All participants were recruited between 2018 and 2019 as part of the Translational Biomarkers of Aging and Dementia (TRIAD) study conducted at McGill University. Subjects consisted of 19 cognitively normal (CN) controls, 11 mild cognitive impairment (MCI), and 8 subjects diagnosed with AD by a physician using accepted criteria. MCI and AD subjects had a positive Aβ PET scan and a Cognitive Dementia Rating Scale (CDR score between 0.5 and 2). CN subjects had no objective cognitive impairment and a CDR score of 0. All subjects were free from significant medical or neurologic disease (other than AD). The study was approved by the Douglas Mental Health University Institute Research Ethics Board, and written informed consent was obtained from all participants.

Three-dimensional volumetric T1-weighted (T1w) MPRAGE were acquired using a Siemens Skyra 3 T scanner with acquisition parameters: repetition time (TR) = 2,300 ms, echo time (TE) = 2.96 ms, flip angle = 9^o, and voxel size 1 × 1 × 1 mm³. T2-FLAIR images were acquired in the same session with an isotropic voxel size of 1 × 1 × 1 mm³ for all participants. Details of the data acquisition were documented in the previous literature (Butler et al., 2024; Pascoal et al., 2021; Macedo et al., 2024).

All MRI scans underwent a standardized preprocessing pipeline in FreeSurfer v7.1 using the recon-all command, which processes T1w into a FS space with standard regions of interest (ROI) parcellations. Specifically, FS output includes the segmentation of lateral ventricles (LV) and CP, providing anatomical priors for the subsequent segmentation process. The accuracy of the LV mask from FS reconstruction was screened by making a snapshot of the LV mask overlaying T1w, and necessary manual edits were performed to make the LV mask accurate. The FS segmented CP was also used to compare with our proposed method. T2-FLAIR is processed with N3 correction to remove the regional brightness inhomogeneity (Tustison et al., 2010), and coregistered to FS T1w space.

The manual segmentation of CP was performed on the T2-FLAIR image in the FS T1w space by drawing the bright voxels located in the lateral ventricles and verified against the T1w scan for anatomical accuracy. This process was conducted by a trained radiologist, Dr. Savard, and re-examined by Dr. Lussier to guarantee the accuracy and consistency.

Our proposed segmentation approach utilizes a Gaussian Mixture Model (GMM) informed by FS-segmented LV + CP. Specifically, to apply GMM, the T2-FLAIR images in the LV + CP region, eroded by one voxel using a sphere kernel, were applied a 3D Gaussian filter to smooth the intensity of voxels, followed by a normalization step that divided the image values by the non-zero voxel mean. A 3-component GMM model was applied to the result image with our empirical initial mean values μ = [ 0.15 , 1.5 , 4 ] and a corresponding standard deviation (SD) σ = [ 0.02 , 0.1 , 1.5 ] with component proportion [ 0.45 , 0.5 , 0.05 ] . Note that these initial means and standard deviations were obtained from testing on typical cases, and they are not fixed and can easily be adapted for a new dataset by testing on several cases. The GMM assigns each voxel a probability of belonging to the CP or the other two components based on its intensity and spatial location. The component/group that has the highest mean image intensity was assigned as the voxels in CP based on the posterior probability calculation and hard clustering rule, that is, assigns voxels with the highest posterior probability to the CP region. All these processing steps were conducted separately on the left and right sides to avoid the effect of the size variation of CP between hemispheres. The detailed steps of the processing were shown in the flowchart in Figure 1. This region-informed approach guides the GMM to focus on the areas where the CP is anatomically expected, enhancing segmentation accuracy. The implementation code of our proposed One-GMM method, written in MATLAB, will be made available as indicated in the “Data Availability” section.

We compared our region-informed GMM (One-GMM) segmented CP to the FS’s automatic CP segmentation and manual segmentation, as well as the previously reported two-stage GMM (Two-GMM) method (Tadayon et al., 2020), using the following three metrics.

Dice similarity coefficient (DSC) (Zou et al., 2004), with formula DSC = 2TP/(2TP + FP + FN), where TP is the true positive, FP the false-positive, and FN the false-negative. DSC is to measure the overlap between the automatic and the manual CP segmentations, the higher the better.

All statistical analyses were performed in R v4.4.3 within RStudio v2022.07.1. We used the Kruskal–Wallis Rank Sum Test with post hoc Dunn’s test of multiple comparisons following a significant Kruskal-Wallis Test for assessing the diagnostic group differences in segmentation accuracy. The impact of diagnosis (CN, MCI, and AD) on model performance was also examined using Tukey’s Honestly Significant Difference (HSD) test. The paired Wilcoxon Signed Rank Test was used to compare the performance of our proposed One-GMM method with the FS method, as well as the previous Two-GMM method, and it was also used to compare the segmented CP volume using the proposed One-GMM, FS, and previous Two-GMM with manual segmentation. A clinical analysis was conducted to evaluate the diagnostic group difference of segmented CP volume with the Kruskal–Wallis Rank Sum Test. The age and sex effects of segmented CP volume were modeled using linear regression. Statistical significance was set at p < 0.05, and all p-values were adjusted for multiple comparisons when necessary.

