Bayesian Estimation of CBF Measured by DSC-MRI in Patients with Moyamoya Disease: Comparison with 15O-Gas PET and Singular Value Decomposition

S. Hara; Y. Tanaka; S. Hayashi; M. Inaji; T. Maehara; M. Hori; S. Aoki; K. Ishii; T. Nariai

doi:10.3174/ajnr.A6248

Article Figures & Data

Figures

Tables

Download figure
Open in new tab
Download powerpoint
Fig 1.
A representative case. A 48-year-old man was incidentally diagnosed with Moyamoya disease by nonspecific symptoms. MRA reveals occlusion of the terminal portion of the left ICA (arrows), bilateral anterior cerebral arteries, and the left MCA. FLAIR shows bilateral white matter hyperintensities. CBF maps obtained by PET, sSVD and Bayesian analysis shows clearly reduced CBF on the left sides (arrows). However, the CBF maps obtained by cSVD and oSVD do not show the laterality of the CBF that is evident on PET (arrows). The CBF% map of Bayesian analysis shows values close to the CBF% of PET. Please note that the color scale used in the raw CBF maps was the same as that used during the visual assessments and that the color scales of the CBF% maps were unified.
Download figure
Open in new tab
Download powerpoint
Fig 2.
The Bland-Altmann analysis between the regional raw CBF values of PET and each of the 4 DSC analyses. All methods showed moderate correlation between DSC and PET, and there was no significant difference between Bayesian analysis and the other SVD methods except for sSVD (versus cSVD,P = .11; versus oSVD, P = .37; versus sSVD, P = .04). sSVD overestimates PET-CBF with a small proportional bias; cSVD underestimates PET-CBF with moderate negative proportional bias; oSVD overestimates PET-CBF without proportional bias; and Bayesian-CBF strongly overestimates PET-CBF with a small proportional bias. LOA, limits of agreement defined as average ± 1.96 × SD. *P < .05 and **P < .005.
Download figure
Open in new tab
Download powerpoint
Fig 3.
The Bland-Altmann analysis between the regional CBF values normalized to cerebellum (CBF%) of PET and the 4 DSC analyses. The correlation of the Bayesian analysis was significantly better than those of the 3 SVD methods (versus sSVD, P = .02; versus cSVD, P < .001; versus oSVD, P < .001). The limit of agreement (LOA, defined as average ±1.96 x SD) of the Bayesian method was also smaller than those of the 3 SVD methods. *P < .05 and **P < .005.

Tables

Figures

Sensitivity, specificity, positive predictive value, and negative predictive value of the visual assessment of the 4 DSC-CBF maps^{^a}

(%)	sSVD	cSVD	oSVD	Bayesian
Sensitivity	62.5 (35.4–84.8)^{^b}	18.8 (4.0–45.6)	31.3 (11.0–58.7)^{^c}	62.5 (35.4–84.8)^{^b}
Specificity	86.4 (65.1–97.1)	86.4 (65.1–97.1)	90.9 (70.8–98.9)	100 (78.1–100)
Positive predictive value	76.9 (46.2–95.0)	50.0 (11.8–88.2)	71.4 (29.0–96.3)	100 (58.7–100)^{^c}
Negative predictive value	76.0 (54.9–90.6)	59.4 (40.6–76.3)	64.5 (45.4–80.8)	78.6 (59.0–91.7)^{^b}