Automatic assessment of DWI-ASPECTS for acute ischemic stroke based on deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Alberta Stroke Program Early Computed Tomography Score (ASPECTS) is a standard for assessing early ischemic changes in acute ischemic stroke.
• Current ASPECTS scoring is subjective, leading to inter-reader variability and impacting clinical decisions.
• There is a need for objective and reliable methods for early stroke assessment.

Purpose of the Study:

• To develop an automated ASPECTS scoring model utilizing diffusion-weighted imaging (DWI).
• To improve the accuracy and consistency of ASPECTS scoring in clinical practice.
• To aid clinicians in making timely and precise treatment plans for acute ischemic stroke patients.

Main Methods:

• A deep learning network, enhanced from U-net, was designed for segmenting brain regions relevant to ASPECTS scoring.
• Hybrid classifiers were employed: a grayscale comparison algorithm for larger brain regions and hybrid feature training for smaller regions.
• The model was trained and validated on data from 82 stroke patients.

Main Results:

• The automated segmentation achieved an average DICE coefficient of 0.864 for the hindbrain.
• The hybrid classifier demonstrated significant performance in both region-level and dichotomous ASPECTS scoring.
• The model achieved a sensitivity of 95.51% and accuracy of 93.43% on the test set.
• An intraclass correlation coefficient (ICC) of 0.87 was observed for dichotomous ASPECTS compared to expert readings.

Conclusions:

• An automated DWI-based ASPECTS scoring model was successfully developed for acute ischemic stroke.
• The proposed segmentation-then-classification approach proved effective and reliable.
• This automated system has the potential to significantly assist physicians in early stroke scoring and diagnosis.