ISLES 2024: The first longitudinal multimodal multi-center real-world dataset in (sub-)acute stroke

Evamaria Olga Riedel, Ezequiel de la Rosa, The Anh Baran, Moritz Roman Hernandez Petzsche, Hakim Baazaoui, Kaiyuan Yang, David Robben, Joaquin Oscar Seia, Roland Wiest, Mauricio Reyes, Ruisheng Su, Claus Zimmer, Tobias Boeckh-Behrens, Maria Berndt, Bjoern H. Menze, Benedikt Wiestler, Susanne Wegener, Jan S. Kirschke

Abstract

Stroke remains a leading cause of global morbidity and mortality, placing a heavy socioeconomic burden. Over the past decade, advances in endovascular reperfusion therapy and the use of CT and MRI imaging for treatment guidance have significantly improved patient outcomes and are now standard in clinical practice. To develop machine learning algorithms that can extract meaningful and reproducible models of brain function for both clinical and research purposes from stroke images - particularly for lesion identification, brain health quantification, and prognosis - large, diverse, and well-annotated public datasets are essential. While only a few datasets with (sub-)acute stroke data were previously available, several large, high-quality datasets have recently been made publicly accessible. However, these existing datasets include only MRI data. In contrast, our dataset is the first to offer comprehensive longitudinal stroke data, including acute CT imaging with angiography and perfusion, follow-up MRI at 2-9 days, as well as acute and longitudinal clinical data up to a three-month outcome. The dataset includes a training dataset of n = 150 and a test dataset of n = 100 scans. Training data is publicly available, while test data will be used exclusively for model validation. We are making this dataset available as part of the 2024 edition of the Ischemic Stroke Lesion Segmentation (ISLES) challenge (https://www.isles-challenge.org/), which continuously aims to establish benchmark methods for acute and sub-acute ischemic stroke lesion segmentation, aiding in creating open stroke imaging datasets and evaluating cutting-edge image processing algorithms.

Introduction: The Challenge of Stroke Evolution

In the high-stakes world of emergency neurology, "time is brain." When a patient suffers an ischemic stroke—a blockage of blood flow to the brain—doctors must quickly decide on a treatment path, such as endovascular reperfusion therapy. These decisions rely on estimating the "core" (irreversibly damaged tissue) versus the "penumbra" (salvageable tissue).

The challenge lies in the dynamic nature of stroke lesions, which expand over time based on factors like collateral circulation and clot location. Current software for estimating these volumes is often inconsistent across vendors. While several high-quality stroke image datasets exist, they are almost exclusively "snapshots" taken at a single point in time. Without longitudinal data—seeing both the initial state and the final outcome—it is incredibly difficult to train AI models to predict infarct growth or long-term patient prognosis.

Methods and Standardization

The researchers curated a dataset from two major stroke centers (Munich and Zurich) using a variety of CT and MRI scanners to ensure the data is robust and generalizable. Key technical steps included:

  • Standardization: All images were converted to the NIfTI format and organized according to BIDS (Brain Imaging Data Structure) conventions.
  • Pre-processing: The team performed image co-registration, temporal resampling of 4D perfusion data, and automated skull-stripping.
  • Vessel Occlusion Masks: Neuroradiologists manually delineated the exact point of vessel blockage in CTA scans—a critical feature for predicting how a stroke might spread.
  • Hybrid Ground Truths: To create "gold standard" labels for the final stroke lesions, the authors used a deep-learning ensemble model followed by rigorous manual inspection and correction by experienced neuroradiologists.

The ISLES’24 Dataset

The ISLES’24 dataset consists of 245 cases divided into a publicly available training set (n=149) and a hidden test set (n=96) reserved for fair benchmarking in the ISLES’24 challenge. Unlike previous benchmarks like ISLES’22 (which focused on MRI) or the ATLAS dataset (which focuses on T1-weighted images), ISLES’24 provides a comprehensive "triad" of acute CT imaging (NCCT, CTA, CTP) paired with follow-up MRI (DWI, ADC, FLAIR) and longitudinal clinical metadata.

