APIS: a paired CT-MRI dataset for ischemic stroke segmentation - methods and challenges

Area of Science:

• Neurology
• Radiology
• Medical Imaging
• Computational Science

Background:

• Stroke is a leading global cause of death, with ischemic stroke requiring rapid diagnosis for effective patient management.
• Non-contrast CT (NCCT) is standard for initial stroke evaluation but has low sensitivity for early ischemic changes.
• Diffusion-weighted MRI (DWI) offers better detection but faces accessibility and cost limitations.

Purpose of the Study:

• To develop and evaluate computational strategies for delineating ischemic stroke lesions on CT scans using paired apparent diffusion coefficient (ADC) information.
• To create the first paired dataset of NCCT and ADC imaging for acute ischemic stroke patients.
• To accelerate stroke patient management through enhanced CT-based analysis.

Main Methods:

• A public challenge was organized for scientists to apply computational strategies to segment stroke lesions on NCCT scans, using paired ADC data.
• A novel dataset combining NCCT and ADC studies of acute ischemic stroke patients was compiled.
• Submitted algorithms were validated against expert radiologist segmentations.

Main Results:

• The best algorithm achieved a Dice score of 0.2 on a test set of 36 patient studies.
• Despite using advanced deep learning tools, computational approaches demonstrated limitations in segmenting small lesions with heterogeneous density.
• This work represents the first effort to create a paired NCCT and ADC dataset for acute ischemic stroke.

Conclusions:

• Integrating ADC information into CT analysis shows promise for improving stroke lesion characterization.
• Current computational methods struggle with the accurate segmentation of subtle and heterogeneous ischemic lesions on CT.
• Further advancements are needed to overcome the limitations of automated stroke lesion segmentation, particularly for early ischemic changes.