A Grok API-Based DICOM File-Attachment Structured Reporting Workflow for Silicone Breast Implant MRI: a Single-Center...

This study aims to evaluate a Grok application programming interface (API)-based file-attachment structured reporting workflow for silicone breast implant rupture detection on magnetic resonance imaging (MRI), using radiologist consensus as the reference standard. This retrospective single-center study included 95 patients with 180 evaluable implants. Anonymized native Digital Imaging and Communications in Medicine (DICOM) instances were submitted through a custom interface using a fixed prompt and attached to the Responses API as input_file objects without application-side rendering. Provider-side DICOM pixel decoding could not be independently verified; metadata-mediated inference could not be excluded. The workflow generated human-readable reports and machine-readable outputs. The primary endpoint was binary rupture detection; secondary endpoints were agreement and three-class performance. Radiologist consensus classified 156 implants as intact and 24 as rupture-positive, including 20 intracapsular and 4 extracapsular ruptures. The workflow produced 141 true-negative, 19 true-positive, 15 false-positive, and 5 false-negative classifications. Sensitivity was 79.2% (95% confidence interval [CI], 57.8-92.9%), specificity was 90.4% (95% CI, 84.6-94.5%), accuracy was 88.9% (95% CI, 83.4-93.1%), positive predictive value was 55.9%, and negative predictive value was 96.6%. Binary agreement was 88.9%, with Cohen's kappa of 0.591 and Gwet's agreement coefficient 1 of 0.848. All four extracapsular ruptures were classified as intracapsular rupture. The workflow achieved 88.9% binary accuracy, but provider-side DICOM pixel decoding and Pixel Data access were unverified, and metadata-mediated inference cannot be excluded. Accordingly, the results reflect workflow-level performance rather than validated pixel-level interpretation and support only supervised reporting or educational use.