In positron emission tomography (PET)/computed tomography (CT), CT is used for attenuation correction (AC). CT-based AC (CTAC) is susceptible to misregistration due to respiratory phase differences between CT and PET, frequently resulting in a “banana artifact (BA)”—an underestimation of tracer uptake immediately below the diaphragm. This study aimed to evaluate a deep learning-based CT-less AC (DLAC), trained on respiratory-phase-matched data using data-driven gated (DDG) PET, for its efficacy in reducing BA. We retrospectively analyzed 18F-fluorodeoxyglucose (FDG) PET/CT datasets from a Discovery MI-25 (GE Healthcare) with DDG PET acquisition. DDG removed respiratory blurring in PET, and no AC errors were observed in cases with respiratory-phase-matched CT. Out of 1137 consecutive clinical cases, 255 well-aligned cases (WA-cases) and 387 cases with BA (BA-cases) were selected. The WA-cases showed no detectable misregistration in the upper abdomen through visual discrimination. The model, trained on pairs of non-AC PET (NAC-PET) and CTAC PET (CTAC-PET) of the WA-cases, processed NAC-PET to generate DLAC PET (DLAC-PET), resulting in reduced AC errors. Performance was assessed via fivefold cross-validation in WA-cases and artifact reduction testing in BA-cases. In WA-cases, DLAC-PET highly correlated with CTAC-PET, with metrics showing excellent agreement: mean absolute error (MAE) 0.029, root mean squared error (RMSE) 0.086, peak signal-to-noise ratio (PSNR) 35.5 dB, structural similarity (SSIM) 0.937, and Pearson correlation coefficient (r) 0.984. MAE and RMSE are expressed in standardized uptake values units. In BA-cases, DLAC-PET successfully suppressed BA in 385 out of 387 cases (99.5%), and no BA was observed in 366 cases (94.6%). Our DLAC model, trained on DDG-derived respiratory-phase-matched NAC-PET and CTAC-PET, effectively and consistently reduced BA in almost all problematic cases, potentially facilitating clinical interpretation in CTAC-failing cases.
TAKAMURA et al. (Sun,) studied this question.