AI-Assisted Sequential Graf Assessment of Infant Hip Ultrasound: A Retrospective Single-Centre Evaluation in a Referral Cohort

first_page settings Order Article Reprints This is an early access version, the complete PDF, HTML, and XML versions will be available soon. Open AccessArticle AI-Assisted Sequential Graf Assessment of Infant Hip Ultrasound: A Retrospective Single-Centre Evaluation in a Referral Cohort by Andrea Audisio 1,2,* [ORCID] , Alessandro Aprato 1,2 [ORCID] , Jessica Hu 3, Giorgia Marullo 4 [ORCID] , Matteo Sabbatini 5, Luca Ulrich 4 [ORCID] , Sandro Moos 4 [ORCID] , Marco Petraz 6 [ORCID] , Enrico Vezzetti 4 [ORCID] and Alessandro Massè 2 [ORCID] 1 Pediatric Orthopaedics and Trauma, Regina Margherita Children’s Hospital, 10126 Turin, Italy 2 Department of Surgical Sciences, University of Turin, 10126 Turin, Italy 3 School of Medicine, University of Turin, 10126 Turin, Italy 4 Department of Management and Production Engineering, Politecnico di Torino, 10129 Turin, Italy 5 Department of Electronics and Telecommunications, Politecnico di Torino, 10129 Turin, Italy 6 Pediatric Radiology, Regina Margherita Children’s Hospital, 10126 Turin, Italy * Author to whom correspondence should be addressed. J. Imaging 2026, 12(11), 554; https://doi.org/10.3390/jimaging12110554 Submission received: 19 August 2026 / Revised: 9 October 2026 / Accepted: 2 November 2026 / Published: 5 November 2026 (This article belongs to the Special Issue Medical Computer Vision: Innovations and Clinical Impact) Download keyboard_arrow_down Review Reports Versions Notes Abstract Graf-method ultrasound supports early assessment of developmental dysplasia of the hip (DDH). This study assessed a sequential artificial-intelligence-assisted workflow and clinicians’ Checklist I decisions with and without computer-aided diagnosis (CAD). This retrospective single-centre study of referred infants used 1803 development images from 679 examinations and separate cohorts of 1024 workflow and 100 reader images; Graf IV hips were excluded. ResTransUNet segmented eight Checklist I structures to assess reportability. Fifteen clinicians assessed the reader images unaided and four weeks later with CAD. Stand-alone Checklist I accuracy was 93.0% (93/100; 95% CI 86.3–96.6). Of 1024 workflow images, 846 received final outputs; Checklist II non-compliance precision was 48.1%. Among accepted images, the non-type-I endpoint had 81.0% accuracy, 82.7% sensitivity and 80.4% specificity; exact Graf-type agreement was 40.1%. Across entering non-type-I images, 167/266 (62.8%) received a correct positive output. Mean absolute errors were 3.59∘ for the 𝛼 angle and 11.91∘ for the 𝛽 angle, with a 𝛽-angle bias of +11.07∘. Clinician accuracy was 72.3% unaided and 90.2% assisted—an observed 17.9 percentage-point difference (95% CI 10.0–26.1); 26 correct decisions became incorrect on AI-error cases. The workflow supported Checklist I assessment, but fixed order precludes causal CAD inference; gate failures and angle errors require refinement and prospective evaluation before clinical implementation.