A diagnostic strategy for Parkinsonian syndromes using quantitative indices of DAT SPECT and MIBG scintigraphy: an investigation using the classification and regression tree analysis

Abstract

Purpose We aimed to evaluate the diagnostic performances of quantitative indices obtained from dopamine transporter (DAT) single-photon emission computed tomography (SPECT) and 123I-metaiodobenzylguanidine (MIBG) scintigraphy for Parkinsonian syndromes (PS) using the classification and regression tree (CART) analysis. Methods We retrospectively enrolled 216 patients with or without PS, including 80 without PS (NPS) and 136 with PS [90 Parkinson’s disease (PD), 21 dementia with Lewy bodies (DLB), 16 progressive supranuclear palsy (PSP), and 9 multiple system atrophy (MSA). The striatal binding ratio (SBR), putamen-to-caudate ratio (PCR), and asymmetry index (AI) were calculated using DAT SPECT. The heart-to-mediastinum uptake ratio (H/M) based on the early (H/M [Early]) and delayed (H/M [Delay]) images and cardiac washout rate (WR) were calculated from MIBG scintigraphy. The CART analysis was used to establish a diagnostic decision tree model for differentiating PS based on these quantitative indices. Results The sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 87.5, 96.3, 93.3, 92.9, and 93.1 for NPS; 91.1, 78.6, 75.2, 92.5, and 83.8 for PD; 57.1, 95.9, 60.0, 95.4, and 92.1 for DLB; and 50.0, 98.0, 66.7, 96.1, and 94.4 for PSP, respectively. The PCR, WR, H/M (Delay), and SBR indices played important roles in the optimal decision tree model, and their feature importance was 0.61, 0.22, 0.11, and 0.05, respectively. Conclusion The quantitative indices showed high diagnostic performances in differentiating NPS, PD, DLB, and PSP, but not MSA. Our findings provide useful guidance on how to apply these quantitative indices in clinical practice.

Explainable AI to improve acceptance of convolutional neural networks for automatic classification of dopamine transporter SPECT in the diagnosis of clinically uncertain parkinsonian syndromes

“…However, quantitative analysis and/ or automatic classification is a useful adjunct when used as an objective second reader, particularly in borderline cases and for less experienced readers [32]. Conventional machine learning methods using support vector machines [33][34][35][36][37][38][39][40][41][42][43], decision trees [44,45], or cluster analyses [46] based on a (small) set of predefined image-derived features have been proposed for this purpose. However, recent work suggests that artificial neural networks, particularly deep CNN, outperform conventional approaches for the automatic classification of DAT-SPECT [18,[47][48][49][50][51][52][53][54][55][56][57][58], partly because artificial neural networks can be less sensitive to camera-and site-specific variability of image quality (e.g., with respect to spatial resolution) [18].…”

Single Photon Emission Computed Tomography/Positron Emission Tomography Molecular Imaging for Parkinsonism: A Fast‐Developing Field

“…A linear SVM model based on voxel values of statistical parametric images differentiated PD from vascular parkinsonism with 90.4% accuracy 59 . Both Nicastro et al and Iwabuchi et al were able to differentiate between the different forms of atypical parkinsonism by respectively using an SVM model or a classification and regression tree analysis 60,61 . Both approaches were based on FP‐CIT uptake in caudate and putamen volumes of interest.…”

“…CBD patients showed a moderate reduction in uptake and a higher asymmetrical index 60 . DLB was associated with a lower level of impaired putamen to caudate ratio compared to the other forms of parkinsonism 61 . A fully automated artificial neural network based on deep‐learning analyses of dopaminergic imaging has also been shown to diagnose PD with accuracies of >90%.…”

“… 59 Both Nicastro et al and Iwabuchi et al were able to differentiate between the different forms of atypical parkinsonism by respectively using an SVM model or a classification and regression tree analysis. 60 , 61 Both approaches were based on FP‐CIT uptake in caudate and putamen volumes of interest. CBD was correctly diagnosed with accuracies of up to 83.7% and PSP with accuracies of up to 94.4%.…”

“… 60 DLB was associated with a lower level of impaired putamen to caudate ratio compared to the other forms of parkinsonism. 61 A fully automated artificial neural network based on deep‐learning analyses of dopaminergic imaging has also been shown to diagnose PD with accuracies of >90%. This is similar to accuracies obtained by experienced physicians.…”