dotatate and ^99m Tc MIP-1404 data, respectively, and are applied to two test data sets (T 1 and T 2 ) consisting of 36 ¹⁷⁷ Lu-Dotatate and 20 ^99m Tc MIP-1404 clinical SPECT/CT studies, respectively. We evaluate the cross-tracer image qualities of the µ-maps using the structural similarity index measure (SSIM) and the line profiles in three different planes. For the SPECT reconstructions using these µ-maps, we also evaluate the quantitative accuracy of the reconstructions by performing statistical testing for the activity concentrations measured on 10 spheres with random locations inside the body and having different volumes from 5mL to 15mL. Cross-tracer SSIM values for µ-maps are: 0.85± 0.05(U 1 ,T 1 ), 0.78±0.02(U 1 ,T 2 ), 0.70±0.04(U 2 ,T 1 ), 0.84±0.04(U 2 ,T 2 ). The SSIM values for attenuated reconstructions are: 0.99 ± 0.008 for ¹⁷⁷ LuDotatate tracer, and 0.98 ± 0.009 for ^99m Tc MIP-1404 tracer. The Mann-Whitney-Wilcoxon test shows that there is no significant difference between the average activity concentrations measured on 10 random spheres (p-value = 0.9). In conclusion, SPECT alone can generate decent µ-maps using trained 3d-UNETs for ¹⁷⁷ Lu-Dotatate and ^99m Tc MIP-1404 tracer compounds. The reconstructions using these µ-maps show very little differences with that using µ-maps derived from CT, either visually or quantitatively. But the UNET trained on one tracer cannot be used directly to the other tracer, although the images from these two tracers cover similar body parts.