Patient demographics and distribution across training and test sets.

Key Insights: Clinical Context and Longitudinal Pairing

The dataset's value extends beyond pixels by incorporating "real-world" clinical context:

  • The Power of Longitudinal Pairing: By linking acute CT (initial presentation) with follow-up MRI (final damage), the dataset allows models to learn the mapping of stroke evolution.
  • Multimodal Clinical Data: The inclusion of onset-to-door times, medication history, and 3-month outcomes (modified Rankin Scale) allows for models that combine imaging with patient history.
  • Vessel-Specific Analysis: Including masks for the point of occlusion and the Circle of Willis allows researchers to study how vascular anatomy influences stroke patterns.
Comprehensive clinical parameters, laboratory values, and treatment timelines.

Results and Validation

The authors validated the quality of their dataset through inter-rater agreement studies. When comparing their ground truth segmentations against external expert raters, they achieved a Dice similarity coefficient of 0.86 to 0.90. This high level of agreement indicates that the labels provided are highly reliable for training machine learning models. The dataset reflects a realistic clinical population, with a mean patient age of approximately 71 years and a diverse range of stroke patterns.

Limitations and Future Work

The authors acknowledge a few constraints:

  • Cropping Issues: CTA scans were cropped to the head for anonymization, potentially missing occlusions lower in the neck.
  • Privacy Noise: Laboratory values and intervention times were randomly altered by ±5% to protect patient identity.
  • Selection Bias: The study specifically included patients who underwent successful recanalization (mTICI 2c or 3). Future datasets might need to include cases where treatment failed to provide a more complete picture of stroke pathology.

The ISLES’24 dataset is currently hosted on the Grand Challenge platform, serving as a foundation for the next generation of stroke-prediction algorithms.

Summary updated

Abstract

Stroke remains a leading cause of global morbidity and mortality, placing a heavy socioeconomic burden. Over the past decade, advances in endovascular reperfusion therapy and the use of CT and MRI imaging for treatment guidance have significantly improved patient outcomes and are now standard in clinical practice. To develop machine learning algorithms that can extract meaningful and reproducible models of brain function for both clinical and research purposes from stroke images - particularly for lesion identification, brain health quantification, and prognosis - large, diverse, and well-annotated public datasets are essential. While only a few datasets with (sub-)acute stroke data were previously available, several large, high-quality datasets have recently been made publicly accessible. However, these existing datasets include only MRI data. In contrast, our dataset is the first to offer comprehensive longitudinal stroke data, including acute CT imaging with angiography and perfusion, follow-up MRI at 2-9 days, as well as acute and longitudinal clinical data up to a three-month outcome. The dataset includes a training dataset of n = 150 and a test dataset of n = 100 scans. Training data is publicly available, while test data will be used exclusively for model validation. We are making this dataset available as part of the 2024 edition of the Ischemic Stroke Lesion Segmentation (ISLES) challenge (https://www.isles-challenge.org/), which continuously aims to establish benchmark methods for acute and sub-acute ischemic stroke lesion segmentation, aiding in creating open stroke imaging datasets and evaluating cutting-edge image processing algorithms.
BibTeX
@misc{riedel2024isles,
  author = {Evamaria Olga Riedel and Ezequiel de la Rosa and The Anh Baran and Moritz Roman Hernandez Petzsche and Hakim Baazaoui and Kaiyuan Yang and David Robben and Joaquin Oscar Seia and Roland Wiest and Mauricio Reyes and Ruisheng Su and Claus Zimmer and Tobias Boeckh-Behrens and Maria Berndt and Bjoern H. Menze and Benedikt Wiestler and Susanne Wegener and Jan S. Kirschke},
  title = {{ISLES 2024: The first longitudinal multimodal multi-center real-world dataset in (sub-)acute stroke}},
  year = {2024},
  eprint = {2408.11142},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.11142},
}
Figure