Scanner-integrated reconstruction versus post-processing deep learning for low-count ¹⁸F-FDG PET/CT: a comparative...

To compare two deep learning (DL) approaches for low-count PET/CT: deep progressive reconstruction (DPR), a scanner-integrated reconstruction-level method, and a deep-learning image-domain post-processing enhancement (POST; RaDynPET).

Sixty-seven patients who underwent whole-body ¹⁸F-FDG PET/CT were enrolled. PET images were reconstructed with ordered-subsets expectation maximization (OSEM) at 30/60/120 s/bed (O30, O60, O120 [clinical reference]) and with DPR at 30/60/90/120 s/bed (D30, D60, D90, D120). POST (RaDynPET) was applied to the unaltered O30 /O60 images to yield P30/P60. Two nuclear medicine physicians rated image quality using 5-point Likert scales. Liver signal-to-noise ratio (SNR), lesion tumour-to-background ratio (TBR), and contrast-to-noise ratio (CNR) were calculated. Non-inferiority (NI) versus O120 was prespecified for overall quality (Δ = -0.5) and lesion CNR (ratio lower bound 0.90). Time-matched DPR versus POST and DL versus OSEM were also assessed. Agreement with O120 was evaluated using Lin's concordance correlation coefficient (CCC) and Bland-Altman analysis.

Both DPR and POST achieved higher reader scores than time-matched OSEM. Inter-reader agreement was substantial to almost perfect. POST was superior at 30 s, whereas DPR was at 60 s. D60 and P30 met both NI margins, whereas D30 failed overall quality and P60 failed CNR. Concordance with O120 was excellent by CCC, and Bland-Altman showed small biases with limited proportional effects. CNR and SNR increased monotonically with DPR, while POST yielded gains at 30 s that attenuated at 60 s. TBR improvements were confined to DPR.

Both DPR and POST improved or preserved image quality while enabling scan-time reduction, with excellent agreement with the clinical reference. POST is supported for 1/4 acquisition time, whereas DPR is favored from 1/2 time onward.