Fig. 1

Four distinct spatial patterns of FP-CIT PET images from unsupervised clustering of the unlabeled cohort and its impact on survival in the separate MSA cohort. a All FP-CIT PET images in the training cohort were clustered into four clusters by an unsupervised manner. The data distribution of the clusters was visualized using t-distributed stochastic neighbor embedding (t-SNE). b, c To characterize the regional uptake pattern of the clusters, the same data were mapped based on binding ratios (BRs) of the caudate nucleus (b) or putamen (c). d All normalized FP-CIT PET images were averaged into a single representative image for each cluster. e To delineate the regional abnormality pattern of clusters, a relative decrease map was produced in each cluster using cluster 2 as a reference, and presented on an MRI template. f Survival outcomes of the four clusters in the separate MSA cohort were demonstrated using Kaplan–Meier curve analysis