Abstract
BACKGROUND: Deep learning models have shown strong potential for automated fracture detection in medical images. However, their robustness under varying image quality remains uncertain, particularly for small and subtle fractures, such as scaphoid fractures. Understanding how different types of image perturbations affect model performance is crucial for ensuring reliable deployment in clinical practice. OBJECTIVE: This study aimed to evaluate the robustness of a deep learning model trained to detect scaphoid fractures in radiographs when exposed to various image perturbations. We sought to identify which perturbations most strongly impact performance and to explore strategies to mitigate performance degradation. METHODS: Radiographic datasets were systematically modified by applying Gaussian noise, blurring, JPEG compression, contrast-limited adaptive histogram equalization, resizing, and geometric offsets. Model accuracy was evaluated across different perturbation types and levels. Image quality was quantified using peak signal-to-noise ratio and structural similarity index measure to assess correlations between degradation and model performance. RESULTS: Model accuracy declined with increasing perturbation severity, but the extent varied across perturbation types. Gaussian blur caused the most substantial performance drop, whereas contrast-limited adaptive histogram equalization increased the false-negative rate. The model demonstrated higher resilience to color perturbations than to grayscale degradations. A strong linear correlation was found between peak signal-to-noise ratio-structural similarity index measure and accuracy, suggesting that better image quality led to improved detection. Geometric offsets and pixel value rescaling had minimal influence, whereas resolution was the dominant factor affecting performance. CONCLUSIONS: The findings indicate that image quality, especially resolution and blurring, substantially influences the robustness of deep learning-based fracture detection models. Ensuring adequate image resolution and quality control can enhance diagnostic reliability. These results provide valuable insights for designing more accurate and resilient medical imaging models under real-world variability.