Table 1.
Distribution of scans used to develop and evaluate the method.
Fig 1.
Visualization of the annotated landmarks overlaid on a maximum intensity projection of a CBCT.
(a) Left side landmarks from the laterally symmetric pairs. (b) Midline landmarks.
Fig 2.
The architecture of the deep learning model a. A U-net style deep learning system architecture with a contracting pathway, an expanding pathway, and connecting pathways between. b. Single residual block consists of three convolutions, each followed by group normalization (GN) and ReLU nonlinearity with elementwise summation.
Fig 3.
Boxplots of landmarking results evaluated with Euclidean distance when grouping by (a) device and (b) cohort. Barplots landmarking success detection rate (SDR) when grouping by (c) device and (d) cohort. Statistical significance was determined using the Mann-Whitney U test with Benjamini–Hochberg correction procedure.
Fig 4.
Boxplot of landmarking results with each row having (a) the left, (b) right and (c) midline landmarks shown individually and as a group.
Statistical significance was determined using the Mann-Whitney U test with Benjamini–Hochberg correction procedure. *)statistically significant difference (p < 0.05).
Table 2.
Comparison of the mean and standard deviation of distances and successful detection rate for the landmarks between the cohorts.
Fig 5.
Boxplot of performance on cephalometric characteristic measures with 3D point-to-point measures with (a) reference to prediction distance and (b) error, components of vector with (c) reference to prediction distance and (d) error, and plane-to-plane angle with (e) reference to prediction distance and (f) error.
Cohort difference between each error measures are compared with the Mann-Whitney U test with Benjamini–Hochberg correction procedure.
Table 3.
Mean and standard deviation of distances and successful detection rate (SDR) from reference to prediction (Ref.—Pred.) for the cephalometric characteristics.
SDR is defined as ≤2 mm or ≤2° of the absolute error.