UNSTRUCTURED
This study evaluated the performance of a deep learning model trained to detect scaphoid fractures in radiographs when various perturbations were added to the images. The datasets were modified by applying Gaussian noise; via blurring, JPEG compression, contrast-limited adaptive histogram equalization (CLAHE), and resizing; and the addition of geometric offsets. Model accuracy declined as the severity of perturbations increased; however, the extent of performance decline varied according to the type of perturbation. The model demonstrated greater resistance to color perturbations than to grayscale perturbations, but the application of Gaussian blur exerted a considerably negative impact on model performance. CLAHE increased the false-positive rate. There was a strong linear relationship between image quality and model performance; the model performed better on higher-quality images. We also found that geometric offset or pixel value rescaling did not affect the performance of the deep learning model. Resolution was the most significant factor influencing model performance; localizing the region of interest may minimize any decrease in accuracy. Overall, this study provides insights into the robustness of deep learning models that detect small fractures, such as scaphoid fractures, in radiographs subjected to various perturbations. The findings could guide the development of more accurate models for medical image analysis. The cumulative insights gained from this study will contribute to the design of more accurate and robust deep learning models tailored for medical image analysis, especially under conditions of varying image quality.