Fig 1.
Schematic structure of the C2-net model construction and evaluation.
a shows the resampling of CT images to standardize pixel spacing across different subject data. b depicts the image segmentation module of the C2-Net model, which is used to segment the C2 pedicle regions from CT images. c illustrates the probability assessment module of the C2-Net model, which is used to evaluate the feasibility of placing screws in the C2 pedicles. The output of the module is the probability of successful and failed placement. d presents the evaluation results of the model, comparing the predicted results with the 3D-printed ground truth and the judgments of surgeons with different levels of experience.
Fig 2.
a: Utilize the “Multi-Planar Reconstruction (MPR) function” of the RadiAnt DICOM Viewer software; b: Employ the MPR function to correct potential skeletal deformities or improper positioning during patient scans to the standard transverse and standard sagittal planes; c: Reconstruct a tilted transverse plane along the longitudinal axis of the vertebral arch root notch in the standard sagittal position (yellow line); d: Reconstruct an oblique coronal plane perpendicular to the longitudinal axis of the vertebral arch root notch in the tilted transverse plane (red line), ensuring that the oblique coronal section is simultaneously perpendicular to the axis (blue line) and sagittal plane (yellow line) of the root arch, and measure the marrow cavity width (MPD) at the narrowest part of the vertebral arch root notch on the oblique coronal section; e: Classify cases with MPD ≥ 4.78 mm into the low-risk group; f: Classify cases with MPD < 4.78 mm into the high-risk group.
Fig 3.
3D-printed model and CT assessment of C2 pedicle screw insertion.
C2 pedicle screw placement in a 3D-printed bone model, followed by CT multiplanar reconstruction to evaluate screw trajectory and cortical integrity.
Table 1.
Summary of clinical and imaging features of study subjects and C2 minimum pediculoisthmic component diameter.
Fig 4.
Flowchart of subject data retrieval and the study.
405 patients (490 C2 pedicles) were enrolled and stratified by pedicle diameter into low- and high-risk groups for screw placement. Both groups were included in the 10-fold cross-validation for C2-NET training and validation, while an independent test set was used to compare the AI model's performance with that of doctors.
Fig 5.
ROC curve of the validation, test set.
ROC curves for the C2-NET model, showing performance in the 10-fold cross-validation (left) and independent test set (right). Each curve corresponds to a validation fold, with AUC values indicating strong and consistent diagnostic accuracy.
Fig 6.
Performance comparison of C2-net model with junior and senior surgeons in C2 pedicle screw placement assessment.
Confusion matrices illustrate the diagnostic outcomes of junior and senior surgeons, while the ROC curves compare their performance to the C2-NET model, demonstrating the model's superior accuracy in identifying high-risk C2 pedicles.
Fig 7.
Attention area in the CT images detected by C2-Net.
Original CT images (top) and corresponding model attention maps (bottom) show the regions of interest the C2-NET model focuses on to evaluate C2 pedicle screw placement risk, with color intensity reflecting the degree of attention.