Reducing acquisition time in pediatric whole-body PET imaging with deep learning enhanced image reconstruction

Introduction: Deep learning-based methods have been widely applied in nuclear medicine and have been recently demonstrated to successfully reduce PET image noise while maintaining spatial resolution. These developments provide opportunities relevant to pediatric imaging, where dose and/or acquisition time reduction are of particular interest. Our study aims to train and evaluate a patch-based artificial neural network (ANN) model to reconstruct reduced-count PET data from a state-of-the-art digital PET/CT system into images of quality comparable to the full-count clinical standard images.

Methods: In this study, we retrospectively reprocessed data for twenty-one pediatric patients (age 9.9 ± 5.4 yr and weight 37.0 ± 24.3 kg) who had undergone whole-body ¹⁸F-FDG (0.12 ± 0.01 mCi/kg) examinations on a GE Discovery MI Gen 2 PET/CT scanner using a 90 sec/bed standard-of-care acquisition protocol. The list-mode time-of-flight data for each patient were trimmed to simulate an acquisition time of 30 sec/bed. The full 90 sec/bed data were reconstructed using manufacturer standard regularized reconstruction (Q.Clear) with a regularization parameter (RP) of 700. The trimmed 30 sec/bed data were reconstructed with three RPs (700, 1000, and 1300). We built a fully-connected patch-based ANN model and trained it with image patches (size 4×4×4, maximum overlapping) from the three 30 sec/bed image sets of one randomly selected patient as the input. Image patches from the 90 sec/bed image of the same patient served as the target output. The trained ANN model was then tested using the 30 sec/bed image sets from the other 20 patients to generate an enhanced image for each patient. Quantitative and qualitative evaluations were performed with the 90 sec/bed images serving as the reference standard. For each testing patient, normalized mean square errors (NMSEs) were calculated for the 30 sec/bed image sets and the ANN generated images to quantify the difference with respect to the 90 sec/bed image. Standardized uptake values (SUVs) were also measured in three regions of interest (ROIs) including the liver, thigh, and largest tumor focus. Dunn’s test was used to compare SUV_mean, SUV_max, and SUV_SD/SUV_mean from the 30 sec/bed and the ANN generated images with the 90 sec/bed image. Independent, blinded, qualitative reviews of the images were performed by two radiologists (R1 and R2) who characterized any preference between the ANN generated and the 90 sec/bed standard image pairs for overall image quality and lesion conspicuity.

Results: Average NMSE values (×10^-5) for the 20 patients were 19.7 ± 7.1, 12.2 ± 4.1, and 9.2 ± 2.9 for the 30 sec/bed with RPs of 700, 1000, and 1300 images, respectively and 4.9 ± 1.9 for the ANN generated images, demonstrating increased similarity to the 90 sec/bed standard image. No significant difference was found in SUV_mean values between the 90 sec/bed and any of the 30 sec/bed images or the ANN generated images. Despite significant differences in SUV_max and SUV_SD/SUV_meanvalues between the 90 sec/bed and all three 30 sec/bed images (p < 0.05), the ANN generated images showed no significant differences from the 90 sec/bed images for any of the three ROIs indicating the capability of the proposed method in noise reduction and resolution preservation. There was no qualitative image quality difference for 43% (40% from R1 and 45% from R2) of the image pairs. For the subset with image quality differences, the 90 sec/bed images were preferred in 74% (73% from R1 and 75% from R2) whereas the ANN images were preferred in 26% (27% from R1 and 25% from R2) pairs. There was no qualitative difference in lesion conspicuity between any image pair.

Conclusions: For pediatric patients imaged on a state-of-the-art digital PET/CT system, ANN-enhanced 30 sec/bed images are quantitatively and qualitatively similar to 90 sec/bed images. The proposed method has the potential to reduce scan time (or dose) by 2/3 without significantly affecting the image quality or quantitative measures.

• Download figure
• Open in new tab
• Download powerpoint