Fig 1.
Flow diagram of data inclusion and allocation.
Table 1.
Patient characteristics.
Fig 2.
The architecture of the deep convolutional neural network used for OPLL segmentation.
OPLL, ossification of the posterior longitudinal ligament.
Table 2.
Results of the observer performance test, including subgroup analysis, according to the morphologic subtype of OPLL.
Fig 3.
A 65-year-old male patient with mixed-type OPLL.
(A) A lateral plain radiograph shows mixed-type OPLL along the posterior side of the vertebra. The lesion at the C7 level is obscured by the shoulder shadow. (b) In the sagittal image of cervical spine CT, ossifications ranging from the C2 to T1 level are clearly demonstrated (window width = 2000 HU, window level = 500 HU). (c) OPLL lesions annotated by a radiologist on the plain radiograph. (d) In the resulting image inferred by the deep-learning model, OPLL lesions at the C2-6 levels are well predicted, but the lesion located at the C7 level was not detected by the model. OPLL, ossification of the posterior longitudinal ligament.
Fig 4.
A 51-year-old female patient with segmental-type OPLL.
(a) A lateral plain radiograph shows segmental-type OPLL (arrow) at the C5-6 level. (b) A sagittal image of cervical spine CT also demonstrates the segmental ossifications (arrow) at C5-6 level (window width = 2000 HU, window level = 500 HU). (c) OPLL lesions annotated by a radiologist on the plain radiograph. (d) The deep-learning model correctly predicted the segmental-type OPLL, which was overlooked by two observers. OPLL, ossification of the posterior longitudinal ligament.
Fig 5.
Comparison of observer performances with and without the DL model (a) ROC curve of the DL model and average observers in per-patient analysis.
The AUC of the DL model alone was 0.851 (95% CI, 0.799–0.903), and that for average observers was 0.841 (95% CI, 0.781–0.901). The AUC of average observers improved to 0.911 (95% CI, 0.876–0.945) when referring to the results of the DL model. (b) Improved diagnostic performance of individual observers in per-patient analysis with the assistance of the DL model. ROC, receiver operating characteristics; DL, deep learning; AUC, area under the curve.
Table 3.
Subgroup analysis according to the vertebral level.