Fractured Bone Diagnostics Using Optimised Convolutional Neural Network Model with Variable Learning Rates

This study explores the human bone fracture diagnostic using Convolutional Neural Network (CNN) models with varying architectures, precisely three and four layers named CNN3 and CNN4, amid varying gradations of learning rates. Two-dimensional bone X-ray images act as the input dataset. The CNN model was trained, validated, tested and evaluated. The evaluation encompasses the outcomes from the intricacies elucidated within the confusion matrix to gauge the models’ effectiveness. The results reveal nuanced trends in the impact of learning rates on the two architectures, with CNN3 demonstrating improved performance as the learning rate decreases. At the same time, CNN4 exhibits sensitivity to the choice of learning rate. This study yielded valuable insights into their performance for a specific task. Lower learning rates, 0.0001 and 0.00001, are associated with enhanced overall performance. These lower learning rates contribute to improved generalisation capabilities. While a higher learning rate (0.001) results in perfect precision, it comes at the expense of lower recall, emphasising the importance of striking a balance. The study concludes that a three-layer CNN model with a learning rate of 0.0001 exhibits exceptional performance with accuracy (0.88), precision (0.98), recall (0.77), F1-score (0.85) and ROC (0.89). Hence, it is the preferred model for addressing the problems under consideration.