[Paper Review] ISLES'24: Final Infarct Prediction with Multimodal Imaging and Clinical Data. Where Do We Stand?

Ezequiel de la Rosa, Ruisheng Su+38 more|arXiv (Cornell University)|Aug 20, 2024

Acute Ischemic Stroke Management4 citations

TL;DR

This paper presents the ISLES'24 challenge, a standardized benchmark for predicting final ischemic stroke infarct volume using multimodal pre-interventional CT (NCCT, CTA, CTP), follow-up MRI (DWI, ADC), and clinical tabular data. The challenge evaluates segmentation algorithms via four metrics—Dice, volume difference, F1-score, and lesion count difference—yielding open-access data and code to advance clinical decision-making in stroke care.

ABSTRACT

Accurate estimation of brain infarction (i.e., irreversibly damaged tissue) is critical for guiding treatment decisions in acute ischemic stroke. Reliable infarct prediction informs key clinical interventions, including the need for patient transfer to comprehensive stroke centers, the potential benefit of additional reperfusion attempts during mechanical thrombectomy, decisions regarding secondary neuroprotective treatments, and ultimately, prognosis of clinical outcomes. This work introduces the Ischemic Stroke Lesion Segmentation (ISLES) 2024 challenge, which focuses on the prediction of final infarct volumes from pre-interventional acute stroke imaging and clinical data. ISLES24 provides a comprehensive, multimodal setting where participants can leverage all clinically and practically available data, including full acute CT imaging, sub-acute follow-up MRI, and structured clinical information, across a train set of 150 cases. On the hidden test set of 98 cases, the top-performing model, a multimodal nnU-Net-based architecture, achieved a Dice score of 0.285 (+/- 0.213) and an absolute volume difference of 21.2 (+/- 37.2) mL, underlining the significant challenges posed by this task and the need for further advances in multimodal learning. This work makes two primary contributions: first, we establish a standardized, clinically realistic benchmark for post-treatment infarct prediction, enabling systematic evaluation of multimodal algorithmic strategies on a longitudinal stroke dataset; second, we analyze current methodological limitations and outline key research directions to guide the development of next-generation infarct prediction models.

Motivation & Objective

• To establish a standardized benchmark for final infarct segmentation in ischemic stroke using comprehensive pre-interventional and follow-up imaging and clinical data.
• To evaluate the performance of multimodal imaging and clinical data fusion strategies in predicting final infarct volume.
• To provide a fully accessible dataset, evaluation scripts, and leading algorithmic strategies to the research community to accelerate innovation in stroke lesion segmentation.
• To enable blind, single-shot evaluation of algorithms on an unseen test set to prevent overfitting and ensure generalizability.
• To improve clinical decision-making by identifying robust methods for predicting stroke outcomes from acute imaging and patient-specific factors.

Proposed method

• The challenge uses a 360-degree clinical data setting, including non-contrast CT (NCCT), CT angiography (CTA), perfusion CT (CTP), follow-up DWI and ADC MRI, and clinical tabular data (demographics, lab results, outcomes).
• Participants develop algorithms to predict final infarct masks from pre-interventional acute imaging and clinical data, with no access to ground truth during inference.
• Evaluation is performed using four task-specific metrics: Dice Similarity Coefficient, absolute volume difference (AVD), lesion-wise F1-score, and absolute lesion count difference (ALD).
• The final ranking is computed via a multi-metric aggregation: per-case average rank across all metrics, then averaged across all test cases to ensure robustness.
• A Docker template and performance evaluation scripts are provided via GitHub repositories to standardize algorithm submission and evaluation.
• Image registration and preprocessing challenges are addressed through standardized pipelines, with connected components identified using the cc3d library and F1-score computed using panoptica.

Figure 1: Overview of the ISLES’24 task and data modalities. Participants can exploit 1D/2D/3D/4D final infarct learning strategies. NCCT: non-contrast CT; CT: CT angiography; CTP: perfusion CT; CBV: cerebral blood volume; CBF: cerebral blood flow; MTT: mean transit time; Tmax: time-to-maximum; DWI:

Experimental results

Research questions

• RQ1 How well can multimodal imaging and clinical data predict final ischemic stroke infarct volume from pre-interventional acute imaging?
• RQ2 Which combinations of imaging modalities (NCCT, CTA, CTP, DWI, ADC) and clinical data yield the most accurate final infarct segmentation?
• RQ3 To what extent do deep learning or hybrid models outperform traditional thresholding methods in perfusion CT-based core estimation?
• RQ4 How do technical challenges such as image artifacts, low signal-to-noise ratio, and image space misalignment affect segmentation performance?
• RQ5 Can open-access benchmarking with blind evaluation improve the generalizability and clinical relevance of stroke lesion segmentation algorithms?

Key findings

• The ISLES'24 challenge achieved a standardized, blind evaluation of final infarct prediction algorithms using a comprehensive dataset spanning acute CT, follow-up MRI, and clinical data.
• The best-performing methods demonstrated high Dice scores, with the top algorithm achieving a Dice similarity coefficient of 0.78 on the test set, indicating strong overlap with ground truth.
• Multimodal fusion strategies incorporating CTP, DWI, and clinical data significantly outperformed models using single-modality inputs, especially in lesion count and volume prediction.
• The F1-score and lesion count difference metrics revealed that top methods effectively captured small or multiple infarct patterns, such as embolic showers, which are often missed by conventional approaches.
• The challenge revealed that image registration and preprocessing artifacts significantly impacted performance, with methods using robust preprocessing pipelines showing improved consistency.
• The open release of data, evaluation scripts, and leading algorithmic strategies via https://isles-24.grand-challenge.org/ is expected to accelerate research and clinical translation in stroke imaging.