Among the total of 40 participants, there were 19 CN, 11 MCI, and 8 AD. Their age, sex, and CP segmentation-related measurements are listed in Table 1. No group differences were found for age, sex, or CP volume from all methods. All methods show no performance difference across diagnostic groups. We observed that the proposed region-informed GMM model outperformed the other methods in all three metrics—DSC, HD95, and VD%—among all diagnostic groups compared with the FS and Two-GMM methods.

Figure 2 shows the examples of three participants (S1, S2, and S3) with varying CP volumes in both axial and coronal views. Figures 2A,C has small (S1), medium (S2), and high (S3) CP volume, respectively. We see that our proposed region-informed One-GMM method gives CP segmentation highly agrees with manually drawn ROIs and outperforms FS and previous Two-GMM methods, whose quantitative comparisons are discussed in the following two subsections.

The paired Wilcoxon Signed Rank Test showed that our proposed region-informed One-GMM method gave consistently slight underestimation of CP volume (V = 11, p < 0.001) compared with manual ground truth, while FS and Two-GMM methods show no over- or under-estimation of CP volume (FS: V = 315, p = 0.21; Two-GMM: V = 456, p = 0.22), as presented in Figure 3A. These findings have no change after accounting for the intracranial volume (ICV), the proposed One-GMM method (V = 14, p < 0.001), FS method (V = 324, p = 0.25), and the previous Two-GMM method (V = 458, p = 0.21), as presented in Figure 3B. Note that the segmentation of CP volume agreement between FS/Two-GMM and manual methods does not mean that the FS/Two-GMM segmentations are accurate. On the contrary, we can see in Figure 3 that CP volumes from FS/Two-GMM methods have both severely underestimation and overestimation, showing their inconsistent performance across participants. The error of CP segmentation from FS/Two-GMM methods is also seen in the second and third columns of Figure 2.

Our region-informed Gaussian Mixture Model (One-GMM) achieved a mean Dice similarity coefficient (DSC) of 0.82 ± 0.04, significantly outperforming (p < 0.001) FS’s automatic CP segmentation with a mean DSC of 0.24 ± 0.11 and the previous Two-GMM method with DSC of 0.66 ± 0.10, as shown in Figure 4A. These results suggest that approximately 82% segmented voxels overlay with manually drawn ROI using our proposed One-GMM method. Additionally, our model demonstrated lower Hausdorff distances (Figure 4B, One-GMM: 6.06 ± 10.32 mm, FS: 26.21 ± 7.32 mm, Two-GMM: 10.58 ± 10.32 mm, all p < 0.001), indicating that our proposed method has better segmentation boundary consistency with the manual method. Moreover, our proposed method has a similar volume difference error (Figure 4C, One-GMM: 20.97 ± 9.53, FS: 24.32 ± 28.13, Two-GMM: 24.27 ± 22.10, all p > 0.05). In summary, all three typical measures for segmentation accuracy showed that our proposed method agrees with manual segmentation well and has superior performance than the FS/Two-GMM methods in the whole cohort and across diagnostic groups.

The Kruskal–Wallis rank sum test showed that in this small cohort there is no diagnostic group difference of CP volume from all four methods (p > 0.05) including manual, FS, the previous Two-GMM, and the proposed One-GMM methods, as shown in Figures 5C–E,G, respectively. In the whole cohort, the linear regression analysis with CP volume as output and age and sex as independent factors showed that the CP volume has no age or sex effect. Restricting to CN group, we observed a normal aging effect of CP volume from manual (t = 2.41, p = 0.02, R² = 0.18; Figure 5B), the previous Two-GMM (t = 4.63, p < 0.01; Figure 5F), and the proposed method (t = 2.05, p = 0.05, R² = 0.12; Figure 5H) show CP volume increases with normal aging, but not saw the same results from FS method-based CP volume (t = 1.29, p = 0.22, R² = 0.01; Figure 5D). No sex effect of CP volume was observed from all methods.

