Deep learning for odontoid fractures: the critical role of patient-level partitioning and Grad-CAM assessment in model generalization

Abstract Missed diagnoses of odontoid fractures on Computed Tomography (CT) scans can lead to severe morbidity. While Deep Learning (DL) shows promise, many studies report inflated performance metrics due to hidden data leakage (image-level splitting). To evaluate six Convolutional Neural Network (CNN) architectures under a strict zero-leakage protocol for odontoid fracture detection, validating diagnostic reliability and anatomical focus using Explainable AI (XAI). A retrospective cohort of 540 cervical spine CT images from 96 patients was evaluated under a strict zero-leakage protocol. Data was partitioned at the patient level into three disjoint subsets: Training (60 patients, 326 images), Validation (16 patients, 92 images for early stopping), and an independent held-out Test set (20 patients, 122 images touched only once for final evaluation). Six CNN architectures were benchmarked. Statistical robustness was validated via patient-level bootstrapping ( N = 1000 iterations) with Holm-Bonferroni family-wise error rate correction, paired with Grad-CAM visual interpretability. On the untouched test cohort, DenseNet121, GoogLeNet, and ResNet18 achieved similar, leading discriminative performance with AUCs of 0.830 (95% CI: 0.718–0.933), 0.829 (95% CI: 0.714–0.929), and 0.829 (95% CI: 0.697–0.934), respectively, with DenseNet121 demonstrating a balanced diagnostic profile (Accuracy: 73.8%, Sensitivity: 73.0%, Specificity: 75.8%). In stark contrast, SqueezeNet 1.1 and VGG19 suffered complete collapse to majority-class prediction (Specificity: 0.0%) under the evaluated training configuration. Pairwise bootstrap hypothesis testing confirmed that DenseNet121 showed a borderline statistically significant difference over SqueezeNet 1.1 after multiple testing correction ( $$\:{p}_{\text{raw}}=0.0090$$ , Holm-adjusted $$\:p=0.0450$$ on $$\:n=20$$ patients), while differences among DenseNet121, GoogLeNet, and ResNet18 remained statistically non-significant ( $$\:{p}_{\text{adj}}=0.9840$$ ). Architectures incorporating deep feature reuse or residual connections exhibit robust generalization under strict patient-level isolation, whereas compact models and unregularized heavyweight architectures can struggle under small-sample constraints. Rigorous zero-leakage methodology and qualitative XAI assessment are essential prerequisites prior to clinical translation of spinal CAD models.