Machine learning to identify suitable boundaries for band-pass spectral analysis of dynamic [ 11 C]Ro15-4513 PET scan and voxel-wise parametric map generation.

Background: Spectral analysis is a model-free PET quantification technique that treats the time-space signal as an impulse response to a bolus injection. Band-pass spectral analysis, considering specific frequency ranges, enables calculation of separate parametric maps of receptor subtype tracer binding for suitable radiopharmaceuticals such as [ $_{}^{11}$ C]Ro15-4513 binding to GABA_A $\alpha$ 1/5 subunits. Frequency ranges are based on inspection of spectra, prior knowledge of receptor distribution, and blocking studies. The process currently requires the manual selection of frequency ranges based on the data. To enhance the efficiency of band-pass spectral analysis and extend its application to a broader range of tracers, we propose employing machine learning to automate the selection of spectral boundaries. Based on these boundaries, voxel-wise parametric maps can be generated. The machine learning models utilized in this study include 1D Convolutional Neural Network, Neural Network, Support Vector Machine, Logistic Regression, K-nearest neighbors, and Fine Tree.

Results: The best machine learning model, Fine Tree, agreed with the manual frequency boundary in 96.92% of 3185 ROIs. The absolute mean error was 3.80% for slow component volume-of-distribution ( ${\text{V}}_{\mathrm{slow}}$ , largely representing $\alpha$ 5) and 4.74% for fast component volume-of-distribution( ${\text{V}}_{\mathrm{fast}}$ , largely representing $\alpha$ 5), while the relative error was 2.83% ± 43.47% for ${\text{V}}_{\mathrm{slow}}$ and $-$ 2.01% ± 78.04% for ${\text{V}}_{\mathrm{fast}}$ . The median test-retest intraclass correlation coefficient across six representative regions was 0.770 for ${\text{V}}_{\mathrm{slow}}$ , 0.670 for ${\text{V}}_{\mathrm{fast}}$ , and 0.502 for total component volume-of-distribution( ${\text{V}}_d$ ). Parametric maps applying different boundaries for different ROIs were generated.

Conclusion: The machine learning model developed provided accurate boundary predictions in 96.92% of regions, with minimal average bias. However, when errors occur, they can be large, owing to the sparsity of peaks. The model enables setting boundaries automatically for the vast majority of regions, followed by manual checking of the outliers. It opens the possibility of accelerating analyses e.g. of GABA_A $\alpha$ 1/2/3/5 subunit binding using [¹¹C]flumazenil and of extending band-pass spectral analysis to other receptor systems.