Figure 4.

Evaluation of the annotation accuracy of SegPath

(A) Annotation accuracy of pathologists and the IF-masks in SegPath (n = 20 patches of 217.5 × 217.5 μm for each tissue or cell type) compared with pGT as ground truth. Dice coefficients of the IF-masks were compared with those of annotations by each pathologist. Two-sided Wilcoxon signed-rank test was used, and p values were adjusted using the Benjamini-Hochberg method. p < 0.05 was considered statistically significant, as shown by asterisks. See also Figures S4 and S5; Table S5.

(B) Ground truth (pGT) cell images annotated by multiple pathologists (pGT⁺P⁺M⁺⁻) and not identified by multiple pathologists but successfully annotated by the masks (pGT⁺P⁻M⁺) in the ten patches. The illustrations and the actual images of the representative cell morphologies and sentences describing the morphologies written in a histology textbook are shown in each cell type.²³ Original illustrations from BioRender were used except for the lymphocyte, whose nucleus was denser and larger than the original illustration. The image was adjusted to be more similar to the representative morphology. For the box plot, the lower and upper hinges correspond to the 25^th and 75^th percentiles, respectively; the upper whisker extends from the hinge to the largest value no further than 1.5× interquartile range (IQR) from the hinge. The lower whisker extends from the hinge to the smallest value at 1.5× IQR of the hinge. pGT, ground truth; P, HE-path; M, IF-mask. See also Figure S6.

(C) Distribution of plasma cells with or without the typical cartwheel-shaped nuclei (n = 41 cells for pGT⁺P⁺M⁺⁻ and n = 44 cells for pGT⁺P⁻M⁺, two-sided Fisher’s exact test).

(D) Nucleus hematoxylin intensity of lymphocytes (n = 63 cells for pGT⁺P⁺M⁺⁻ and n = 25 cells for pGT⁺P⁻M⁺, two-sided Mann-Whitney U test). See also Figure S7.

(E) Distance (μm) from the endothelial cell to the closest RBC (n = 32 cells for pGT⁺P⁺M⁺⁻ and n = 29 cells for pGT⁺P⁻M⁺). ∗∗∗p < 0.0001, ∗∗p < 0.01. See also Figure S8.