If you’re interested in testing DRIPS on your own data or collaborating on new applications, feel free to reach out.

Grateful to Luna Bitar and Mario Díaz for their outstanding work, and to all collaborators who contributed through data collection, manual segmentation, and writing — this project truly was a collective effort.

The segmentation performance of DRIPS remained stable regardless of white matter hyperintensity (WMH) burden, unlike other methods that were biased by their presence.

DRIPS generally outperformed both classical filtering approaches and more recent deep-learning methods, achieving consistent PVS segmentation across imaging data from participants with diverse health conditions.

Out-of-sample, DRIPS accurately segmented PVS across real T1- and T2-weighted scans, isotropic and anisotropic resolutions, and even an ex vivo histology reconstruction — without needing retraining or fine-tuning.

DRIPS synthesises brain images with PVS-like structures and corresponding label maps on the fly, combining anatomical priors with simulated artefacts such as noise, motion, and bias fields to teach networks robust representations.

Current segmentation methods often fail when moving across scanners or cohorts — they depend on extensive manual annotations, retraining, or contrast-specific tuning. This is not sustainable in the long run.

As a team, we developed targeted recommendations aimed at accelerating the integration of ML+CSVD into clinical practice.

In need of improvement: Serious methodological issues—such as inconsistent reporting, limited testing of generalisability, and other potential biases—remain common and continue to hinder broader adoption.

Positive: There is growing interest in incorporating CSVD markers into ML models for classifying neurodegenerative diseases, and current performance indicates that this approach merits further investigation.

We carried out a systematic review and meta-analysis to examine the use of neuroimaging CSVD markers in machine learning–based diagnosis and prognosis of cognitive impairment and dementia, while also identifying methodological trends over time and barriers to clinical translation.

Take home 🏠 message:
Cerebrovascular and ageing-related atrophic changes are interrelated and shape cognitive trajectories. Trajectories may nonetheless be modified for the better by managing cerebrovascular and mental health while fostering cognitive engagement.

(B) WMH progression and ageing-related atrophy were independently associated with changes in cognitive performance, underscoring their relevance for brain maintenance.

(C) Neuroticism, depression, and low cognitive engagement predicted worse domain-specific trajectories.

(A) WMH not only co-occur with medial temporal lobe atrophy but may also contribute to and accelerate its progression. We observed this coupling also in a prior study (doi.org/10.1186/s131...). This finding suggests that preserving microvasculature may reduce vulnerability to neurodegeneration.

Using four-year data from 543 cognitively unimpaired participants from the DELCODE study and latent growth curve modelling, we examined the co-evolution of white matter hyperintensities (WMH), medial temporal lobe to ventricle ratio, and cognitive performance.

I had no clue you were working on CSVD!!!