(G-251) ISIT-DAT: An in silico imaging trial to objectively evaluate the performance of a deep-learning-based attenuation compensation method for dopamine transporter SPECT

Introduction:: Parkinson’s disease (PD) is a prevalent neurodegenerative disorder that impacts millions of individuals worldwide [1]. Abnormalities in dopamine transporter (DaT) concentration within the caudate, putamen, and globus pallidus (GP) regions are being investigated as indicative of neurodegeneration in patients with PD [2]. Single-photon emission computed tomography (SPECT) is a widely used clinical imaging technique that allows for quantifying DaT uptake in these regions [3]. However, accurate quantification is challenging in DaT-scan SPECT due to various image-degrading effects in SPECT, including limited resolution, attenuation, scatter, collimator-detector response, and noise.

Attenuation compensation (AC) methods have been developed to mitigate the effects of attenuation in SPECT. Typically, these methods require an attenuation map obtained from a computed tomography (CT) scan. However, the use of a separate CT scan introduces additional radiation exposure, imaging costs, and patient discomfort. Misalignment between the SPECT and CT scans can lead to errors in quantification [4]. Moreover, a significant portion of the SPECT market consists of SPECT-only systems [5], including emerging solid-state-detector-based SPECT systems [6].

Various transmission-less AC (Tx-less AC) methods for SPECT have been proposed [7-10], including deep learning (DL) techniques [10-15]. In this study, we proposed a deep learning-based attenuation compensation method for DaT-scan SPECT. We conducted an in silico imaging trial, namely ISIT-DAT, to evaluate the proposed method on fidelity-based figures of merit (FoM) and a clinical task of quantifying DaT uptake in the caudate, putamen, and GP regions. The evaluation study demonstrated the effectiveness of the proposed method for transmission-less AC in DaT-scan SPECT.

Materials and Methods:: The proposed method, DaT-SLAC, first obtains an initial estimate of the attenuation map from scatter-window projection using an ordered subsets expectation maximization (OSEM)-based technique. This initial estimate is then segmented into different regions. The segmented regions are assigned pre-defined attenuation coefficient values, yielding an attenuation map used in AC.

We collected images from N = 220 3T T1-weighted MRI and CT studies. MR images were registered with the corresponding CT scans and segmented into caudate, putamen, GP, and other soft tissues, yielding anatomical templates of the brain. Realistic DaT activity uptakes were assigned to these regions. CT images were then segmented into soft tissue, skull, scalp with head holder, and background. Projections in the photopeak and scatter windows were collected by a realistically simulated 3D SPECT system using SIMIND [16]. The projections were scaled to the clinical count level and associated with Poisson noise.

We used 175 pairs of initial attenuation estimations and corresponding segmented CT images for the network training, and the rest for testing. We evaluated DaT-SLAC through both proof-of-concept (POC) and task-specific evaluations, of which the workflow shown in Fig. 1. DaT-SLAC was compared with the standard CT-based (CTAC) method and a uniform-attenuation-map-based AC (UAC) method. In the POC evaluation, we compared the estimated attenuation maps and the reconstructed activity images with those obtained by CTAC using normalized RMSE and SSIM. In ISIT-DAT, we estimated regional uptake in the caudate, putamen, and GP regions obtained by DaT-SLAC and compared with the true uptake using normalized RMSE.

Results, Conclusions, and Discussions:: Table I shows the normalized RMSE and SSIM between the estimated attenuation maps and the reconstructed activity maps obtained by DaT-SLAC and CTAC, as well as between UAC and CTAC. DaT-SLAC method achieved lower normalized RMSE and higher SSIM than UAC, showing higher fidelity-based similarity to the CTAC method.

Table I. Evaluation results on fidelity-based FoMs

	Methods	Normalized RMSE	SSIM
Activity maps	DaT-SLAC	0.3929 (0.3872, 0.3986)	0.9725 (0.9710, 0.9739)
	UAC	0.4952 (0.4882, 0.5022)	0.9648 (0.9632, 0.9664)
Attenuation maps	DaT-SLAC	0.3916 (0.3840, 0.3991)	0.7889 (0.7809, 0.7969)
	UAC	0.6265 (0.6160, 0.6369)	0.6460 (0.6398, 0.6523)

The mean activity uptake from the caudate, putamen, and GP regions yielded by DaT-SLAC, CTAC, and UAC approach are shown in Fig. 2. DaT-SLAC performed similarly to CTAC approach and outperformed the UAC approach.

In ISIT-DAT, we conducted an in silico imaging trial to objectively evaluate the performance of DaT-SLAC on a clinical task of quantifying DaT uptake in the caudate, putamen, and GP regions. The results of the evaluations demonstrated that DaT-SLAC has a similar performance to CTAC method. Moreover, DaT-SLAC demonstrated superior performance compared to UAC. Further evaluation with clinical data and extension to different scanner configurations and SPECT applications are motivated by these results.

In conclusion, we objectively evaluated a scatter-window projection and deep learning-based method to perform transmission-less AC for DaT-scan SPECT. The method yielded similar performance as that obtained using a CT-based AC approach on the task of quantifying regional activity uptake. Further, the method outperformed an AC approach that uses a uniform attenuation map.

