An MRI-based artificial intelligence system that learned to grade enlarged perivascular spaces performed best in a test against three other approaches, according to an arXiv preprint. The multi-task model recorded a mean average precision, or mAP, of 64.08%, compared with 60.22% for a conditional convolutional neural network, 52.11% for a baseline CNN and 49.32% for logistic regression.
Mean average precision is the paper's summary score for how well a system predicted the score categories. The winning system was called multi-task because it was trained to predict both the PVS score and a map of PVS locations.
The task was narrower than it sounds
The target was PVS burden in two brain regions: the basal ganglia, or BG, and the centrum semiovale, or CSO. The researchers drew on five MRI datasets; subjects from four of the five had observer scores for both regions, and some subjects also had PVS segmentation masks.
Because scores of 0 and 4 were rare, the researchers merged 0 into 1 and 4 into 3. The models therefore worked on a modified Potters/Wardlaw scale, rather than the full set of original score categories.
The data were divided into training, validation and test sets in a 70%, 10% and 20% split, with the division stratified by BG and CSO score. Six VALDO cases without scores were added to training for segmentation and multi-task learning.
The comparison included a baseline CNN without segmentation information; a conditional CNN that used U-Net PVS predictions; a multi-task CNN that predicted scores and masks together; and logistic regression using segmentation-derived features plus age or sex.
A stronger score, with a mixed picture
The multi-task approach was reported as best or joint-best across the wider set of metrics. It was also the only method reported to show a statistically significant difference from logistic regression on mean F1, another summary of classification performance, with p = 0.03. To estimate uncertainty, the researchers used 10,000 bootstrap samples of the test set to calculate 95% confidence intervals. P-values came from pairwise bootstrap differences, with a Bonferroni correction across six comparisons.
The result was accompanied by a localization finding. The authors reported that the multi-task model localized individual PVS in a way not shown by the other CNNs.
The model's confidence was also lower in the middle category. On both BG and CSO, it was less confident for score 2 than for scores 1 or 3. Score 1 included 10 or fewer PVS, score 2 represented 11 to 20, and score 3 represented 21 or more.
In an ordinal-logistic analysis, a method for comparing ordered score categories, predicted scores and observer scores showed similar normalized beta directions and matching correlation directions. WMH volume was the most statistically significant predictor in those analyses.
That result describes a shared pattern in the data, not a cause-and-effect finding. It indicates that the model's scores tracked relationships seen in observer scores, but the analysis does not establish that PVS burden changes WMH volume or any other clinical measure.
What the preprint does not settle
Human readers still agreed with one another more closely than the model agreed with them. Using Cohen's kappa as its agreement measure, the analysis found observer-observer values of 0.66 for BG and 0.36 for CSO, while model-observer agreement was lower.
The paper's clearest caution is that it used a single data split, a limitation the authors identified despite using bootstrapping and ensembling. The segmentation U-Net's overlap scores against silver-standard masks were 65.06% on training data, 63.45% on validation data and 62.84% on test data.
The evidence supports a narrower conclusion: in this comparison, the multi-task approach produced the strongest score-prediction result, while model-observer agreement remained below observer-observer agreement. The modified scale and single split define important boundaries around that conclusion. The manuscript is an arXiv preprint, version 1, dated 29 Sep 2026.
Paper data and sources
Original title: Multi-task learning for the automatic grading of enlarged perivascular space burden using MRI
Authors: Jesse Phitidis, William N. Whiteley, Joanna M. Wardlaw et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text