Automated detection of bone fractures over x-ray images with deep learning methods

This thesis presents a detailed comparative analysis of five advanced convolutional neural network (CNN) architectures tailored for the detection of bone fractures from X-ray images. Specifically, the study examines the performance of Xception-FracAtlas, VGG16-FracAtlas, ResNet50-FracAtlas, InceptionV3-FracAtlas, and EfficientNetv2b2-FracAtlas, utilizing a dataset distinctly categorized into Non-Fractured and Fractured states. These models are evaluated based on precision, recall, F1-score, overall accuracy and inference time, offering a holistic view of their capabilities and limitations in medical image analysis. EfficientNetv2b2-FracAtlas emerged as the superior model, demonstrating exceptional performance across multiple metrics. Notably, it achieved the highest accuracy rate of 94.91% and an impressive F1-score which indicates its strong precision in identifying Fractured states. Moreover, it recorded the fastest inference time of 25.83 milliseconds and the highest ROC value of 0.95 according to the 80% training and 20% testing values. These results highlight its potential for real-time, accurate medical diagnostics, a critical requirement in clinical settings. In contrast, while the other models also showed promising results, they did not perform uniformly across all metrics. For instance, VGG16-FracAtlas excelled in recall for Non-Fractured states and demonstrated rapid inference times but did not achieve comparable accuracy or F1-scores. Similarly, Xception-FracAtlas and ResNet50-FracAtlas showed proficiency in certain specific metrics but fell short in achieving the all-around effectiveness exhibited by EfficientNetv2b2-FracAtlas. By integrating EfficientNetv2b2-FracAtlas into clinical workflows, healthcare providers can enhance the speed and accuracy of bone fracture detection, potentially reducing the rate of misdiagnosis and facilitating faster treatment responses. Moreover, this study encourages ongoing research to further optimize CNN architectures. Finally, this thesis underscores the transformative potential of deep learning in medical imaging and sets the stage for its increased adoption in clinical settings, promising significant advancements in patient care and treatment outcomes.