Generalizable and equitable automated ischaemic stroke lesion segmentation with vision transformers

Ischaemic stroke, a leading cause of death and disability, relies on neuroimaging for characterising the anatomical pattern of injury. Diffusion-weighted MRI (DWI) provides the most anatomically specific signal in acute ischaemic stroke but poses substantial challenges for automated lesion segmentation due to susceptibility artefacts, lesion morphological heterogeneity, comorbid pathology, instrumental variability, and limited labelled data. Current U-Net-based models therefore underperform, a problem accentuated by evaluation metrics that neglect anatomical, subpopulation and acquisition-dependent variability.

We train 3D vision transformer–based segmentation models on a multi-site DWI dataset comprising 3563 annotated lesion-positive and 6900 lesion-negative volumes, using balanced cross-validation splits. We compare these models with U-Net baselines and an nnU-Net configuration under harmonised augmentation and training schemes, and introduce an evaluation framework that quantifies fidelity, anatomical precision, robustness to image corruption and equity across demographic and lesion-defined subtypes. Here, we show that transformer-based models with our proposed control-image regularisation achieve higher segmentation performance than U-Net-based approaches on clinically realistic data, while substantially reducing false positives in lesion-negative images.

They exhibit more stable performance across lesion sizes, anatomical territories, image quality and patient subgroups, indicating improved epistemic equity relative to convolutional architectures. This work reconciles model expressivity with domain-specific challenges and redefines performance benchmarks to prioritise equity and generalisability–critical for personalised medicine and mechanistic research. These findings establish vision transformer architectures, combined with equity-aware validation, as a powerful approach for ischaemic stroke lesion segmentation on DWI.

Stroke is a leading cause of long-term disability. Doctors use brain scans to see which parts of the brain have been damaged, but manually outlining stroke lesions is slow and varies between experts. We trained an artificial intelligence system on more than ten thousand brain scans from multiple hospitals to automatically identify and measure stroke lesions.

Our system is based on a modern “vision transformer” architecture, and we carefully compared it to established methods under matched conditions. The transformer-based model was more accurate, more robust to image quality differences and showed fewer systematic performance differences between patient groups and lesion types. We release our code and trained models so that other teams can build on this work and explore clinical translation.

Supported by the Wellcome (213038), the NIHR UCLH Biomedical Research Centre (NIHR-INF-0840), EPSRC (2252409), and the Medical Research Council (MR/X00046X/1). This work benefited from the support of the project HoliBrain of the French National Research Agency (ANR-23-CE45-0020-01). Moreover, this project is supported by the Precision and global vascular brain health institute funded by the France 2030 investment plan as part of the IHU3 initiative (ANR-23-IAHU-0001).