A Unified Framework for Joint Detection of Lacunes and Enlarged Perivascular Spaces

Lucas He; Krinos Li; Hanyuan Zhang; Runlong He; Silvia Ingala; Luigi Lorenzini; Marleen de Bruijne; Frederik Barkhof

cs.CVMar 4, 2026

A Unified Framework for Joint Detection of Lacunes and Enlarged Perivascular Spaces

Authors: Lucas He, Krinos Li, Hanyuan Zhang, Runlong He, Silvia Ingala, Luigi Lorenzini, Marleen de Bruijne, Frederik Barkhof,

Organizations: Hawkes Institute, University College London, UK · Unit for Lifelong Health and Aging, University College London, UK · Bioengineering Department and Imperial-X, Imperial College London, UK · Department of Diagnostic Radiology, Copenhagen University Hospital, Denmark ·

Abstract

Cerebral small vessel disease (CSVD) markers, specifically enlarged perivascular spaces (EPVS) and lacunae, present a unique challenge in medical image analysis due to their radiological mimicry. Standard segmentation networks struggle with feature interference and extreme class imbalance when handling these divergent targets simultaneously. To address these issues, we propose a morphology-decoupled framework where Zero-Initialized Gated Cross-Task Attention exploits dense EPVS context to guide sparse lacune detection. Furthermore, biological and topological consistency are enforced via a mixed-supervision strategy integrating Mutual Exclusion and Centerline Dice losses. Finally, we introduce an Anatomically-Informed Inference Calibration mechanism to dynamically suppress false positives based on tissue semantics. Extensive 5-folds cross-validation on the VALDO 2021 dataset (N=40) demonstrates state-of-the-art performance, notably surpassing task winners in lacunae detection precision (71.1%, p=0.01) and F1-score (62.6%, p=0.03). Furthermore, evaluation on the external EPAD cohort (N=1762) confirms the model's robustness for large-scale population studies. Code will be released upon acceptance.

Figures & tables

Figure 1: Overview of the Proposed Morphology-Decoupled Framework.

Task/Model	VALDO						EPAD
Task/Model	Segmentation		Detection
Task 1: EPVS	DSC (%) $\uparrow$	NSD (%) $\uparrow$	Recall (%) $\uparrow$	Precision (%) $\uparrow$	F1 (%) $\uparrow$	FP/Subject $\downarrow$	$\rho$ (BG) $\uparrow$	$\rho$ (CSO) $\uparrow$	$\rho$ (MB) $\uparrow$
DYN Unet	$36.9\pm 9.7$	$56.9\pm 12.5$	$50.2\pm 1.8$	$54.9\pm 16.2$	$48.6\pm 9.8$	$21.0\pm 11.9$	$0.13\pm 0.02$	$0.17\pm 0.02$	$0.10\pm 0.02$
MedNeXt	$37.9\pm 10.2$	$55.6\pm 12.0$	$52.4\pm 7.5$	$46.1\pm 13.2$	$45.7\pm 8.7$	$30.6\pm 19.5$	$0.09\pm 0.02$	$0.09\pm 0.02$	$0.06\pm 0.02$
Swin-UNETR V2	$41.9\pm 9.2$	$62.2\pm 11.9$	$\mathbf{58.4\pm 5.6}$	$55.0\pm 19.7$	$50.5\pm 10.7$	$30.4\pm 23.6$	$0.02\pm 0.02$	$0.05\pm 0.02$	$0.03\pm 0.02$
VISTA-3D	$36.3\pm 8.0$	$57.6\pm 9.8$	$51.0\pm 8.3$	$49.7\pm 10.7$	$47.7\pm 9.1$	$23.7\pm 7.6$	$0.14\pm 0.03$	$0.17\pm 0.02$	$0.10\pm 0.03$

Table 1: Comparison with SOTA methods on the VALDO and EPAD datasets. Bold indicates best numerical results. * denotes statistical significance. Spearman’s rank correlation and Pearson’s correlation coefficient are denoted by

\rho

and

r

, respectively.

Figure 2: Qualitative visualization on the VALDO dataset. Ground truth lesions (cyan), true positive predictions (green), and false positive predictions (red) are shown.

Task/model			Segmentation		Detection
$\mathbf{L_{MT}}$	$\mathbf{L_{exc}}$	$\mathbf{L_{clDice}}$	DSC (%) $\uparrow$	NSD (%) $\uparrow$	Recall (%) $\uparrow$	Precision (%) $\uparrow$	F1 (%) $\uparrow$	FP/Subject $\downarrow$
Task 1: EPVS
✓			$35.6\pm 9.4$	$54.5\pm 12.0$	$45.6\pm 7.6$	$58.9\pm 17.3$	$47.7\pm 9.6$	$17.1\pm 14.2$
✓	✓		$35.7\pm 9.5$	$54.2\pm 11.9$	$46.9\pm 7.7$	$59.2\pm 17.5$	$48.9\pm 9.8$	$15.3\pm 13.6$
✓		✓	$\mathbf{36.9\pm 9.7}$	$\mathbf{54.7\pm 12.1}$	$46.6\pm 7.7$	$\mathbf{62.6\pm 19.1}$	$49.2\pm 9.9$	$\mathbf{16.4\pm 12.3}$
✓	✓	✓	$34.3\pm 8.0$	$51.1\pm 12.4$	$\mathbf{47.1\pm 7.5}$	$60.4\pm 10.7$	$\mathbf{50.0\pm 6.2}$	$16.5\pm 11.6$

