<![CDATA[Identifying bone fractures in X-ray images is a complex task that requires special expertise from radiologists and can be time-consuming in clinical workflows. Deep learning offers a significant automated diagnostic solution to improve accuracy and efficiency. This study aims to analyze the performance of three Convolutional Neural Network (CNN) architectures namely, ResNet50, DenseNet169, and EfficientNet-B3 and specifically compare the performance of models trained using augmented data with that of models trained without augmentation. The research method utilizes a local dataset, which is divided equally between the fractured and non-fractured classes. Preprocessing techniques such as Contrast Limited Adaptive Histogram Equalization (CLAHE) were applied, and the models were evaluated on a separate test set (hold-out test set). Model evaluation was conducted using accuracy, precision, recall, F1-score, ROC-AUC metrics, as well as analysis through confusion matrix, classification report, sensitivity, specificity, and calibration curve to assess overall performance. The experimental results show that the application of data augmentation consistently improves the accuracy and robustness of all three models. In the augmentation scenario, EfficientNet-B3 showed the best performance, achieving an accuracy of 93.33%. This study concludes that the combination of the EfficientNet-B3 architecture with the data augmentation strategy is the most optimal and recommended approach for developing a reliable automatic detection system on local X-ray image datasets.]]>
Copyrights © 2025
