AI-aided methods to automatically segment the Choroid Plexus from brain MRI: the evolution of generalizability, from fine-tuning to federated continuous learning

The Choroid Plexus (ChP), a highly vascularized structure located within the four cerebral ventricles, constitutes a key component of the blood–cerebrospinal fluid barrier and plays a crucial role in regulating brain homeostasis, immune signaling, and waste clearance. Recent evidence has linked alterations in the ChP to a range of neurological and psychiatric conditions, including Multiple Sclerosis (MS), and major depressive disorder. In these contexts, changes in ChP volume (ChPV) have been hypothesized to be associated with neuroinflammation. While most studies have focused on pathological conditions, emerging research aims to characterize ChPV in healthy individuals, with the goal of capturing its variability across the lifespan and advancing its development as a reliable biomarker of brain aging and disease. Given the increasing relevance of ChPV as a potential biomarker, the accurate and reproducible segmentation of the ChP from structural Magnetic Resonance Imaging (MRI) has become a critical methodological objective. However, manual segmentation—currently considered the gold standard—is time-consuming and operator-dependent, thus impractical for large cohorts. Existing automated approaches, including FreeSurfer and Gaussian Mixture Models, provide limited accuracy and generalizability. Advanced Deep Neural Networks (DNN) architectures, such as UNETR or nnU-Net, as well as training strategies such as fine-tuning or federated learning, have not yet been systematically applied to this task, underscoring the need for a robust and scalable DNN-based toolbox for ChP segmentation. The aim of the present research is to develop an automatic, accurate, reliable, and efficient AI-based toolbox for the semantic segmentation of the ChP, thereby promoting the translational potential of ChPV in clinical settings. The work is organized into three main sections. The first section investigates the optimal MRI sequence for ChP imaging, addressing the current lack of systematic evaluation in the literature. While Gadolinium-enhanced T1-weighted (T1-w) MRI is considered the gold standard, its invasive nature limits its applicability, making non-contrast T1-w and Fluid-Attenuated Inversion-Recovery (FLAIR) sequences promising alternatives. Using an MS dataset, this study quantitatively evaluated whether non-contrast acquisitions can reliably substitute the contrast-enhanced imaging without compromising spatial or volumetric accuracy in ChP segmentation. The second section constitutes the core of the thesis: the development of ASCHOPLEX (Automatic Segmentation of CHOroid PLEXus), a DNN-based toolbox designed to achieve accurate and generalizable automatic segmentation of the ChP. ASCHOPLEX generated an ensemble through majority voting across the five best fold configurations derived from five-fold cross-validation training. An implemented fine-tuning procedure enhanced adaptability to new datasets. To address the limitations of the fine-tuning procedure, a second version of ASCHOPLEX was integrated into Dafne, a federated incremental learning framework that enables continuous learning without loss of previously acquired knowledge. This alternative was trained and tested on 2,284 subjects, including both MS patients and healthy controls, demonstrating its potential as a scalable and reproducible tool for ChP segmentation. Finally, the last section describes how ASCHOPLEX was applied to establish normative reference standards for ChPV in healthy adult populations. Employing Normative Modeling (NM), the study delineated normative variability and age-related trajectories of ChPV across 1,036 healthy adults. The model’s validity was further confirmed in diseased cohorts. This work represents the first application of NM to ChPV across the adult lifespan, enabling the precise quantification of individual deviations from normative ranges.