Automated Diagnosis of Infantile Skull Fractures From X-Ray Images Using an Ensemble Deep Learning Model.
Authors
Affiliations (8)
Affiliations (8)
- Department of Neurosurgery, Sheikh Khalifa Specialty Hospital, Ras Al Khaimah, United Arab Emirates.
- Interdisciplinary Program for Bioinformatics, Graduate School, Seoul National University College of Medicine, Seoul, Korea. [email protected].
- Medical Artificial Intelligence Center, Hallym University Medical Center, Anyang, Korea.
- Artificial Intelligence Research Center, Hallym University Chuncheon Sacred Heart Hospital, Chuncheon, Korea.
- Department of Biomedical Informatics, Hallym University Chuncheon Sacred Heart Hospital, Chuncheon, Korea.
- Department of Radiology, Hallym University Sacred Heart Hospital, Hallym University College of Medicine, Anyang, Korea.
- Department of Neurosurgery, Hallym University Sacred Heart Hospital, Hallym University College of Medicine, Anyang, Korea.
- Department of Ophthalmology, Hallym University Sacred Heart Hospital, Hallym University College of Medicine, Anyang, Korea. [email protected].
Abstract
To develop an artificial intelligence (AI)-assisted model for detecting skull fractures in neonates and infants using plain radiographs, enhancing diagnostic accuracy while minimizing radiation exposure. A retrospective dataset of skull X-rays from 1,184 patients with head trauma (2010-2021) was collected. Images underwent preprocessing, including background removal, Gaussian blurring, binarization, and CLAHE-based contrast enhancement. Three convolutional neural network architectures (ResNet-50, DenseNet-121, EfficientNet-B5) were trained, with DenseNet-121 optimized using CLAHE. An ensemble model combining anterior-posterior (AP) and lateral views was constructed. The model's performance was evaluated using internal (4,298 images) and external (460 images) datasets. DenseNet-121 with CLAHE preprocessing achieved the highest mean area under the curve (AUC) of 0.926 for the AP view and 0.911 for the lateral view, demonstrating robust performance. The ensemble model, which combined both AP and lateral views, further enhanced the model's diagnostic performance. It achieved an overall AUC of 0.938, with a classification accuracy of 91.6% on the external validation dataset. Expert annotation by neurosurgeons and radiologists significantly improved the model's reliability, enabling accurate differentiation between skull fractures and normal anatomical structures, such as cranial sutures, reducing false positives. This study presents a highly accurate AI model for pediatric skull fracture detection, incorporating CLAHE-enhanced preprocessing, deep learning, and expert-annotated data. The model improves clinical decision-making, reduces unnecessary computed tomography scans, and provides valuable anatomical insights into the challenges of distinguishing fractures from normal cranial variations.