Deep Learning in Radiology: A Comparative Study of CNN Architectures for Automated Detection of Bone Fractures

The research introduces an automatic bone fracture detection method through deep learning networks which evaluates three CNN architectures including Custom CNN and MobileNetV2 and ResNet50. The investigators established a binary classification objective on orthopedic X-ray data in order to address orthopedic needs for rapid diagnostic tools in resource-constrained areas. The models ran for ten epochs with check pointing and early stopping strategies after normalization and uniform resizing to 224×224 pixels. The evaluation metrics consisted of accuracy together with sensitivity values and area under the ROC curve (AUC). The MobileNetV2 achieved clinical deployment readiness through its 100% performance metrics combined with 100% sensitivity in addition to a perfect AUC value of 1.000. The Custom CNN model showed compelling performance levels with an accuracy rate of 97.74% along with sensitivity of 95.50% and an AUC score of 0.9915 because of its lightweight design. Abilities of ResNet50 were lower than MobileNetV2 and Custom CNN for this task as it achieved 73.18% accuracy, 80.50% sensitivity and 0.7820 value for AUC. The research proposal includes future work recommendations for multi-class classification and attention mechanism model fine-tuning as well as explainable AI methods to improve both model interpretation and clinical implementation capabilities.