Four-fold cross-validation results for different model and loss combinations^a

Model	Loss	MAE (HU)	MSE (×10³ HU)	R	Avg. Bone Precision	Avg. Bone Recall	Avg. Bone Dice
Light_U-Net	MAE	95.6 (94.4–96.9)	54.3 (53.1–55.5)	0.872 (0.869–0.875)	0.665 (0.661–0.669)	0.519 (0.505–0.533)	0.567 (0.558–0.576)
Light_U-Net	MSE	106.0 (103.5–108.4)	51.5 (50.0–53.0)	0.878 (0.875–0.881)	0.621 (0.614–0.629)	0.548 (0.526–0.570)^{^b}	0.558 (0.544–0.573)
Light_U-Net	Mix	97.6 (96.6–98.7)	51.3 (50.4–52.2)	0.878 (0.876–0.880)^{^b}	0.641 (0.636–0.646)	0.538 (0.529–0.546)	0.568 (0.562–0.573)^{^b}
VGG U-Net	MAE	101.5 (99.8–103.3)	60.1 (58.3–61.9)	0.859 (0.856–0.863)	0.667 (0.662–0.672)	0.454 (0.431–0.476)	0.516 (0.497–0.534)
VGG U-Net	MSE	111.5 (106.2–116.7)	55.1 (52.2–58.0)	0.869 (0.864–0.875)	0.614 (0.606–0.622)	0.521 (0.498–0.543)	0.538 (0.517–0.558)
VGG U-Net	Mix	103.4 (100.9–105.9)	55.7 (53.6–57.9)	0.869 (0.865–0.873)	0.643 (0.637–0.648)	0.492 (0.471–0.513)	0.532 (0.514–0.55)
VGG U-Net TL	MAE	99.2 (97.8–100.6)	58.0 (56.6–59.4)	0.864 (0.861–0.867)	0.668 (0.663–0.674)^{^b}	0.470 (0.450–0.490)	0.530 (0.514–0.546)
VGG U-Net TL	MSE	111.7 (108.7–114.6)	55.0 (54.0–56.1)	0.869 (0.867–0.872)	0.619 (0.611–0.627)	0.503 (0.491–0.514)	0.527 (0.517–0.536)
VGG U-Net TL	Mix	103.8 (101.9–105.7)	55.9 (54.4–57.5)	0.867 (0.864–0.870)	0.630 (0.620–0.640)	0.506 (0.489–0.523)	0.540 (0.528–0.552)

Note:—TL indicates transfer learning; Avg., average.
↵a Ninety-five percent confidence intervals across 10 separate training iterations are shown in parentheses. Loss is computed in Hounsfield units, with lower values better for MAE and MSE and higher values better for Pearson R, bone precision, bone recall, and bone Dice scores.
↵b The best score within a column.