From quantification to classification: comparative analysis of two software applications for machine learning–based prediction of early Parkinson’s disease using 123I-Ioflupane metrics

Purpose To compare semiquantitative striatal indices produced by two widely used quantification platforms — BasGanV2™ and NeuroTrans3D (Oasis©) — for ¹²³I-Ioflupane DAT SPECT, to evaluate their diagnostic performance for Parkinson’s disease (PD) versus non-degenerative presentations, and to determine whether supervised machine-learning (ML) classifiers can clarify relative utility of the two methods. Methods Retrospective analysis of 90 consecutive subjects (48 PD, 42 not-PD) undergoing routine ¹²³I-Ioflupane SPECT. Striatal binding ratios (caudate and putamen, left and right) were computed with both tools. Distributional differences were tested (Mann–Whitney U and Brown–Forsythe/Levene tests). Three supervised classifiers (Gaussian naïve Bayes, k-nearest neighbours, and SVM-RBF) were trained on four regional metrics from each tool using stratified shuffle-split (60:40 train: test) repeated 50 times; performance was summarised as mean (95% percentile CI) accuracy, sensitivity, specificity and AUC. Results Both quantification methods discriminated PD from not-PD (all regions p < 0.0001). NeuroTrans3D/Oasis© produced consistently higher mean binding values and significantly greater variance than BasGanV2™. ML models trained on either tool achieved excellent AUCs (0.917–0.960) and similar accuracy (≈ 86–91%), with overlapping confidence intervals and no single tool/classifier combination clearly outperforming the others. Conclusion BasGanV2™ and NeuroTrans3D (Oasis©) are comparably effective for distinguishing PD from non-degenerative cases when combined with straightforward ML classifiers, despite systematic differences in absolute values and dispersion between tools. ML aids comparison by quantifying discriminative performance but, in this dataset, does not indicate a clear superiority of one quantification pipeline over the other.