Developing a Deep Learning Segmentation Tool for the Choroid Plexus – FastPlex

Poster Session C - Sunday, April 14, 2024, 5:00 – 7:00 pm EDT, Sheraton Hall ABC

Paulo Lizano¹ (plizano@bidmc.harvard.edu), Ling-yu Huang¹, Victor Zeng¹, David Kuegler², Yuan Cao³, Deepthi Bannai¹, Martin Reuter²; ¹BIDMC, ²DZNE, ³Sichuan University

The choroid plexus (ChP) is located within the four ventricles of the brain and forms the blood-cerebrospinal fluid (CSF) barrier. The ChP plays an important role in cognition and neuropsychiatric disorders. Neuroanatomical alterations of the ChP have been identified in autism, schizophrenia and Alzheimer’s disease. The delay in ChP research is due to a lack of readily accessible, automated tools that can reliably label the ChP in structural images. Manual segmentation is currently the “gold-standard”, but as the field moves toward multi-site consortiums with large samples, it is time consuming and non-sustainable. Here, we introduce an ongoing effort to create a fast, reliable, machine learning-based tool for ChP labeling on T1-weighted MRI images. Sampling 22,000+ brain MRIs from open-source neuroimaging databases such as the Human Connectome Project (HCP), Connectome Studies Related to Human Disease (CRHD), Adolescent Brain Cognitive Development (ABCD), and others we plan to manually segment the ChP from a representative subset (N=700). This subset will be generalizable since it includes a wide variety of image resolutions, scanner manufacturers, subjects’ age and sex, as well as clinical status. The representative subset will be used to train, test and validate the deep learning algorithm (FastPlex) and it will be made compatible with the FastSurfer pipeline. The manual labels and the FastPlex will both be publicly available upon completion, with the goal to help speed up investigations of ChP’s functional role in aging, development, cognition, and neuropsychiatric conditions.

Topic Area: METHODS: Neuroimaging