RT Journal Article SR Electronic T1 Accelerated SPECT Image Reconstruction with a Convolutional Neural Network JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 1351 OP 1351 VO 60 IS supplement 1 A1 Martijn Dietze A1 Woutjan Branderhorst A1 Max Viergever A1 Hugo de Jong YR 2019 UL http://jnm.snmjournals.org/content/60/supplement_1/1351.abstract AB 1351Purpose: Image-guided procedures involving radionuclides, such as hepatic radioembolization, would benefit from the availability of SPECT/CT during intervention by gaining direct feedback on the activity distribution. To ensure a smooth workflow for such interventional scanning, three objectives should be achieved: i) a SPECT/CT scanner is required in the intervention room to eliminate patient transportation; ii) scanning should be performed quickly; and iii) reconstructions should be of high quality and obtained within minutes. We have proposed a mobile compact device (1) that can retrieve accurate quantitative measures with substantially reduced scan times compared to clinical practice (2). This leaves reconstruction time as the final hurdle. Filtered backprojection (FBP) is able to generate images within seconds, but lacks in quality due to high noise levels and possible artifacts. Iterative reconstructors, combined with model-based corrections, retrieve results closer to the true activity distribution, but need substantially more time to generate their results. We propose to use a convolutional neural network to upgrade FBP image quality, so that high quality reconstructions can be obtained within seconds. Methods: SPECT/CT scans of 129 99mTc-MAA hepatic radioembolization pre-treatment procedures were available. Projections were first reconstructed using a validated Monte Carlo-based iterative reconstructor and the obtained reconstructions were set as ground truth. Out of the ground truth reconstructions, projections were simulated for 20, 5, and 1 min scan time. These projections were then again reconstructed, now using the following methods: i) FBP; ii) current clinical reconstruction; iii) Monte Carlo-based reconstruction; and iv) using the neural network. The neural network was trained using 100 FBP reconstructions with the associated ground truth distributions. Quantitative accuracy was determined by calculating the mean squared error (MSE) on the validation set and the diagnostic value of the images was furthermore assessed by means of an observer study. Results: Reconstructions of a representative patient distribution are shown in the accompanying figure. FBP generated reconstructions in 2 seconds and the neural network required 5 seconds. The clinical reconstructor took 5 minutes and the Monte Carlo-based reconstructor 19 minutes. FBP created reconstructions with severe artifacts and tumor contrast was low. Clinical reconstruction had no visible artifacts, but also had low contrast. The neural network provided images comparable to the Monte Carlo-based reconstruction. Quantitatively, the MSE of the neural network reconstructions was between that of the Monte Carlo-based and the clinical reconstructions, for all scan times. The observer study showed that full diagnostic information was retrieved for 3.8% of the FBP reconstructions, 45.0% of the clinical reconstructions, 93.8% of the neural network reconstructions, and 98.8% of the Monte Carlo-based reconstructions. Conclusions: Enhancement of FBP-based reconstruction by means of a convolutional neural network generates high-quality reconstructions within seconds, which paves the way for use of interventional SPECT/CT in time critical clinical practice. References 1. Van der Velden S, Kunnen B, Koppert WJC, et al. A Dual Layer Detector for Simultaneous Fluoroscopic and Nuclear Imaging. Radiology. 2019. 2. Van der Velden S, Dietze MMA, Viergever MA, de Jong HWAM. Fast technetium-99m liver SPECT for evaluation of the pretreatment procedure for radioembolization dosimetry. Med Phys. 2018.