Deep learning framework for the assessment of odontoid fractures on cervical CT: optimization of sensitivity and specificity

Abstract Objective Odontoid fractures are injuries that may be overlooked on cervical CT and can lead to serious clinical consequences. In this study, four deep learning models were evaluated independently for the assessment of odontoid fractures on selected 2D CT slices, with particular attention to false-negative and false-positive predictions. Methods A balanced dataset was used (training: 1208 images; test: 604 images; 302 fracture and 302 healthy images in the test set). Fracture cases were annotated with bounding boxes. YOLOv10s was used for detection, ResNet18 for classification, and nnU-Net v2 2D and SwinUNETR 2D for weakly supervised segmentation. Each model was evaluated independently. SwinUNETR’s image-level threshold analysis was exploratory. Results ResNet18 classification using reference-box crops for fracture-positive images and central crops for healthy images achieved accuracy of 98%, sensitivity of 99.3%, specificity of 96.7%, and F1 of 98%. YOLOv10s-derived image-level decisions yielded specificity of 97.4%, precision of 97.1%, sensitivity of 90.1%, and F1 of 93.5%. SwinUNETR 2D mask-derived decisions yielded sensitivity of 99.3%, specificity of 40.4%, accuracy of 69.9%, and F1 of 76.7%. Corresponding nnU-Net v2 2D values were 70.9%, 88.4%, 79.6%, and 77.7%. These image-level metrics describe task-specific outputs and decision rules. Conclusion The findings support preliminary technical feasibility within a single-center, selected-slice evaluation. The proposed alarm–confirmation strategy was not tested as a combined system. External validation on complete CT examinations and comparison with radiologists are needed to establish clinical effectiveness.