Acknowledgements (Optional): : This work was supported by National Institute of Biomedical Imaging and Bioengineering of National Institute of Health (NIH) under grant numbers R21-EB024647, R01-EB031051, and R01EB031962. Support is also acknowledged from the NVIDIA GPU grant. We also thank the Washington University Center for High Performance Computing for providing computational resources. The center is partially funded by NIH grants 1S10RR022984-01A1 and 1S10OD018091-01.

References (Optional): : [1] E. Dorsey, T. Sherer, M. S. Okun, and B. R. Bloem, "The emerging evidence of the Parkinson pandemic," J. Parkinsons Dis., vol. 8, no. s1, pp. S3-S8, 2018.
[2] B. Ciliax et al., "The dopamine transporter: immunochemical characterization and localization in brain," The Journal of Neuroscience, vol. 15, no. 3, pp. 1714-1723, 1995, doi: 10.1523/jneurosci.15-03-01714.1995.
[3] I. Gayed et al., "The impact of DaTscan in the diagnosis of Parkinson disease," Clin. Nucl. Med., vol. 40, no. 5, pp. 390-393, 2015.
[4] M. Sakoshi, N. Matsutomo, T. Yamamoto, and E. Sato, "Effect of misregistration between SPECT and CT images on attenuation correction for quantitative bone SPECT imaging," Nihon Hoshasen Gijutsu Gakkai Zasshi, vol. 74, no. 5, pp. 452-458, 2018.
[5] "Technavio. (2021) Global SPECT market 2017-2021." https://www.technavio.com/report/global-medicalimaging-global-spect-market-2017-2021 (accessed.
[6] K. Tsuchiya et al., "Basic performance and stability of a CdTe solid-state detector panel," Ann. Nucl. Med., vol. 24, no. 4, pp. 301-311, 2010.
[7] T. S. Pan, M. A. King, D. S. Luo, S. T. Dahlberg, and B. J. Villegas, "Estimation of attenuation maps from scatter and photopeak window single photon-emission computed tomographic images of technetium 99m-labeled sestamibi," (in English), J. Nucl. Cardiol., vol. 4, no. 1, pp. 42-51, Jan-Feb 1997, doi: Doi 10.1016/S1071-3581(97)90048-9.
[8] M. Nunez, V. Prakash, R. Vila, F. Mut, O. Alonso, and B. F. Hutton, "Attenuation correction for lung SPECT: evidence of need and validation of an attenuation map derived from the emission data," Eur. J. Nucl. Med. Mol. Imaging, vol. 36, no. 7, pp. 1076-89, Jul 2009, doi: 10.1007/s00259-009-1090-4.
[9] H. Zaidi and B. Hasegawa, "Determination of the attenuation map in emission tomography," J. Nucl. Med., vol. 44, no. 2, pp. 291-315, Feb 2003. [Online]. Available: https://www.ncbi.nlm.nih.gov/pubmed/12571222.
[10] L. Shi, J. A. Onofrey, H. Liu, Y.-H. Liu, and C. Liu, "Deep learning-based attenuation map generation for myocardial perfusion SPECT," Eur. J. Nucl. Med. Mol. Imaging, vol. 47, no. 10, pp. 2383-2395, 2020/09/01 2020, doi: 10.1007/s00259-020-04746-6.
[11] X. Chen et al., "CT-free attenuation correction for dedicated cardiac SPECT using a 3D dual squeeze-and-excitation residual dense network," J. Nucl. Cardiol., vol. 29, no. 5, pp. 2235-2250, 2022.
[12] X. Chen et al., "Direct and indirect strategies of deep-learning-based attenuation correction for general purpose and dedicated cardiac SPECT," Eur. J. Nucl. Med. Mol. Imaging, vol. 49, no. 9, pp. 3046-3060, 2022/07/01 2022, doi: 10.1007/s00259-022-05718-8.
[13] H. Liu et al., "Post-reconstruction attenuation correction for SPECT myocardium perfusion imaging facilitated by deep learning-based attenuation map generation," J. Nucl. Cardiol., pp. 1-12, 2021.
[14] A. D. Shanbhag et al., "Deep Learning-based Attenuation Correction Improves Diagnostic Accuracy of Cardiac SPECT," J. Nucl. Med., p. jnumed.122.264429, 2022, doi: 10.2967/jnumed.122.264429.
[15] T. Hagio et al., "“Virtual” attenuation correction: improving stress myocardial perfusion SPECT imaging using deep learning," Eur. J. Nucl. Med. Mol. Imaging, vol. 49, no. 9, pp. 3140-3149, 2022/07/01 2022, doi: 10.1007/s00259-022-05735-7.
[16] M. T. Bahreyni Toossi, J. P. Islamian, M. Momennezhad, M. Ljungberg, and S. H. Naseri, "SIMIND Monte Carlo simulation of a single photon emission CT," (in eng), J. Med. Phys., vol. 35, no. 1, pp. 42-47, 2010, doi: 10.4103/0971-6203.55967.