Patch-Wise Approach with Vision Transformer for Detecting Implant Failure in Spinal Radiography

Spinal implant failures are becoming more common due to the increasing number of surgeries and an aging population. Early detection is crucial but often missed on radiographs due to subtle features and limited radiologist availability. This study aims to develop and validate a vision transformer-based deep learning model for detecting spinal implant fractures and to assess its impact on radiologist performance. The training and testing datasets included 9924 spinal radiographs (3492 studies) from 798 patients, between 2003 and 2023. A DINOv2-based model first detected spinal implants in radiographs, then analyzed 224 × 224-pixel patches within those regions to classify the presence or absence of implant fractures. Three radiologists (1 to 16 years of experience) independently reviewed the test set (1538 images) with and without AI assistance. Performance was measured using accuracy, F1 score, precision, recall, and generalized estimating equation (GEE) analysis. The AI model achieved a recall of 0.94, precision of 0.37, F1 score of 0.54, and accuracy of 0.83. With AI assistance, radiologist recall improved from 0.70 to 0.95, with the greatest gain seen in the least experienced reader (0.49 to 0.92). GEE analysis confirmed significant diagnostic improvement (OR = 2.82, p < 0.001), particularly in rod fractures. This patch-wise transformer-based approach demonstrated high sensitivity and improved reader performance, supporting its use as a triage tool in spinal implant surveillance.