Figure 6. (A-B) Plots and linear regressions of (A)
pathologist-estimated global HS vs. pathologist-estimated MaS and (B)
the algorithm-estimated global HS vs. pathologist-estimated MaS. (C-D)
ROCs of dual-variable binary predictions on ≥10% steatosis discrepancy
between (C) pathologist-estimated global HS and pathologist-estimated
MaS and (D) algorithm-estimated global HS and pathologist-estimated MaS.
The Raman and the reflectance channels were applied as dual predictors.
Considering the inherently high correlation with global HS (assessed by
the expert pathologist and the algorithm) and the pathological
assessments are to the nearest 5% steatosis, we set a compromised
steatosis discrepancy threshold at 10%. Information from the
reflectance channel was paired with the Raman-estimated fat content for
dual-variable predictions. Figure 6(C) shows the ROC curves of
binary predictions on ≥10% steatosis discrepancy between
pathologist-estimated global HS and pathologist-estimated MaS, with an
AUROC of 0.66. This could be due to data points’ degradation
(overlapping), as the expert pathologist’s estimations were to the
nearest 5%.
The ROC curve of binary predictions on ≥10% steatosis discrepancy
between the algorithm-estimated global HS (positive pixel area) and
pathologist-estimated MaS presented a better AUROC of 0.80, as shown inFigure 6(D) . This may be attributed to the algorithm-based
estimation of global HS, which quantified the areas of ORO-stained lipid
droplets in liver specimens, yielding more precise percentage readings.