Better Diffusion Segmentation in Acute Ischemic Stroke Through Automatic Tree Learning Anomaly Segmentation

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Stroke is a leading cause of death and disability worldwide, with ischemic strokes accounting for the majority.
• Prompt assessment of irreversible brain tissue damage (core lesion) is crucial for effective acute stroke management.
• Magnetic Resonance Imaging (MRI), specifically diffusion-weighted imaging, is vital for quantifying lesion volume.

Purpose of the Study:

• To introduce ATLAS, a fully automated machine learning algorithm for rapid and standardized delineation of the ischemic stroke core lesion.
• To compare the performance of ATLAS against established threshold-based methods and the state-of-the-art COMBAT Stroke algorithm.
• To evaluate ATLAS's accuracy in quantifying lesion volume for improved patient triaging.

Main Methods:

• ATLAS, a machine learning algorithm using decision trees and spatial regularization, was developed and trained on expert delineations.
• The algorithm was tested on acute anterior circulation stroke patient data from the I-Know multicenter study using leave-one-out cross-validation.
• Performance was quantified using the Dice index for lesion overlap and standard deviation of residuals for volume accuracy.

Main Results:

• ATLAS achieved a significantly higher median Dice coefficient (0.6122) compared to COMBAT Stroke (0.5636) and threshold-based methods (0.3951-0.2839).
• The ATLAS algorithm demonstrated superior accuracy in volume quantification, with a lower standard deviation of residuals (10.25 ml) versus COMBAT Stroke (17.53 ml).
• Leave-one-out cross-validation confirmed robust performance on independent patient data.

Conclusions:

• ATLAS provides a highly accurate and automated solution for delineating ischemic stroke core lesions.
• The algorithm's superior performance in lesion segmentation and volume quantification can enhance clinical decision-making.
• ATLAS has the potential to optimize patient triaging for timely and appropriate acute stroke therapies.