Program Information
Toward Fully 4D PET Image Reconstruction
I Häggström1*, Y Lin2 , A Krol3 , Y Xu2 , C Schmidtlein1 , (1) Mem. Sloan Kettering Cancer Ctr, New York, NY, (2) Sun Yat-sen University, Guangzhou, China, (3) SUNY Upstate Med. University, Syracuse, NY
Presentations
TH-EF-605-1 (Thursday, August 3, 2017) 1:00 PM - 3:00 PM Room: 605
Purpose: This study uses sparse representation methods to demonstrate the feasibility of fully 4D PET for dynamic PET imaging through 2D spatial + 1D temporal (3DT) simulations. Both perceptual and quantitative image quality metrics are used to evaluate improvements in ungated data.
Methods: Three different 3DT temporal and spatial penalties have been incorporated into a fixed-point proximity gradient algorithm for 3DT PET image reconstruction: 3DT higher order total variation (HOTV), and two patch based ones, discrete cosine transform (DCT), and tensor singular value decomposition (tSVD). These penalty functions were compared to two independent frame methods; 2D ordered subset expectation maximization (OSEM) and HOTV. Ten replicates of two dynamic phantoms were simulated; one brain phantom with kinetic uptake, and one moving cardiac/lung phantom. The dynamic PET data was reconstructed using all algorithms, and the agreement in terms of structural similarity index (SSIM) to ground truth was investigated. Image derived left ventricular (LV) volumes were also obtained, on ungated (0.1 s frames) and cardiac gated data (10 bins of 15x0.1 s frames each).
Results: For the brain phantom, the patch based 3DT-tSVD algorithm yielded optimal uptake images with a 25% higher SSIM compared to the standard 2D-OSEM (0.55 vs. 0.44), and a 14% lower relative RMSE (0.65 vs. 0.76). 3DT-DCT yielded optimal images for the cardiac/lung phantom with a 40% higher SSIM (0.62 vs. 0.44), and 26% lower RMSE (0.48 vs. 0.65), compared to 2D-OSEM. The LV volume relative true was on average 62, 53 and 19% for 3DT-DCT, gated 2D-OSEM, and 2D-OSEM, respectively.
Conclusion: DCT and tSVD patch based methods that include temporal penalties strongly improve dynamic PET image quality compared to standard 2D methods. In turn, this enables more robust and accurate LV volume extraction at higher temporal sampling without the need for reducing number of time bins through gating.
Funding Support, Disclosures, and Conflict of Interest: Partial support by MSK Cancer Center Support Grant/Core Grant P30 CA008748
Contact Email: