Abstract
P1257
Introduction: Whole-body dynamic PET with multi-pass protocols enables parametric imaging for scanners with short axial field of view (FOV). The discontinuous data collection and inherent low sensitivity nature lead to low-count statistics for each bed, resulting in noisy parametric images with the conventional indirect voxel-based Patlak analysis. Direct parametric image reconstruction alleviates the image noise issue, which however, is complex in algorithm design and computationally expensive. In the past, we have proposed a deep progressive reconstruction (DPR) algorithm to achieve high image quality even for a low-dose static PET study. Neural networks trained with uEXPLORER data were embedded in the DPR algorithm for image denoising and enhancement. In this work, we applied the DPR algorithm to the whole-body indirect parametric imaging. The objective is to quantitatively evaluate the noise reduction performance of the proposed method.
Methods: Four participants (3/1 M/F, 57-67 yrs, 53-75 kg) from Xijing Hospital (Xi’an, China) underwent 60-min whole-body multi-pass dynamic 18F-FDG scans (6.4-9.5 mCi) on the uMI Panorama PET/CT system. Each scan started with a 10-min single-bed scan centered at the cardiac region, followed by a multi-pass scan (referred to as the full-dose scan): 5 pass × 4 bed × 90 sec and 1 pass × 4 bed × 300 sec. All the data of the cardiac bed were reconstructed to generate image-derived input function (IDIF) from the descending aorta region. Low-dose scans at 1/2 and 1/3 dose were simulated by down-sampling the list-mode data for each bed in the multi-pass scan, as shown in Fig 1. IDIF was also extracted from the low-dose reconstructions of the cardiac bed.
In our proposed method, the dynamic images were reconstructed with DPR and then were processed with Patlak model (t* = 10 min). The proposed method (referred to as DL Indirect) was compared with the conventional indirect Patlak analysis (OSEM Indirect) and direct Patlak reconstruction (Direct). Nested algorithm with 10 nested iterations was used to accelerate the convergence in the Direct method. Regions of interest (ROI) including thalamus, cerebellum, myocardium, liver, and lesions (N = 32) were drawn manually to quantitatively evaluate the target to background ratio (TBR), coefficient of variation (CoV), and contrast to noise ratio (CNR) of Ki values.
Results: For the full-dose study, excellent correlation (R2 > 0.97) of Ki values for all the ROIs and all the participants was shown between the DL Indirect and the OSEM Indirect methods, as well as between the Direct and the OSEM Indirect methods (Fig 2). Lower CoV value in the liver region was shown in the Ki image with DL Indirect (Ki CoV = 0.29) than that with OSEM Indirect (0.56), and higher TBR value in the lesion region was shown in the Ki image with DL Indirect (Ki TBR = 6.5) than that with Direct method (3.8). Even for the 1/2 dose and 1/3 dose study, DL Indirect remained lower noise as compared with the OSEM Indirect method and higher contrast as compared with the Direct method (Fig 3). In comparison of CNR values for all the lesions (Fig 4), DL Indirect showed the comparable or slightly higher CNR compared with the Direct method for both full-dose and low-dose studies. Comparable lesion CNR was found between the DL Indirect with 1/3 dose (19.8 ± 12.3) and OSEM Indirect with full dose (18.8 ± 12.9).
Conclusions: We proposed a deep learning-based reconstruction algorithm for whole-body Patlak parametric imaging denoising. The proposed DL Indirect method showed lower noise (average 38% for 1/3 dose liver CoV) compared with the OSEM Indirect method, and higher contrast (average 30% for 1/3 dose lesion TBR) compared with the Direct method, and overall the best CNR. The DL Indirect method with only 1/3 dose yielded similar noise level as compared to the full-dose OSEM Indirect method (CoV = 0.59 vs. 0.56).
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.