Optimization and application of non-local means (NLM) filtering of SPECT/CT
Authors: Jiabei Zheng1, Se Young Chun2, Jeffrey A. Fessler1, and Yuni K. Dewaraja1. 1. University of
Michigan, Ann Arbor, USA; 2. Ulsan National Institute of Science and Technology, Ulsan, South
Korea
Objective: NLM filtering with high-quality side information (CT in our case) provides excellent image
denoising. We optimized NLM filtering for SPECT/CT dosimetry applications where accurate
quantification of total target activity and 3D activity distribution is of interest.
Method: SPECT and CT NLM weight maps were combined by multiplication (NLM-CTM) or by
summation (NLM-CTS). With XCAT simulation, we optimized filter parameters for SPECT (σf) and
CT (σa) by maximizing an objective function based on recovery coefficients (RC) and normalized root
mean squared errors (NRMSE) of spherical lesions. By scaling SPECT range to CT range when
forming weight maps, we removed count-rate dependency of the optimal parameters. The optimized
filters were applied to OSEM reconstructed I-131 SPECT of both experimental phantoms (with known
uniform or nonuniform activity in spheres) and patient studies. For phantom studies, σf was scaled to
account for differences in sphere and background CT numbers compared with XCAT. Results of
filtered and unfiltered OSEM were compared.
Result: In XCAT, RCs and NRMSEs calculated for different σa and σf had similar distributions for
different lesion sizes and count rates, making it possible to find two broadly applicable filters (optimal
NLM-CTM and optimal NLM-CTS). For high- and low-count cases, optimal filters increased RCs by
3.4% to 8.1% and reduced NRMSEs by 20.8% to 29.6%. The dose-rate-volume histograms (DVH) also
better matched truth. In the experimental phantom with uniform spheres, RCs increased by -0.7% to
8.9% and NRMSEs decreased by 14.6% to 37.2%. In the phantom with nonuniform spheres, RCs
remained unchanged while NRMSEs decreased by -1.5% to 14.1%. In the patient study, changes in
image quality and DVHs were observed.
Conclusion: We proposed two optimized NLM filters for SPECT/CT that were effective for a range of
count levels and lesion sizes and practical for clinical application. They improved estimates of both
total target activity and 3D activity distribution in simulated and experimental studies.
Optimal NLM filters also
improve RCs and NRMSEs
for measured data of a
phantom with 6 uniform
activity spheres. As shown for XCAT, distributions of RCs and NRMSEs under
different 𝜎! and 𝜎! are similar for different lesion sizes and
count levels. Low and high counts are chosen according to
pre-treatment tracer count and treatment count of patients. The
cross marks the universal optimal filter, which was applied to
all measured data. Recovery Coefficient
Method
Left
Center
Right
No filter
82.6% 96.3% 100.0%
NLM-CTM 81.5% 98.3%
99.7%
NLM-CTS 82.5% 98.6% 100.8%
Normalized RMSE
Method
Left
Center
Right
No filter
39.0% 47.9%
47.3%
NLM-CTM 39.6% 41.2%
42.8%
NLM-CTS 39.2% 42.3%
44.4%
Results of a measured phantom with 2 non-uniform activity
spheres. Improved NRMSEs are achieved without sacrificing
recovery.
CT (Patient)
Results of patient data.
Changes in DVH and
activity distribution were
observed. Proposed optimal NLM filters significantly improve both RCs
and NRMSEs for all 5 XCAT lesions for high- and low-count
I-131 SPECT images. Improvement of DVH is also observed.