Article Figures & Data
Figures
- FIG 1.
Flow chart of patient inclusion.
- FIG 2.
Representative GRE images of the cervical spine (top) with insets (middle) and examples of spinal canal segmentations (bottom) performed by trained radiology physicians demonstrating examples of no/mild stenosis (A), moderate stenosis (B), and severe stenosis/cord compression (C).
- FIG 3.
Vision transformer analysis pipeline. Images are divided into flattened patches, which are then fed into the transformer encoder along with positional encoding. For each section, a prediction of severity is generated.
- FIG 4.
Receiver operating characteristic curves comparing patient-level performance of ViT, DenseNet121, and ResNet50 in detecting cord compression.
- FIG 5.
False-positive and -negative cases. A, False-positive: Diffuse congenital moderate stenosis with small superimposed annular bulge, incorrectly categorized as positive for cord compression. B, False-negative: Large annular bulge with severe spinal stenosis and cord compression and abnormal T2 hyperintense signal within the cord, incorrectly categorized as negative for cord compression.
Tables
- Table 1:
Patient demographics and clinical setting in the training/validation and testing cohorts
Characteristic Training/Validation (n = 113) Testing (n = 29) Demographics Male 56 (49.5%) 11 (37.9%) Age 57.5 53.1 Clinical setting ER 64 (56.6%) 18 (62.1%) Inpatient 16 (14.2%) 3 (10.3%) Outpatient 33 (29.2%) 8 (27.6%) History of trauma 49 (43.4%) 14 (48.3%) - Table 2:
Section-level and patient-level test results for ViT, ResNet50, and DenseNet121 models
ViT ResNet50 DenseNet121 Section-level classification accuracy 82% 72% 78% Patient-level accuracy 93% 62% 62% Patient-level sensitivity 0.90 1.0 0.8 Patient-level specificity 0.95 0.42 0.52 Patient-level PPV 0.90 0.47 0.47 Patient-level NPV 0.95 1.0 0.83
