Automatic Detection of Pediatric Craniofacial Fractures in Computed Tomography Using a 2.5D Deep Learning Model and Grad-CAM

Pediatric craniofacial fractures represent a significant diagnostic challenge due to age-dependent anatomical variability, subtle fracture lines, and the limited availability of specialists during emergency care. Computed tomography (CT) remains the gold standard for fracture assessment; however, small or non-displaced fractures may be easily overlooked. Advances in artificial intelligence offer opportunities for automated support in injury detection. In this study, we developed a 2.5D deep learning model that uses multiplanar CT reconstructions (axial, coronal, sagittal) as a three-channel input to a ResNet-18 network with transfer learning. A total of 63 high-quality CT examinations were included and split into a training set (n = 44) and a validation set (n = 19). The preprocessing pipeline included isotropic resampling, bone-window normalization, and the generation of 384 × 384 images. Model performance was evaluated using AUCROC, AUCPR, sensitivity, specificity, precision, F1-score, and threshold optimization using Youden’s index. The proposed model achieved AUCROC = 0.822 and AUCPR = 0.855. At the optimal decision threshold (t = 0.54), sensitivity reached 0.556, specificity 1.000, and F1-score 0.714. Grad-CAM visualizations correctly highlighted fracture-related structures, which was confirmed in an independent clinical assessment. The results demonstrate the potential of the proposed model as a decision-support tool for pediatric craniofacial trauma, providing a foundation for future extensions toward radiomics and hybrid deep-radiomic pipelines.