Browsing by Author "Oyindamola Eniola Ajiboye"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    An Explainable ResNet-101 Framework for Hand Fracture Classification in X-ray
    (Co hosted by Baze University Big Data Analytics and Innovation Research Group, American University Yola, and Sule Lamido University, 2026-04-23) Ayodeji Jubril Alabi; Shakirat Ronke Yusuff; Sulaiman Olaniyi Abdulsalam; Ismail Idowu Akuji; Oyindamola Eniola Ajiboye
    : Fracture detection in radiographic images continues to be a difficult task due to subtle bone disruptions and variations in anatomical structures, especially within the hand region. This study proposes an explainable deep learning framework for automated hand fracture detection using a ResNet-101 architecture. The model was trained and evaluated on the hand subset of the publicly available MURA dataset, with preprocessing methods and data balancing strategies applied to mitigate class imbalance and enhance model generalization. Bayesian hyperparameter optimization was utilized to determine the most suitable training parameters, including learning rate, dropout rate, and network capacity, thereby improving model stability and overall performance. The optimized model achieved an accuracy of 79.7% with an area under the curve (AUC) of 0.75, indicating dependable classification performance. To enhance interpretability, Grad-CAM++ was incorporated to visualize the regions of interest that influence the model’s predictions, allowing verification of clinically meaningful focus areas. The results demonstrate that the model effectively detects fracture regions while maintaining stable performance across diverse radiographic samples. In addition, the system produces automated PDF-based clinical reports that combine predictions with corresponding heatmaps, thereby improving its practical relevance in clinical environments. Overall, the proposed approach offers a robust, interpretable, and clinically valuable solution for hand fracture detection, with strong potential for integration into computer-aided diagnosis systems. Future work will focus on extending the model to multi-centre datasets and incorporating additional clinical features to further enhance generalization and diagnostic reliability