EPVS Challenge shows top deep‑learning models reach Dice >0.80 on known sites but lose >20% accuracy on unseen Shanghai cohort

The Numbers

Wu and colleagues evaluated 200 MRI scans—100 for training, 50 for validation, and 50 for testing—collected across the UK, Singapore, and China. The top‑performing MedNeXt model achieved a Dice similarity coefficient (DSC) exceeding 0.80 on the validation and test sets drawn from sites represented in the training data. When applied to the previously unseen Shanghai cohort, DSC dropped by more than 20 %, falling below 0.64. Reported secondary metrics included absolute volume difference, recall, and precision, though exact numeric values were not disclosed.

Context: What This Study Is

The work was led by Wu at the Centre for Sleep and Cognition (CSC) and published in Medical Image Analysis (2026). It describes the EPVS Challenge organized for MICCAI 2024, a community‑wide benchmark that solicited automated segmentation algorithms for enlarged perivascular spaces (EPVS) in brain MRI. Teams submitted deep‑learning pipelines—most based on U‑Net variants, with the winner employing a MedNeXt dual 2D/3D architecture. All 200 scans were annotated according to the STRIVE protocol, providing a standardized ground truth across diverse acquisition parameters.

What This Result Means

Achieving DSC >0.80 demonstrates that current convolution‑ and transformer‑based models can reliably delineate EPVS when training and test data share similar scanner settings and population characteristics. The sharp decline on the Shanghai cohort reveals a domain‑shift problem: subtle differences in pulse sequences, voxel anisotropy, or demographic factors can disrupt feature extraction, leading to mis‑segmentation. This underscores the need for algorithms that learn invariant representations rather than overfitting to site‑specific image textures.

Study Limitations

• Absence of quantitative values for secondary metrics (absolute volume difference, recall, precision) limits a full performance profile.


• The challenge dataset, while multi‑site, contains only 200 scans; rare acquisition protocols may be under‑represented.


• Evaluation focused on offline benchmark scores; real‑time clinical integration and robustness to pathological variability were not tested.

Practical Application

It's too early to apply these findings to a personal protocol for EPVS assessment. Validation on larger, prospectively collected cohorts and demonstration of consistent performance across a broader range of scanners are required before clinicians can rely on automated EPVS quantification in routine practice.

Disclaimer: This article is for informational and educational purposes only. The information presented does not constitute medical advice, diagnosis, or treatment. Consult a qualified healthcare professional before modifying your diet, supplementation, or exercise routines. The scientific studies cited reflect the state of knowledge at their publication date and may be subject to revision.