The effects of image manipulation on classification of cervical spondylosis X-ray images using deep learning

Cervical spondylosis is a prevalent degenerative disorder of the cervical spine and a leading cause of chronic neck pain and neurological impairment worldwide. Diagnostic imaging, particularly X-rays, remains the first-line tool for detection; however, reliance on manual interpretation contributes to delays, variability, and potential diagnostic error. This study addresses the challenge of accurately diagnosing cervical spondylosis from X-rays by assessing the effects of various preprocessing procedures on the efficacy of classical convolutional neural networks (CNNs) for image classification. To determine whether preprocessing improves diagnostic outcomes, we trained CNNs with transfer learning on the Cervical Spine X-ray Atlas (CXSA) dataset, applied techniques such as wide-area cropping and color enhancement, and measured classification accuracy, precision, and F1-scores. We hypothesized that image processing via the wide cropping of images would significantly increase the performance of our model when tasked with detecting cervical spondylosis in comparison with other typical preprocessing techniques. Our most effective approaches, wide-area cropping and color enhancement, achieved a maximum accuracy of 95.73%, with wide-area cropping rooting out the most false negatives and positives. These findings demonstrate that image preprocessing improves diagnostic accuracy and efficiency, offering potential for clinical translation. More broadly, the methods developed here could be extended to other musculoskeletal conditions, supporting more reliable, individualized treatment plans and advancing the role of deep learning in medical diagnostics.