Clinical Potential for Artificial Intelligence in PET Imaging: Phase 1 Result of Dose Reduction using Deep Learning Reconstruction

Objectives: Positron emission tomography/computed tomography (PET/CT) plays a central role in the management of cancer. However, the radiation exposure to patients from the radiopharmaceutical and the CT remains of concern in clinical practice. Efforts directed towards reducing the FDG dose to reduce radiation dose to patients referred for PET/CT have been done, but, both the relatively short half-life of FDG (109 minutes) and patients’ factors, mainly large body habitus may negatively affect the image quality if a reduce dose of FDG is administered. We decided to test the hypothesis that a reduced FDG dose will not affect the diagnostic quality of PET imaging when dedicated imaging parameters are used together with a deep learning reconstruction (DLR) algorithm. DLR is intended to reduce noise from FDG-PET images by using a technique made of artificial neural networks to produce better signal to noise ratio (SNR), where the neural network is trained to learn the noise from low-count FDG images by using high-count FDG images as gold standard. The aim of this study is to evaluate the SNR performance of DLR in clinical practices then to obtain an optimized protocol with reduced dose (equivalent time/bed position’s reduction) for oncological FDG PET scanning when using DLR.

Methods: This is a prospective trial planned in two phases with a primary purpose of assessing the feasibility of obtaining high quality and diagnostic images with a reduced (dose level of phase2 is decided based on result of phase1) dose of FDG in clinical PET protocol. The trial is conducted on consecutive participants that are referred for routine PET/CT imaging. Hereby we describe the phase 1 of this study. Participants were administered with a standard FDG dose [approximately 5.6MBq /kg (0.15mCi/kg)] as per EANM Dosage Calculator. The whole-body FDG oncological imaging was acquired in a PET-CT system (uMI 550, United-Imaging Healthcare, Shanghai, China), and 3.0mins per bed and overlap >35% were used. Each participant’s whole body FDG images were reconstructed into 3 bedtime configurations (3.0mins, 1.5mins and 1.0min), where 1.5mins and 1.0min are technically equivalent to dose reduction of ½ and 1/3) using DLR and OSEM with gaussian smooth (3mm) respectively. The mean SNR for whole body were calculated from signal and noise images separated by block-matching and 3D filtering algorithm. Paired t-test was used to compare the SNRs.

Results: Fifty participants entered the trial and data from 44 [mean age ± SD= 53.3 ± 17.4 years, range:12 years to 82 years; median weight= 70kg (IQR=60kg & IQR3=78kg), range:39kg to 112kg; and median BMI=24.4 (IQR1=21.8 & IQR3=26.6] were analyzed. The mean injected FDG dose was 377.77±69.56 MBq (10.21 ± 1.88 mCi). DLR’s SNRs are all significantly higher than OSEM’s SNRs (p<0.0001) for all bed times. The mean SNR for DLR is 1.28, 1.34, 1.33 times higher than OSEM for corresponding bed time of 3.0mins, 1.5mins to 1.0min. The SNR of DLR 1.0min bed time is 1.02 times higher than OSEM 3.0mins bed time. The DLR’s SNR over body weight also follows the same inverse square power relationship of the OSEM. The visual observation is consistent with the statistical results and shows improved quality of output images, particularly in participants with high body mass index (BMI).

Conclusions: The DLR for oncological FDG imaging is feasible with 1/3 reduction of equivalent FDG dose from the standard dose used in clinical OSEM protocol. A phase2 trial that compares same individuals at two time points (T1 &T2 within a week period) will soon start with a standard FDG dose at T1 and a 1/3 reduction of FDG dose at T2.

• Download figure
• Open in new tab
• Download powerpoint