Ensemble Deep Learning Approaches for Multiclass Classification of Hip Region Fractures in X-Ray Images

Abstract

Hip region fractures, including pelvic, femoral neck, intertrochanteric, and subtrochanteric fractures, are critical medical conditions, especially when diagnosed early. These fractures impair mobility, increase risks, and cause complications. Early diagnosis using X-ray imaging is vital for effective treatment. Recent advances in computer vision, particularly ensemble pretrained models, have revolutionized fracture detection by combining various models to improve classification accuracy and stability. This research developed and evaluated ensemble deep learning methods for multiclass classification of hip fractures on X-ray images. The dataset consists of 1000 X-ray images from Sri Lankan hospitals (20222023), categorized into two types: non-fracture and fracture. Preprocessing and data augmentation techniques are used to increase dataset diversity. The data was split into 70:15:15 for training, validation, and testing to evaluate performance. The pretrained model architectures include ResNet- 101, ResNet-50, EfficientNetB0, and EfficientNetV2, with ResNet-10 taken with different levels of parameterization. ResNet101 achieves the highest test accuracy of 0.8000, followed by ResNet- 50 (0.7786), EfficientNetB0 (0.7286), and EfficientNetV2 (0.7500). These pretrained models are induced as ensemble learning models and enhance multiclass hip fracture classification, yielding more accurate results compared to customized vision models. This approach has potential clinical applications, aiding early and reliable diagnosis. Further, it can extend to differentiate the components of the hip region individually with sophisticated data augmentation techniques that help for the classification. This research proves that pretrained models can be effective in biomedical rather than building and training them from scratch.