Deep learning segmentation of the choroid plexus from structural magnetic resonance imaging (MRI): validation and normative ranges across the adult lifespan

Abstract

Background: The choroid plexus functions as the blood-cerebrospinal fluid (CSF) barrier, plays an important role in CSF production and circulation, and has gained increased attention in light of the recent elucidation of CSF circulation dysfunction in neurodegenerative conditions. However, methods for routinely quantifying choroid plexus volume are suboptimal and require technical improvements and validation. Here, we propose three deep learning models that can segment the choroid plexus from commonly-acquired anatomical MRI data and report performance metrics and changes across the adult lifespan.

Methods: Fully convolutional neural networks were trained from 3D T₁-weighted, 3D T₂-weighted, and 2D T₂-weighted FLAIR MRI using gold-standard manual segmentations in control and neurodegenerative participants across the lifespan (n = 50; age = 21-85 years). Dice coefficients, 95% Hausdorff distances, and area-under-curve (AUCs) were calculated for each model and compared to segmentations from FreeSurfer using two-tailed Wilcoxon tests (significance criteria: p < 0.05 after false discovery rate multiple comparisons correction). Metrics were regressed against lateral ventricular volume using generalized linear models to assess model performance for varying levels of atrophy. Finally, models were applied to an expanded cohort of adult controls (n = 98; age = 21-89 years) to provide an exemplar of choroid plexus volumetry values across the lifespan.

Results: Deep learning results yielded Dice coefficient = 0.72, Hausdorff distance = 1.97 mm, AUC = 0.87 for T₁-weighted MRI, Dice coefficient = 0.72, Hausdorff distance = 2.22 mm, AUC = 0.87 for T₂-weighted MRI, and Dice coefficient = 0.74, Hausdorff distance = 1.69 mm, AUC = 0.87 for T₂-weighted FLAIR MRI; values did not differ significantly between MRI sequences and were statistically improved compared to current commercially-available algorithms (p < 0.001). The intraclass coefficients were 0.95, 0.95, and 0.96 between T₁-weighted and T₂-weighted FLAIR, T₁-weighted and T₂-weighted, and T₂-weighted and T₂-weighted FLAIR models, respectively. Mean lateral ventricle choroid plexus volume across all participants was 3.20 ± 1.4 cm³; a significant, positive relationship (R² = 0.54-0.60) was observed between participant age and choroid plexus volume for all MRI sequences (p < 0.001).

Conclusions: Findings support comparable performance in choroid plexus delineation between standard, clinically available, non-contrasted anatomical MRI sequences. The software embedding the evaluated models is freely available online and should provide a useful tool for the growing number of studies that desire to quantitatively evaluate choroid plexus structure and function ( https://github.com/hettk/chp_seg ).

Keywords: Cerebrospinal fluid; Choroid plexus; Deep learning; Glymphatic; Neurofluids; Segmentation.

Fig. 1

Overview of the processing pipeline of the anatomical magnetic resonance imaging (MRI) utilized in the proposed deep learning method. Examples are shown for a T₁-weighted MRI, but this pipeline also was utilized for T₂-weighted and T₂-weighted FLuid-Attenuated Inversion Recovery (FLAIR) MRI. Input images were registered to MNI152 space and cropped around the choroid plexus based off a probabilistic atlas generated from ground truth manual choroid plexus segmentations. Cropped images were then used as training input for the 3D U-NET fully convolutional neural network. The number of inputs for each trained model was 1 and the number of output structures was 1 (i.e., choroid plexus). Cropped outputs were then decropped and inverse transformed to the native imaging space. Example images are shown from a 69 year old male with Parkinson’s disease. (MRI: magnetic resonance imaging; Conv: convolution; ReLu: rectified linear unit; Tanh: hyperbolic tangent)

Fig. 2

Example choroid plexus segmentations from machine learning models in a 53 year old male with mild cognitive impairment. From left to right, columns show results from T₁-weighted images, T₂-weighted images, and T₂-weighted FLAIR images. The first row (panels a–c) shows the anatomical MRI sequence utilized in this study for deep learning training, and the second row (d–f) shows these same images magnified on the lateral ventricles where the majority of the choroid plexus resides. The remaining rows show the manual segmentations (g–i), machine learning output segmentations (j–l), and the overlay of these segmentations in axial slices (m–o) and 3D renderings (p–r) for each type of MRI contrast. The 3D renderings show the manual segmentation in blue (i.e., under-segmentation), the machine learning segmentation in red (i.e., over-segmetntation, and the overlap between the two in white. The Dice scores of each model (T₁-weighted: 0.78, T₂-weighted: 0.78, T₂-weighted FLAIR: 0.80) are shown and reflect consistently accurate performance across MRI sequences. (MRI: magnetic resonance imaging; FLAIR: FLuid-Attenuated Inversion Recovery)

Fig. 3

Example T₁-weighted sequence from a 53 year old male with mild cognitive impairment (a–d) and manual tracings (e–f) utilized in training of the machine learning methods. Example outputs from the T₁-weighted trained machine learning model (g–h) are shown compared to FreeSurfer segmentations (i–j). Dice scores are shown for machine learning and FreeSurfer outputs and reflect an improvement in segmentation accuracy for the proposed method

Fig. 4

Regression plots for machine learning dice score (a), 95% Hausdorff distance (b), and AUC (c) against testing subject lateral ventricular volume. Overall, models performed consistently across lateral ventricular volume. The only model metric that was significantly related to lateral ventricular volume was the T₁-weighted model’s 95% Hausdorff distance (ß value = 0.015; p-value = 0.05). (MRI: magnetic resonance imaging; FLAIR: FLuid-Attenuated Inversion Recovery)

Fig. 5

Bland-Altman plots for choroid plexus volumes generated from T₁-weighted deep learning methods (a), T₂-weighted deep learning methods (b), T₂-weighted FLAIR deep learning methods (c), and FreeSurfer (d) compared to the ground truth manual segmentation volumes. The intraclass correlation coefficient between T₁-weighted and ground truth choroid plexus volumes was 0.83, T₂-weighted and ground truth choroid plexus volumes was 0.82, T₂-weighted FLAIR and ground truth choroid plexus volumes was 0.82, and FreeSurfer and ground truth choroid plexus volumes was 0.00

Fig. 6

a Regression plot displaying choroid plexus volume against participants’ age for each MRI modality in adult controls. McFadden’s R² values are reported for each regression model. b Case examples for younger, middle, and older-aged controls showing an increase in choroid plexus volume with age. Results show consistently across MRI types that choroid plexus volume increases with age across the adult lifespan. 3D renderings are shown from the T₂-weighted segmentations to provide further support of this finding. (FLAIR: FLuid-Attenuated-Inversion-Recovery)

Fig. 7

Bland-Altman plots for choroid plexus volumes generated from T₁-weighted and T₂-weighted FLAIR deep learning methods (a), T₁-weighted and T₂-weighted deep learning methods (b), and T₂-weighted and T₂-weighted FLAIR deep learning methods (c). The intraclass correlation coefficient between T₁-weighted and T₂-weighted FLAIR choroid plexus volumes was 0.95, T₁-weighted and T₂-weighted choroid plexus volumes was 0.95, and T₂-weighted and T₂-weighted FLAIR choroid plexus volumes was 0.96