Artificial Intelligence-Based Analysis of Central Nervous System Vasculopathy in Pediatric Sickle Cell Anemia.
Authors
Affiliations (9)
Affiliations (9)
- The Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, Georgia, USA.
- Department of Pediatrics, Division of Pediatric Hematology/Oncology, Aflac Cancer Center and Blood Disorders Service of Children's Healthcare of Atlanta, Emory University School of Medicine, Atlanta, Georgia, USA.
- Winship Cancer Institute of Emory University, Atlanta, Georgia, USA.
- Parker H. Petit Institute of Bioengineering and Bioscience, Georgia Institute of Technology, Atlanta, Georgia, USA.
- Institute for Matter and Systems, Georgia Institute of Technology, Atlanta, Georgia, USA.
- Center for Machine Learning, Georgia Institute of Technology and Emory University, Atlanta, Georgia, USA.
- Division of Hematology, Cincinnati Children's Hospital Medical Center, Cincinnati, Ohio, USA.
- Caribbean Institute for Health Research, The University of the West Indies, Mona Kingston, Jamaica.
- University of West Indies, Kingston, Jamaica.
Abstract
In children with sickle cell anemia (SCA), central nervous system (CNS) complications such as chronic vasculopathy, silent cerebral infarcts, and overt stroke cause significant morbidity and mortality, and remain difficult to predict. Here, we coupled pediatric magnetic resonance angiography (MRA) neuroimaging data from a completed clinical trial with artificial intelligence (AI)-based techniques to investigate associations between vascular morphology and clinical risk groups. Using an automated computer vision workflow, we generated quantitative metrics describing individual vessels, including small vessels not captured by conventional radiologic scoring. We then applied open-source machine learning algorithms, including: clustering to identify natural groupings within the data, classification to differentiate scans by clinical risk category, and scaled event-based modeling to order vascular features according to relative changes observed across the cohort. Across these approaches, vessel remodeling at branch points and increased vessel tortuosity consistently emerged as among the earliest observed vascular features distinguishing high-risk groups, independent of transcranial Doppler ultrasound velocities. These retrospective findings describe associations between MRA-derived vascular features and clinical groups, but do not establish causality or predict future events. Our results demonstrate that quantitative vascular metrics be extracted from already-obtained imaging data, without requiring additional patient procedures, and may complement existing risk stratification methods. In addition, as current guidelines recommend neuroimaging in school-aged children with SCA, and as MRA technologies continue to advance, our results support larger-scale prospective studies to validate computer vision-based biomarkers. Ultimately, these approaches may inform future studies aimed at improving SCA-related CNS complications and support more refined clinical characterization.