Artificial Intelligence for Automated Detection of Joint Bleeding via Ultrasound in Hemophilia: Advancing Standardization.
Authors
Affiliations (6)
Affiliations (6)
- Fondazione IRCCS Ca' Granda, Ospedale Maggiore Policlinico di Milano, Angelo Bianchi Bonomi Hemophilia and Thrombosis Center, Milan, Italy; Department of Pathophysiology and Transplantation, Università degli Studi di Milano, Milan, Italy.
- Fondazione IRCCS Ca' Granda, Ospedale Maggiore Policlinico di Milano, Angelo Bianchi Bonomi Hemophilia and Thrombosis Center, Milan, Italy.
- Department of Computer Science, Università degli Studi di Milano, Milan, Italy.
- Fondazione IRCCS Ca' Granda, Ospedale Maggiore Policlinico di Milano, Diagnostic and Interventional Radiology Department, Milan, Italy.
- Fondazione IRCCS Ca' Granda, Ospedale Maggiore Policlinico di Milano, Diagnostic and Interventional Radiology Department, Milan, Italy; Department of Oncology and Hematology-Oncology, Università degli Studi di Milano, Milan, Italy.
- Fondazione IRCCS Ca' Granda, Ospedale Maggiore Policlinico di Milano, Angelo Bianchi Bonomi Hemophilia and Thrombosis Center, Milan, Italy; Department of Pathophysiology and Transplantation, Università degli Studi di Milano, Milan, Italy. Electronic address: [email protected].
Abstract
Musculoskeletal ultrasound (US) is a non-invasive tool for joint assessment in persons with hemophilia. Early detection of joint bleeding using a remote US system operated by patients or caregivers and reviewed by Comprehensive Care Centers could improve personalized management. A computer-aided diagnosis (CAD) system for automatic detection of joint effusion may support clinicians in prioritizing interventions. This study aimed to validate a novel CAD system using a deep-learning algorithm to identify joint capsule distension in musculoskeletal US images. Longitudinal scans of the subquadricipital recess (SQR) of the knee were collected from people with hemophilia and varying degrees of arthropathy and labeled by an expert. The multi-task learning algorithm was trained to detect the recess and classify images as distended or not. A total of 8,634 images (2,267 scans) were acquired from 158 adult persons with hemophilia (mean age 44.7 ± 18.6 years) and 66 age-matched healthy controls. After selecting longitudinal SQR images, 814 images were used, of which 711 for training and 103 for testing, ensuring a patient-based split. The model achieved a classification accuracy of 89.2% and a balanced accuracy of 93.9% compared to expert annotations. No significant differences were observed in classification performance between male and female healthy controls, supporting its broader applicability. The CAD system for automatic detection of joint capsule distension is feasible and reliable. It represents an important step toward telemedicine in hemophilia, enabling early recognition of joint bleeding and supporting personalized, timely therapeutic interventions to prevent further joint damage.