Impact of a combination of quantitative indices representing uptake intensity, shape, and asymmetry in DAT SPECT using machine learning: comparison of different volume of interest settings

Abstract

Background: We sought to assess the machine learning-based combined diagnostic accuracy of three types of quantitative indices obtained using dopamine transporter single-photon emission computed tomography (DAT SPECT)-specific binding ratio (SBR), putamen-to-caudate ratio (PCR)/fractal dimension (FD), and asymmetry index (AI)-for parkinsonian syndrome (PS). We also aimed to compare the effect of two different types of volume of interest (VOI) settings from commercially available software packages DaTQUANT (Q) and DaTView (V) on diagnostic accuracy.

Methods: Seventy-one patients with PS and 40 without PS (NPS) were enrolled. Using SPECT images obtained from these patients, three quantitative indices were calculated at two different VOI settings each. SBR-Q, PCR-Q, and AI-Q were derived using the VOI settings from DaTQUANT, whereas SBR-V, FD-V, and AI-V were derived using those from DaTView. We compared the diagnostic value of these six indices for PS. We incorporated a support vector machine (SVM) classifier for assessing the combined accuracy of the three indices (SVM-Q: combination of SBR-Q, PCR-Q, and AI-Q; SVM-V: combination of SBR-V, FD-V, and AI-V). A Mann-Whitney U test and receiver-operating characteristics (ROC) analysis were used for statistical analyses.

Results: ROC analyses demonstrated that the areas under the curve (AUC) for SBR-Q, PCR-Q, AI-Q, SBR-V, FD-V, and AI-V were 0.978, 0.837, 0.802, 0.906, 0.972, and 0.829, respectively. On comparing the corresponding quantitative indices between the two types of VOI settings, SBR-Q performed better than SBR-V (p = 0.006), whereas FD-V performed better than PCR-Q (p = 0.0003). No significant difference was observed between AI-Q and AI-V (p = 0.56). The AUCs for SVM-Q and SVM-V were 0.988 and 0.994, respectively; the two different VOI settings displayed no significant differences in terms of diagnostic accuracy (p = 0.48).

Conclusion: The combination of the three indices obtained using the SVM classifier improved the diagnostic performance for PS; this performance did not differ based on the VOI settings and software used.

Keywords: 123I-FP-CIT; 123I-Ioflupane; DAT SPECT; Machine learning; Parkinson’s syndrome; Support vector machine.

Fig. 1

Two types of VOI settings using DaTQUANT and DaTView. a DaTQUANT uses separated VOIs delineated to fit to the striatum. The reference VOIs are set to the bilateral occipital lobes. b DaTView uses large pentagonal prism-shaped VOIs encompassing a wide area around the striatum. The reference VOI is set to the whole brain. Abbreviation: VOI, volume of interest

Fig. 2

Box-and-whisker plots of the SBR-Q (a), PCR-Q (b), AI-Q (c), SBR-V (d), FD-V (e), and AI-V (f). There were significant differences in all six quantitative indices between NPS and PS, as assessed using the Mann-Whitney U test. Abbreviations: AI, asymmetry index; FD, fractal dimension; NPS, non-parkinsonian syndrome; PCR, putamen-to-caudate ratio; PS, parkinsonian syndrome; SBR, specific binding ratio

Fig. 3

ROC analysis of the three types of indices. a The AUCs for SBR-Q and SBR-V were 0.978 and 0.906, respectively. b The AUCs for PCR-Q and FD-V were 0.837 and 0.972, respectively. c The AUCs for AI-Q and AI-V were 0.802 and 0.829, respectively. Abbreviations: AI, asymmetry index; AUC, area under the curve; FD, fractal dimension; PCR, putamen-to-caudate ratio; ROC, receiver-operating characteristics; SBR, specific binding ratio

Fig. 4

Scatter plots of the PS (red triangles) and NPS (blue circles) cases. The blue planes are the boundary surfaces determined by the SVM classifier. The apostrophe represents the means after standardization. a The SVM-Q computed with SBR-Q, PCR-Q, and AI-Q. b The SVM-V computed with SBR-V, FD-V, and AI-V. Abbreviations: AI, asymmetry index; FD, fractal dimension; NPS, non-parkinsonian syndrome; PCR, putamen-to-caudate ratio; PS, parkinsonian syndrome; SBR, specific binding ratio; SVM, support vector machine

Fig. 5

Box-and-whisker plots and ROC analysis of SVM-Q and SVM-V. There were significant differences between NPS and PS in both the SVM-Q (a) and SVM-V (b), as assessed using a Mann-Whitney U test. c The AUCs for the SVM-Q and SVM-V were 0.988 and 0.994, respectively. Abbreviations: AUC, area under the curve; NPS, non-parkinsonian syndrome; PS, parkinsonian syndrome; ROC, receiver-operating characteristics; SVM, support vector machine