Pediatric fracture classification in plain radiographs using EfficientNetV2 with proximal policy optimization fine-tuning.
Authors
Affiliations (11)
Affiliations (11)
- Information Technology Department, Cybersecurity Program, Faculty of Information Technology & Computer Sciences, Yarmouk University, Irbid, Jordan.
- Department of Diagnostic Radiology and Nuclear Medicine, Faculty of Medicine, King Abdullah University Hospital, Jordan University of Science and Technology, P.O.Box 3030, Irbid, 22110, Jordan.
- Computer Science Department, Faculty of Information Technology and Computer Sciences, Yarmouk University, Irbid, 21163, Jordan.
- Department of Business Administration, Trine University, Angola, IN, USA.
- Center of Advanced Research for Complementary Medicine, Department of Artificial Intelligence, University of Zawia, Zawia, Libya.
- Faculty of Medicine, Jordan University of Science & Technology, P.O.Box 3030, Irbid, 22110, Jordan.
- Computer Engineering Department, King Abdullah II School of Engineering, Princess Sumaya University for Technology, Amman, 11941, Jordan.
- Department of Industrial Engineering, College of Engineering, King Khalid University, P.O. Box 394, 61421, Abha, Saudi Arabia.
- Center for Engineering and Technology Innovations, King Khalid University, 61421, Abha, Saudi Arabia.
- Computer Science & Engineering, College of Engineering, American University of Sharjah, Sharjah, 26666, UAE.
- Computer Science Department, Faculty of Information Technology and Computer Sciences, Yarmouk University, Irbid, 21163, Jordan. [email protected].
Abstract
Pediatric fracture detection in plain radiographs presents distinct clinical challenges due to the presence of growth plates, incomplete ossification, and subtle cortical irregularities that can closely mimic normal developmental bone appearances particularly in young children. These features contribute to documented inter-reader variability and missed fracture rates in busy emergency settings, underscoring the need for reliable automated decision support. This paper presents a binary classification framework for pediatric radiograph analysis, combining EfficientNetV2-Small as a visual feature extraction backbone with a Proximal Policy Optimization (PPO) based actor-critic refinement module trained using a two-stage optimization framework. In the first stage, the model undergoes supervised warm-up using cross-entropy loss to establish a stable fracture-versus-normal classification boundary. In the second stage, PPO fine-tuning refines the diagnostic policy through a confidence-aware asymmetric reward function that imposes stronger penalties on high-confidence misclassifications directly targeting the clinical risk of overconfident missed fractures. The framework is developed and evaluated on the KAUH Pediatric Fracture X-ray Dataset, comprising 1221 real-world clinical radiographs collected from King Abdullah University Hospital, Jordan, spanning multiple anatomical sites of the appendicular skeleton and annotated by a certified pediatric radiologist. Contrast Limited Adaptive Histogram Equalization is applied as a preprocessing step to enhance local bone contrast and improve the visibility of subtle fracture-related features. The proposed model achieves a test accuracy of 91.86%, an AUC of 95.90%, and a sensitivity of 98.38%, achieving higher performance than the evaluated CNN baselines under the reported experimental setting MobileNetV2, ResNet50, Xception, DenseNet121, and base EfficientNetV2-Small. The current framework should be regarded as a research prototype, and prospective multi-center validation is necessary before clinical deployment can be considered.