Enhancing Bone Fracture Detection in Medical Imaging: Implementing Transfer Learning and Adversarial Training

Abstract

The demand for efficient and robust medical image analysis poses significant challenges in accurately detecting bone fractures within specific anatomical regions. Deep learning (DL) models are often vulnerable to adversarial attacks and exhibit limited generalization across diverse datasets, raising concerns about their reliability in clinical settings. This paper aims to compare the performance of U-Net and ResNet50 architectures for bone fracture detection in X-ray images under varying data conditions. Using transfer learning and adversarial training with the Fast Gradient Sign Method (FGSM), three experiments were conducted: training on clean, pre-processed datasets; training on adversarial examples; and training on a combination of clean and adversarial datasets. Two pre-processing techniques, Gaussian Filter and Canny Edge Detection were also applied on the datasets. The performance was evaluated using accuracy, precision, recall, and F1-score. Results show that U-Net outperformed ResNet50 across all experiments and achieved an accuracy of 80%, and F1-score of 80% with adversarial training on combined Gaussian Filter datasets, demonstrating great robustness and generalization capability. On the other hand, ResNet50 showed limited adaptability to adversarial perturbations, especially when processing edge-based features. This paper highlights the effectiveness of adversarial training in enhancing model robustness and highlights U-Net's potential as a reliable tool for bone fracture detection.

Ähnliche Arbeiten