Title: StrokeTimer: A Robust Representation Learning Approach for Estimating Ischemic Stroke Onset Time via Non-contrast CT

Abstract

Ischemic stroke represents a significant global health burden. Clinical management is critically dependent on timing, as the window for reperfusion therapies is determined by the duration between the initial stroke event and the delivery of treatment. In many clinical scenarios, the exact moment of onset remains unknown, making imaging-based evaluations of tissue age a necessary surrogate indicator. However, early signs of ischemia on standard non-contrast CT (NCCT) scans are frequently faint. Furthermore, real-world data often suffers from severe class imbalance regarding onset times and significant variability linked to different scanning centers and equipment.

To address these challenges, we introduce StrokeTimer, a completely automated system designed to estimate onset times in acute ischemic stroke cases. This framework combines self-supervised disentanglement learning with energy-guided contrastive learning. This hybrid approach allows the model to detect subtle ischemic patterns while effectively managing long-tailed data distributions and variability across different acquisition methods. For clinical utility, onset times are divided into three distinct, relevant time windows.

We evaluated StrokeTimer using a substantial multi-center NCCT dataset derived from two national cohorts: the MR CLEAN Registry and the MR CLEAN LATE study. The results demonstrate that StrokeTimer attained a macro AUC of 0.69 and a macro F1-score of 0.57. This performance marks an improvement of nearly 50% over the best-performing baseline model (p < 0.005). In this complex and realistic environment, standard baseline methods struggled, performing at levels close to random chance.

Further analysis of model explanations revealed attention to subtle gray-white matter blurring and hypodense areas, which align with established radiological biomarkers. These outcomes suggest that StrokeTimer holds considerable promise for assisting in treatment decisions for acute ischemic stroke. The source code for this project is publicly accessible at https://github.com/BrainVas/StrokeTimer.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC