Enabling artificial intelligence-driven blind joint ultrasound self-scans in hemophilia: an annotation protocol.
Authors
Affiliations (6)
Affiliations (6)
- AI & Analytics, Novo Nordisk A/S, Søborg, Denmark.
- Hamilton Niagara Hemophilia Treatment Center, McMaster Children's Hospital, Hamilton Health Sciences, Department of Rehabilitation Science, McMaster University, Hamilton, Ontario, Canada.
- Angelo Bianchi Bonomi Hemophilia and Thrombosis Center, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico, Department of Pathophysiology and Transplantation, University of Milan, Milan, Italy.
- Department of Diagnostic & Interventional Radiology, The Hospital for Sick Children, Medical Imaging Department, University of Toronto, Toronto, Canada.
- Dipartimento di Scienze della Salute, University of Genoa, Genoa, Italy.
- Children's Hospital of Central Switzerland, Luzern, Switzerland.
Abstract
Early detection and management of joint bleeding are critical for preventing long-term joint damage in people with hemophilia. However, current practices often require healthcare professionals to assess bleeding, limiting patient autonomy and potentially leading to delayed or unnecessary treatment. Current literature primarily focuses on clinical protocols for ultrasound (US) imaging but lacks annotation protocols tailored for training artificial intelligence (AI) systems. To propose a structured annotation protocol to support the development of an AI-driven, ultraportable US system that enables people with hemophilia to perform self-guided scans of their joints, focusing specifically on the posterior recess of the elbow, anterior subquadricipital recess of the knee, and anterior recess of the ankle. The annotation protocol defines eight criteria to assess the appropriateness of US images and five levels for region-of-interest clarity. A pool of six readers rated US videos of elbows, knees, and ankles from 100 healthy subjects, and their annotations were analyzed using the intraclass correlation coefficient (ICC). The application of the proposed protocol demonstrates moderate to good agreement based on the ICC when assessing the reliability of single ratings from individual readers (ICC(2,1) = 0.7643), with the knee exhibiting the highest reliability. When considering average ratings across multiple readers, we observe overall excellent reliability, with an ICC of 0.9075. This annotation protocol supports the development of future AI-guided image-acquisition systems for joint self-scanning, marking a significant step toward at-home joint assessment for people with hemophilia. Future research should address the evaluation of the AI guidance system, patient-performed scanning, and diagnostic accuracy for hemarthrosis.