Accurate segmentation of the CP is critical for volumetric/structural and functional analyses and understanding its role in neurodegeneration (Bergsland et al., 2024; Lu et al., 2023; Butler et al., 2023; Hidaka et al., 2024). Our data showed that the CP volume from the manual and proposed methods has a normal aging effect, which is consistent with previous studies (Eisma et al., 2024; Sun et al., 2024), while the FS-based CP volume has no such effect. Although we did not observe the diagnostic group difference in CP volume, which has been reported before (Choi et al., 2022; Jiang et al., 2024), we think this might be due to the small sample size. Another reason for these negative findings here probably arises from the fact that the anatomic change of CP might be later than its functional change (Gião et al., 2022; Delvenne et al., 2024). For instance, a recent study demonstrated that there is no group difference in CP volume between amyloid-positive and negative groups, but there is a group difference in their diffusion tensor imaging-based free water fraction (DTI-FWf) (Xu et al., 2025). Given that our proposed method produces more accurate CP segmentation with a DSC score of approximately 0.82, the group difference might be observed in a larger cohort using our proposed method. The improved precision offered by our model facilitates more reliable investigations into CP-related biomarkers, such as CP volume relative to intracranial volume or other function-related CP inflammation signals (Althubaity et al., 2022; Butler et al., 2023). This is particularly relevant in AD, where subtle changes in CP volume and structure may reflect disease progression or therapeutic response, and its volume alterations in might implicate mechanistic and functional impairment (Cserr, 1971; Althubaity et al., 2022; Choi et al., 2022; Bouhrara et al., 2024; Bouhrara et al., 2024; Rmeily-Haddad et al., 2011). CP is the key region that produces CSF in the brain, which plays crucial roles in protecting the central nervous system (CNS) and is widely considered as the neurofluid in the glymphatic system that works to remove metabolic wastes, including Aβ and tau proteins in AD (Zhou et al., 2024; Agarwal et al., 2024; Iliff et al., 2012; Ozsahin et al., 2025). CP volume changes may be associated with functional alterations, including decreased CSF production and/or brain clearance (Li et al., 2022; Zhou et al., 2024). Accurate assessment of CP volume will facilitate improved understanding of the role of CP in brain fluid clearance and other key functions, such as regulation of neuroinflammation, critical to the pathophysiology of neurodegenerative diseases like AD.

Our proposed CP segmentation method depends on the LV mask from FS reconstruction. While it has been reported that FS v4.5-based LV segmentation for elder and Alzheimer’s disease subjects has an 11% failure rate (Kempton et al., 2011), due to significantly enlarged LV size. In this study, we used FS v7.1, which performs better than its previous versions. In an elderly cohort of 623 participants at the Brain Health Imaging Institute of Weill Cornell Medicine, we found that there are 35 (35/623 = 5.6%) participants whose FS reconstruction has errors in LV segmentation. However, we found none of those 35 failed LV segmentations affects the CP segmentation using our proposed One-GMM method, since the failure of LV segmentation generally underestimates the volume of CP, and the missing parts are usually superior slices of LV and sometimes small regions in inferior LV that are away from CP. Even though we still recommend carefully checking LV segmentation accuracy by taking snapshots of the LV mask overlay on T1w or T2w images. The FS segmented LV and CP masks are adjacent to each other, after merging them there is no gap in between and the erosion steps in our method will only remove voxels at the outer boundary of the merged mask and will not generate gaps between CP and LV, facilitating an accurate segmentation of CP in the following steps.

Our model outperforms FreeSurfer and previous Two-GMM approaches in DSC by integrating regional information that directs the model’s focus to anatomically plausible regions (Storelli et al., 2024; Tadayon et al., 2020; Zhao et al., 2020; Eisma et al., 2024). Our data showed that the proposed One-GMM method yielded to DSC of 0.82 compared with an FS of 0.25 and the previous Two-GMM of 0.66. Although we have not done a comparison with other more sophisticated deep learning-based methods, one previous study has reported DSC 0.72 in a cohort of 98 participants aged between 21 and 89 years (Eisma et al., 2024) using fully connected U-Net and T1w, T2w, and TFLAIR images; another study reported DSC 0.72 in multiple sclerosis patients (Yazdan Panah et al., 2023) using a two-step 3D U-Net and T1w. Our proposed One-GMM method is lightweight and highly adaptive to larger datasets across vendors and study sites. This is especially valuable in AD patients, where ventricular enlargement can mislead models that rely solely on intensity gradients. We compared our method with the previous Two-GMM model mainly because both of them are GMM based and easy and fast to implement for wide clinical applications. It might be ideal to compare the proposed method with more advanced deep learning-based methods. We did not compare our results using deep learning methods, either because the model has no open-source code (Yazdan Panah et al., 2023) or because the model used different data, such as all of T1w, T2w, and T2-FLAIR (Eisma et al., 2024), as we do not have T2w in our dataset. By guiding the model with spatial priors, we reduce false-positives and improve boundary delineation as estimated by DSC, HD95, and VD%. Our results indicate that the performance of our proposed method has no diagnostic group differences within the AD continuum. In contrast, FS segmentation has a higher DSC in AD than the CN group, which is not ideal in clinical applications, as it introduces diagnostic group bias and might introduce bias to the follow-up analysis. In terms of the computational burden, the proposed method is fully automatic and utilizes the outputs of the standard FreeSurfer reconstruction, which is a typical step in current medical imaging research, followed by a light-weight subject-specific GMM processing. With the proposed region-informed segmentation, the One-GMM method takes less than 10 s per subject. Additionally, while our approach utilized FS reconstruction for LV + CP priors, any other image processing pipeline with accurate LV + CP segmentation could also be used with the proposed method.