Deep learning-based segmentation of choroid plexus calcification on quantitative susceptibility mapping: a feasibility study
TL;DR
This framework provides a tool for direct CPC quantification on QSM that shows strong correspondence with CT-derived volumes, enabling large-scale, longitudinal investigations into the clinical implications of CPC.
Abstract
Choroid plexus calcification (CPC) is a common age-related imaging finding associated with neuroinflammation and cognitive decline. Its accurate quantification remains challenging on computed tomography (CT), the current clinical gold standard, due to concerns regarding ionizing radiation exposure. Quantitative susceptibility mapping (QSM) offers a non-invasive, radiation-free alternative for accurate assessment, but manual segmentation is impractical for large cohorts. We developed a deep learning-based framework for automated CPC segmentation using QSM. Using data from 106 subjects, we trained nnU-Net v2 models to evaluate input combinations comprising QSM, 3D T1-weighted images, and choroid plexus (CP) masks. The automated pipeline achieved its highest segmentation accuracy using QSM with CP and CPC masks and the strongest volumetric correlation with manual ground-truth annotations ( r = 0.964) using both QSM and T1-weighted images. QSM-derived CPC volumes were strongly correlated with CT-derived volumes in a validation subset of 22 subjects ( r = 0.728). In an additional cohort of 91 subjects, the framework revealed age-related changes in CPC, with volume increasing and susceptibility values becoming more diamagnetic with age. This framework provides a tool for direct CPC quantification on QSM that shows strong correspondence with CT-derived volumes, enabling large-scale, longitudinal investigations into the clinical implications of CPC.