An Effective Scaphoid Fracture Recognition Using Spatially Adaptive Denormalization and Machine Learning Methods with Multi-Task Loss Function

Automated identification of scaphoid fractures leverages medical imaging to reveal small fractures in the scaphoid bone of the wrist, which are usually hard to locate. Scaphoid fractures are common wrist injuries that often remain undiagnosed initially due to subtle signs in X-ray images. This study offers a solution through a Spatial Feature- and SAD-based Adaptive Multimodal Model (SFSAMM) for automated scaphoid fracture recognition. A dataset from Kaohsiung Chang Gung Memorial Hospital included 486 images of 154 adult patients, with 178 showing fractures and 308 depicting normal conditions. After localization through the YOLO-v4 object detection algorithm, achieving 96% accuracy, a Spatially Adaptive Denormalization (SAD) module processes the extracted sub-images to preserve crucial spatial features. These representations are classified using a feedforward Artificial Neural Network (ANN) with one hidden layer of 120 neurons optimized via backpropagation and multi-task learning predictions for fractures versus normal cases. The evaluation metrics are ROC-AUC curve, error, Threshold-based ROC, loss and accuracy.