Poster #18: Elizabeth Levitis PhD

ClinSeg: Robust Brain Segmentation for Clinically Acquired Pediatric MRI

Elizabeth Levitis, PhD1,2,3, Henry F. J. Tregidgo, PhD4, Dabriel Zimmerman, BA1,2,3, Benjamin Jung, PhD1,2,3, Shivaram Karandikar, BA1,2,3, Margaret Gardner, BS1,2,3, Philip Mattisson, BS1,2,3, Eren Kafadar, BS1,2,3, Anna Zapaishchykova, PhD5,6, Benjamin H. Kann, MD5,6, Susan T. Sotardi, MD7,8, Arastoo Vossough, MD, PhD7,8, Hao Huang PhD7,8, Benjamin Billot, PhD9, Juan Eugenio Iglesias PhD10, Daniel C. Alexander, PhD4, Aaron Alexander-Bloch, MD, PhD1,2,3*, Jakob Seidlitz, PhD1,2,3* Affiliations: 1. Department of Child and Adolescent Psychiatry and Behavioral Sciences, Children's Hospital of Philadelphia, Philadelphia, Pennsylvania 2. Neurodevelopment & Psychosis Section, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania 3. Lifespan Brain Institute, Children's Hospital of Philadelphia and Penn Medicine, Philadelphia, Pennsylvania. 4. UCL Hawkes Institute, Department of Computer Science, University College London, London, United Kingdom. 5. Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Boston, MA, USA 6. Department of Radiation Oncology, Dana-Farber Cancer Institute, Brigham and Women's Hospital, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA 7. Department of Radiology, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania 8. Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA 9. Université Côte d’Azur, Inria, Epione team, Sophia Antipolis, France 10. Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Boston, MA, United States 11. Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, United States

Clinical brain MRIs from pediatric health systems offer an unprecedented resource for modeling real-world neurodevelopmental trajectories and identifying early disease risk. While existing automated tools now enable the extraction of volumetric data from such scans, they largely fail in infants below 6 months of age due to rapid myelination, low signal-to-noise ratio, and dynamic tissue contrast, resulting in the exclusion of early-life data. To resolve this bottleneck, we present ClinSeg, a deep-learning framework engineered for pediatric clinical MRI segmentation across highly variable orientations, spatial resolutions, and contrast profiles. By combining deep-learning image registration and synthetic data generation with a public repository of paired infant scans and manually corrected segmentations, we built a training corpus for a 3D U-Net that captures broad tissue-contrast diversity, normative clinical heterogeneity, and morphological abnormalities from a pediatric hospital cohort. We validated ClinSeg against manual segmentations on research-grade datasets and evaluated its performance in real-world clinical scans against state-of-the-art pipelines across the 0-21 year age range. ClinSeg significantly outperformed existing tools in infants aged 0–1 years, achieving superior segmentation fidelity while matching established models in childhood, adolescence, and adulthood. On routine clinical scans, ClinSeg reliably processed data across all age groups while markedly surpassing current tools in early infancy. ClinSeg enabled the construction of normative brain growth trajectories in a sample of 11,699 individuals aged 0–21 years. Finally, applying ClinSeg to clinical MRIs of youth with 22q11.2 deletion syndrome (22q11DS) allowed the inclusion of previously unanalyzable infant scans, successfully replicating known case-control differences while uncovering novel age-by-diagnosis interactions. ClinSeg bridges a critical gap in pediatric neuroimaging by enabling segmentation from infancy through early adulthood within a single framework. By incorporating infant data into normative growth models, ClinSeg facilitates benchmarking of individual scans and advances the discovery of early neurodevelopmental phenotypes in real-world populations.