Hossein Arabi1, Habib Zaidi2. 1. Division of Nuclear Medicine and Molecular Imaging, Department of Medical Imaging, Geneva University Hospital, CH-1211 Geneva 4, Switzerland. 2. Division of Nuclear Medicine and Molecular Imaging, Department of Medical Imaging, Geneva University Hospital, CH-1211 Geneva 4, Switzerland; Geneva Neuroscience Center, Geneva University, CH-1205 Geneva, Switzerland; Department of Nuclear Medicine and Molecular Imaging, University of Groningen, University Medical Center Groningen, 9700 RB Groningen, Netherlands; Department of Nuclear Medicine, University of Southern Denmark, DK-500 Odense, Denmark. Electronic address: habib.zaidi@hcuge.ch.
Abstract
PURPOSE: Attenuation correction (AC) is essential for quantitative PET imaging. In the absence of concurrent CT scanning, for instance on hybrid PET/MRI systems or dedicated brain PET scanners, an accurate approach for synthetic CT generation is highly desired. In this work, a novel framework is proposed wherein attenuation correction factors (ACF) are estimated from time-of-flight (TOF) PET emission data using deep learning. METHODS: In this approach, referred to as called DL-EM), the different TOF sinogram bins pertinent to the same slice are fed into a multi-input channel deep convolutional network to estimate a single ACF sinogram associated with the same slice. The clinical evaluation of the proposed DL-EM approach consisted of 68 clinical brain TOF PET/CT studies, where CT-based attenuation correction (CTAC) served as reference. A two-tissue class consisting of background-air and soft-tissue segmentation of the TOF PET non-AC images (SEG) as a proxy of the technique used in the clinic was also included in the comparative evaluation. Qualitative and quantitative PET analysis was performed through SUV bias maps quantification in 63 different brain regions. RESULTS: The DL-EM approach resulted in 6.1 ± 9.7% relative mean absolute error (RMAE) in bony structures compared to SEG AC method with RMAE of 16.1 ± 8.2% (p-value <0.001). Considering the entire head region, DL-EM led to a root mean square error (RMSE) of 0.3 ± 0.01 outperforming the SEG method with RMSE of 0.8 ± 0.02 SUV (p-value <0.001). The region-wise analysis of brain PET studies revealed less than 7% absolute SUV bias for the DL-EM approach, whereas the SEG method resulted in more than 14% absolute SUV bias (p-value <0.05). CONCLUSIONS: Qualitative assessment and quantitative PET analysis demonstrated the superior performance of the DL-EM approach over the segmentation-based technique with clinically acceptable SUV bias. The results obtained using the DL-EM approach are comparable to state-of-the-art MRI-guided AC methods. Yet, this approach enables the extraction of interesting features about patient-specific attenuation which could be employed not only as a stand-alone AC approach but also as complementary/prior information in other AC algorithms.
PURPOSE: Attenuation correction (AC) is essential for quantitative PET imaging. In the absence of concurrent CT scanning, for instance on hybrid PET/MRI systems or dedicated brain PET scanners, an accurate approach for synthetic CT generation is highly desired. In this work, a novel framework is proposed wherein attenuation correction factors (ACF) are estimated from time-of-flight (TOF) PET emission data using deep learning. METHODS: In this approach, referred to as called DL-EM), the different TOF sinogram bins pertinent to the same slice are fed into a multi-input channel deep convolutional network to estimate a single ACF sinogram associated with the same slice. The clinical evaluation of the proposed DL-EM approach consisted of 68 clinical brain TOF PET/CT studies, where CT-based attenuation correction (CTAC) served as reference. A two-tissue class consisting of background-air and soft-tissue segmentation of the TOF PET non-AC images (SEG) as a proxy of the technique used in the clinic was also included in the comparative evaluation. Qualitative and quantitative PET analysis was performed through SUV bias maps quantification in 63 different brain regions. RESULTS: The DL-EM approach resulted in 6.1 ± 9.7% relative mean absolute error (RMAE) in bony structures compared to SEG AC method with RMAE of 16.1 ± 8.2% (p-value <0.001). Considering the entire head region, DL-EM led to a root mean square error (RMSE) of 0.3 ± 0.01 outperforming the SEG method with RMSE of 0.8 ± 0.02 SUV (p-value <0.001). The region-wise analysis of brain PET studies revealed less than 7% absolute SUV bias for the DL-EM approach, whereas the SEG method resulted in more than 14% absolute SUV bias (p-value <0.05). CONCLUSIONS: Qualitative assessment and quantitative PET analysis demonstrated the superior performance of the DL-EM approach over the segmentation-based technique with clinically acceptable SUV bias. The results obtained using the DL-EM approach are comparable to state-of-the-art MRI-guided AC methods. Yet, this approach enables the extraction of interesting features about patient-specific attenuation which could be employed not only as a stand-alone AC approach but also as complementary/prior information in other AC algorithms.
Authors: Kerstin Kläser; Thomas Varsavsky; Pawel Markiewicz; Tom Vercauteren; Alexander Hammers; David Atkinson; Kris Thielemans; Brian Hutton; M J Cardoso; Sébastien Ourselin Journal: Med Image Anal Date: 2021-04-16 Impact factor: 8.545