Figure 5. Boxplots of Raman-estimated fat content vs. (A) grade of
pathologist-estimated large droplet macrosteatosis, (B)
pathologist-estimated all steatosis in percentage (to the nearest 5%),
and (C) the algorithm-estimated Oil Red O-stained pixels over pixel area
(of Oil Red O-stained slides) in percentage. The boxes represent the
interquartile range, with the horizontal lines inside the box indicating
the median fat content values. The blank stars inside the box represent
the mean fat content values. The whiskers extend from the boxes to
represent the minimum and maximum fat content values, excluding any
outliers (the 1.5xIQR rule applied) plotted as individual dots outside
the whiskers.
Logistic regression (using the “logit ” command) revealed
significant (all p-values < 0.0001) differences in the
Raman-estimated fat content among all groups. As shown in Table
2 , the Raman-estimated fat content, as a single predictor (using the
“predict ” command), effectively differentiated minimum risk
(≤10%), low-risk (10%-30%), high-risk (30%-60%), and maximum-risk
(≥60%) MaS and global HS, with areas of the operating characteristic
curve (AUROC) between 0.88 and 0.90.
Utilizing both Raman and reflectance signals enhanced the AUROC of
almost all binary classifications; however, the differences between the
Raman and dual-channel predictions were insignificant (see Table
S1 ).
Most specimens in the pathological validation stage had a mixed form of
MaS and MiS. As Figure 6(A-B) shows, the degree of MaS of the
studied liver specimens was strongly positively correlated with the
degrees of pathologist-estimated global HS and the algorithm-estimated
positive pixel area, limiting examinations of the investigated Raman
system on distinguishing MaS from global HS.