Non-invasive regional cerebral blood flow quantification in the 123I-IMP autoradiography using artificial neural network

Abstract

Purpose: Regional cerebral blood flow (rCBF) quantification using 123I-N-isopropyl-p-iodoamphetamine (123I-IMP) requires an invasive, one-time-only arterial blood sampling for measuring the 123I-IMP arterial blood radioactivity concentration (Ca10). The purpose of this study was to estimate Ca10 by machine learning (ML) using artificial neural network (ANN) regression analysis and consequently calculating rCBF and cerebral vascular reactivity (CVR) in the dual-table autoradiography (DTARG) method.

Materials and methods: This retrospective study included 294 patients who underwent rCBF measurements through the 123I-IMP DTARG. In the ML, the objective variable was defined by the measured Ca10, whereas the explanatory variables included 28 numeric parameters, such as patient characteristic values, total injection 123I-IMP radiation dose, cross-calibration factor, and the distribution of 123I-IMP count in the first scan. ML was performed with training (n = 235) and testing (n = 59) sets. Ca10 was estimated in testing set by our proposing model. Alternatively, the estimated Ca10 was also calculated via the conventional method. Subsequently, rCBF and CVR were calculated using estimated Ca10. Pearson's correlation coefficient (r-value) for the goodness of fit and the Bland-Altman analysis for assessing the potential agreement and bias were performed between the measured and estimated values.

Results: The r-value of Ca10 estimated by our proposed model was higher compared with the conventional method (0.81 and 0.66, respectively). In the Bland-Altman analysis, mean differences of 4.7 (95% limits of agreement (LoA): -18-27) and 4.1 (95% LoA: -35-43) were observed using proposed model and the conventional method, respectively. The r-values of rCBF at rest, rCBF after the acetazolamide challenge, and CVR calculated using the Ca10 estimated by our proposed model were 0.83, 0.80 and 0.95, respectively.

Conclusion: Our proposed ANN-based model could accurately estimate the Ca10, rCBF, and CVR in DTARG. These results would enable non-invasive rCBF quantification in DTARG.

Fig 1. Participant flowchart.

Fig 2. Dual-table autoradiography procedures.

Note. SPECT, Single-photon emission computed tomography.

Fig 3. The regions of interest identified by the NEURO FLEXER software.

Note. R, right; L, left; A, anterior; P, posterior; 1, hemisphere; 2, anterior cerebral artery; 3, middle cerebral artery (MCA); 4, anterior part of MCA; 5, posterior part of MCA; 6, posterior cerebral artery; 7, basal ganglia; 8, thalamus; 9, pons; 10, vermis; 11, cerebellum.

Fig 4. Our proposed artificial neural network model consisted of a 28-item input layer and 3 hidden layers with 3, 32, and 9 nodes, respectively.

The activation functions defined the log-sigmoid transfer function between hidden layers 1 and 2, the hyperbolic tangent function between hidden layers 2 and 3, and the log-sigmoid transfer function between hidden layer 3 and the output layer.

Fig 5

(a) Scatterplot with regression line between the true measured and estimated Ca10 by our proposed model revealed Pearson’s r-value of 0.81. (b) Bland–Altman plots of the differences versus the mean values between the measured and estimated Ca10 revealed that the short-dashed lines denote the 95% limits of agreement, and these values had a small and balanced dissemination. (c) Scatterplot with regression line between the true measured and estimated Ca10 by conventional method revealed Pearson’s r-value of 0.66. (d) Bland–Altman plots of the differences versus the mean values between the measured and estimated Ca10 revealed that these values had a relatively large and scattered dissemination.

Fig 6. Scatterplot with regression line between the true measured and estimated rCBF at rest (a) by our proposed model revealed Pearson’s r-value of 0.83 and (d) by conventional method revealed Pearson’s r-value of 0.68.

Scatterplot with regression line between the true measured and estimated rCBF after acetazolamide challenge (b) by our proposed model revealed Pearson’s r-value of 0.80 and (e) by conventional method revealed Pearson’s r-value of 0.63. Scatterplot with regression line between the true measured and estimated CVR (c) by our proposed model revealed Pearson’s r-value of 0.95 and (f) by conventional method revealed Pearson’s r-value of 0.90.