Abstract
<title>Abstract</title> <p>Background Skull fractures are critical indicators of traumatic brain injury (TBI), yet their detection on computed tomography (CT) scans remains challenging due to overlapping anatomical structures and high radiologist workload. Though Deep learning models offer automated detection potential their clinical adoption is limited by opaque decision-making. Methods A convolutional neural network (ResNet-50) and a data-efficient vision transformer (DeiT-Tiny) for slice-level skull fracture classification in CT scans were compared. Axial CT slices from 15 patients (4,914 CT slices) were collected under ethical approval at the National Hospital, Kandy, Sri Lanka. Preprocessing included Hounsfield unit windowing and augmentation of fractured slices. Both models were trained with ImageNet-pretrained weights and evaluated using a patient-level train-validation-test split. Explainability was assessed using Grad-CAM (ResNet-50) and Attention Rollout (DeiT-Tiny), with quantitative alignment measured against radiologist-annotated fracture regions. Results ResNet-50 achieved superior classification performance (accuracy 0.726, AUC-ROC 0.800, F1-score 0.428) compared to DeiT-Tiny (accuracy 0.673, AUC-ROC 0.746, F1-score 0.384). The AUC difference was statistically significant (mean difference = 0.054; 95% CI: 0.035–0.074; p < 0.001). DeiT-Tiny demonstrated marginally better spatial alignment in explainability maps (mean IoU 0.021 vs 0.006; mean Dice 0.041 vs 0.012), though absolute values remained low. Conclusions ResNet-50 provides statistically robust classification performance under limited-data conditions, while DeiT-Tiny offers modest advantages in the spatial coherence of visual explanations. These findings highlight a trade-off between predictive accuracy and interpretability in deep learning-based clinical decision support. Future work should focus on multi-centre validation, patient-level classification, and improved explainability frameworks for small medical imaging datasets.</p>