Radiographic Classification of Low-Energy Distal Radius Fracture Status Using Automated Hand Segmentation

This work presents a radiographic classification framework using hand and wrist radio- graphs to distinguish low-energy distal radius fracture cases from non-fracture controls. The framework integrates advanced preprocessing techniques with deep learning architectures. The proposed method begins with hand keypoint detection, employing a suite of image preprocessing techniques and the application of a pre-trained landmark de- tector. Subsequently, the hand region of each image is extracted for further analysis. To effectively learn the features and patterns distinguishing the two classes, a transfer learning strategy is adopted by fine-tuning multiple pretrained network architectures. Comprehensive experiments validate the framework's effectiveness, with VGG-16 achieving a peak accuracy of 98.16% and ResNet-18 reaching 97.63%, supported by AUC values exceeding 99% for both models. Additionally, graphical visualizations such as PCA plots and ROC curves demonstrate the models' ability to capture meaningful features and accurately differentiate between healthy and osteoporotic conditions. These results are anticipated to pave the way for significant advancements in clinical applications of medical imaging.