A novel computer-assisted image analysis of [123I]β-CIT SPECT images improves the diagnostic accuracy of parkinsonian disorders

The purpose of this study was to develop an observer-independent algorithm for the correct classification of dopamine transporter SPECT images as Parkinson's disease (PD), multiple system atrophy parkinson variant (MSA-P), progressive supranuclear palsy (PSP) or normal. A total of 60 subjects with clinically probable PD (n = 15), MSA-P (n = 15) and PSP (n = 15), and 15 age-matched healthy volunteers, were studied with the dopamine transporter ligand [I-123]beta-CIT. Parametric images of the specific-to-nondisplaceable equilibrium partition coefficient (BPND) were generated. Following a voxel-wise ANOVA, cut-off values were calculated from the voxel values of the resulting six post-hoc t-test maps. The percentages of the volume of an individual BPND image remaining below and above the cut-off values were determined. The higher percentage of image volume from all six cut-off matrices was used to classify an individual's image. For validation, the algorithm was compared to a conventional region of interest analysis. The predictive diagnostic accuracy of the algorithm in the correct assignment of a [I-123]beta-CIT SPECT image was 83.3% and increased to 93.3% on merging the MSA-P and PSP groups. In contrast the multinomial logistic regression of mean region of interest values of the caudate, putamen and midbrain revealed a diagnostic accuracy of 71.7%. In contrast to a rater-driven approach, this novel method was superior in classifying [I-123]beta-CIT-SPECT images as one of four diagnostic entities. In combination with the investigator-driven visual assessment of SPECT images, this clinical decision support tool would help to improve the diagnostic yield of [I-123]beta-CIT SPECT in patients presenting with parkinsonism at their initial visit.