Abstract
Stroke remains a leading cause of global morbidity and mortality, imposing a heavy socioeconomic burden. Advances in endovascular reperfusion therapy and CT and MR imaging for treatment guidance have significantly improved patient outcomes. Developing machine learning algorithms that can create accurate models of brain function from stroke images for tasks like lesion identification and tissue survival prediction requires large, diverse, and well-annotated public datasets. While several high-quality image datasets in stroke exist, they typically include only single time point data. Data over different time points are essential to accurately identify lesions and predict prognosis. Here, we provide comprehensive longitudinal stroke data, including (sub-)acute CT imaging with angiography and perfusion, follow-up MRI after 2–9 days, and acute and longitudinal clinical data up to a three-month outcome. The dataset also includes vessel occlusion masks from acute CT angiography and delineated infarction masks in follow-up MRI. This multicenter dataset consists of 245 cases and provides a solid basis for developing powerful machine-learning algorithms to facilitate clinical decision-making.
Introduction: The Need for Longitudinal Stroke Data
In the high-stakes environment of acute stroke care, clinical decisions regarding invasive procedures like mechanical thrombectomy rely on estimating the "ischemic core" (irreversibly damaged tissue) versus the "penumbra" (salvageable tissue). While AI has shown promise in segmenting these areas, most existing public datasets are cross-sectional "snapshots." This limitation hinders the development of models capable of predicting infarct growth—how much more brain tissue will die if a specific treatment is or is not administered. To transition from descriptive to predictive models, researchers require longitudinal data that connects initial emergency imaging to the final clinical outcome.
Methods and Data Curation
The ISLES’24 dataset was curated from two major stroke centers (Munich and Zurich) using a rigorous multi-step pipeline:
- Imaging Triad: Each case includes (sub-)acute CT imaging (Non-contrast CT, Angiography, and Perfusion) and follow-up MRI (DWI and ADC) taken 2–9 days later.
- Expert Annotation: Neuroradiologists manually marked vessel occlusions in CTA scans. For follow-up MRI, a hybrid approach was used: an AI model generated initial lesion masks, which were then refined and verified by human experts. Validation tests showed high agreement between expert raters, with Dice scores ranging from 0.86 to 0.90.
- Standardization: Images were converted to NIfTI format and organized according to BIDS conventions. Pre-processing included head-motion correction, co-registration of CT and MRI scans, and "defacing" to protect patient privacy.
- Anatomic Context: The dataset includes "silver ground-truth" segmentations of the Circle of Willis (CoW), providing structural context for the arterial network.
Dataset Characteristics
The ISLES’24 dataset consists of 245 cases, divided into a training set of 149 cases and a hidden test set of 96 cases for benchmarking. Unlike previous benchmarks like ISLES’22 (MRI-focused) or ATLAS (T1-weighted), ISLES’24 is specifically designed for the CT-to-MRI prediction task, mirroring the real-world clinical workflow where CT is the primary emergency diagnostic tool.
The cohort is well-balanced, with a mean age of approximately 71 years and a near 50/50 gender split. The inclusion of data from different scanner manufacturers (Siemens, Philips) and field strengths (1.5T and 3T) ensures that models trained on this dataset are more likely to generalize across different clinical environments.
Clinical Parameters and Longitudinal Outcomes
A primary strength of ISLES’24 is the inclusion of detailed clinical parameters and long-term outcomes. By providing 3-month clinical outcomes via the modified Rankin Scale (mRS), the dataset enables researchers to train models that predict functional recovery rather than just imaging changes.
As shown in Table 2, the dataset captures critical time-to-treatment metrics (e.g., onset-to-door, door-to-imaging) and baseline clinical scores (NIHSS, premorbid mRS). The outcome parameters, including mRS at discharge and at three months, provide the necessary "ground truth" for prognostic modeling.
Key Insights and Heterogeneity
- Blood Flow Dynamics: By including raw CT Perfusion (CTP) data and derived maps (Tmax, CBF), the dataset allows for the analysis of hemodynamics rather than just static anatomy.
- Diverse Stroke Patterns: The dataset intentionally includes diverse stroke patterns—from massive territorial infarcts to scattered micro-occlusions (33% of cases). This variety is crucial for developing robust AI.
- Vessel-Level Context: Masks for the exact point of vessel occlusion allow AI to learn the relationship between specific arterial blockages and resulting tissue damage patterns.
Limitations and Future Work
While substantial, the dataset has limitations. Privacy-preserving cropping of CTA scans may miss occlusions in the lower neck. Additionally, the Circle of Willis segmentations are "silver" standard (model-generated) rather than manually traced. Future work involves using this dataset as the foundation for the 2024 ISLES Challenge, pushing the community toward accurate post-treatment infarct prediction using only pre-treatment CT data.
Summary updated