Table 2: Ablation of objective functions on VALDO. Bold indicates numercial best

Task/model			Segmentation		Detection
Method	Decoder	Anatomical	DSC (%) $\uparrow$	NSD (%) $\uparrow$	Recall (%) $\uparrow$	Precision (%) $\uparrow$	F1 (%) $\uparrow$	FP/Subject $\downarrow$
Task 1: EPVS
STL-Full	–	–	$12.4\pm 17.3$	$20.4\pm 27.5$	$25.9\pm 17.9$	$21.4\pm 18.0$	$18.3\pm 14.8$	$94.8\pm 91.9$
STL-Mix	–	–	$36.9\pm 9.7$	$56.9\pm 12.5$	$\mathbf{50.2\pm 1.8}$	$54.9\pm 16.2$	$48.6\pm 9.8$	$21.0\pm 11.9$
MTL	Shared		$34.3\pm 8.0$	$51.1\pm 12.4$	$47.1\pm 8.5$	$60.4\pm 10.7$	$50.0\pm 6.2$	$16.5\pm 11.6$
MTL	Shared	✓	$35.3\pm 7.1$	$52.8\pm 10.8$	$47.1\pm 8.5$	$60.8\pm 12.7$	$49.8\pm 6.7$	$16.1\pm 11.1$

Table 3: Comparison of supervision strategies, cross-task attention decoupling, and anatomical calibration on the VALDO dataset. Bold indicates numerical best results.

Neuroimaging

Neuroimaging uses brain measurements and computational methods to map anatomy, activity, connectivity, and disease.

See all 235 papers

Multi-task learning for the automatic grading of enlarged perivascular space burden using MRI
Sep 29, 2026Jesse Phitidis, William N. Whiteley, Joanna M. Wardlaw +11Multi-Task Learning Medical Image Classification
Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, UK · Canon Medical Research Europe, Edinburgh, UK · UK Dementia Research Institute, Centre at The University of Edinburgh, Edinburgh, UK +5
Adaptive Post-Processing Drives Instance-Level Detection in Stroke Lesion Segmentation
Aug 17, 2026Qinghui Liu, Jon André Ottesen, Atle Bjørnerud +1Medical Image Segmentation Medical Image Analysis
Oslo University Hospital, Norway
VasGuideNet: Vascular Topology-Guided Couinaud Liver Segmentation with Structural Contrastive Loss
Feb 25, 2026Chaojie Shen, Jingjun Gu, Zihao Zhao +4Medical Image Segmentation 3D Medical Image Segmentation
Pujian Medical Technology Co., Ltd., Hangzhou, China · Keyi College of Zhejiang Sci-tech University, ShaoXing, China · College of Computer Science and Technology, Zhejiang University, Hangzhou, China

Sep 29, 2026J/K move · Enter open · S save

Multi-task learning for the automatic grading of enlarged perivascular space burden using MRI

Jesse Phitidis, William N. Whiteley, Joanna M. Wardlaw, Miguel O. Bernabeu+10

Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, UK · Canon Medical Research Europe, Edinburgh, UK · UK Dementia Research Institute, Centre at The University of Edinburgh, Edinburgh, UK+5

Multi-Task Learning Medical Image Classification Medical Image Segmentation

Open paper AskPDF arXiv

Enlarged perivascular spaces (PVS) visible in brain magnetic resonance imaging (MRI) are increasingly thought to be linked to poor brain health. PVS are elongated structures of less than 3 mm in diameter and can be numerous. To reflect the incidence of PVS, radiologists visually score their burden following a clinical grading scale - a task that would benefit from automation to accelerate analyses and overcome the influence of inter-observer differences. We developed and evaluated methods for training machine learning models to score PVS incidence in the basal ganglia (BG) and centrum semiovale (CSO) leveraging the Potters/Wardlaw scale. The novelty in our work lies in the use of imperfect, semi-automatically generated "silver-standard" PVS segmentation masks during training, in addition to PVS radiological scores. We comparatively evaluated a conditional convolutional neural network (CNN) which accepts PVS masks as an extra input channel, a multi-task CNN which performs both PVS segmentation and scoring, and a logistic regression model which utilises features derived from PVS masks to predict PVS scores. Multi-task learning was the most effective method, achieving a mean average precision of 64.08% compared to 60.22% for the conditional CNN, 52.11% for a baseline CNN trained only to predict PVS scores, and 49.32% for the logistic regression model. The multi-task model showed an ability to localise individual PVS not shown by the other CNNs, and behaved in a probabilistically sensible way, predicting with lower confidence on inherently harder classes. Age, sex, hypertension status, white matter hyperintensity volume, and ischaemic stroke lesion status were shown to be associated with the multi-task model's PVS score predictions and the ground truth in a similar way.

Abstract

Neuroimaging

Explore similar work

Multi-task learning for the automatic grading of enlarged perivascular space burden using MRI

Adaptive Post-Processing Drives Instance-Level Detection in Stroke Lesion Segmentation

VasGuideNet: Vascular Topology-Guided Couinaud Liver Segmentation with Structural Contrastive Loss