Fig. 4

Comparison of ground truth annotations and CNN detection with expert scoring. A Expert annotation distribution boxplot in test WSI cohort (n = 56), Mann–Whitney two-tailed U test between score groups p values; Kruskal–Wallis H test of annotated LTS between expert score groups, and Spearman correlation between LTS burden and expert scores. B CNN, 40 × 40 patches positive for LTS distribution boxplot in test WSI cohort (n = 56), Mann–Whitney two-tailed U test between score groups p values; Kruskal–Wallis H test of 40 × 40 patches between expert score groups, and Spearman correlation between 40 × 40 patches positive for LTS burden and a score in test cohort. Scoring was performed as follows: each WSI was classified as positive or negative for α-synuclein pathology and assigned an LTS score ranging from 0 to 3 (Fig. 1), where 0 refers to being negative for α-synuclein, and scores 1–3 refer to scoring density of sparse (1) moderate (2), and frequent (3). CNN convolutional neural network, SMG submandibular gland, LTS Lewy-type synucleinopathy