CASCADE-FSL: Few-shot learning for collateral evaluation in ischemic stroke.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Assessing collateral circulation is crucial for ischemic stroke treatment planning.
• Poor collaterals pose risks like bleeding and death, necessitating accurate identification.
• Current computer-aided methods struggle with limited and imbalanced stroke datasets.

Purpose of the Study:

• To develop a novel computer-aided method for effectively distinguishing poor collateral circulation in ischemic stroke patients.
• To address the challenge of scarce and imbalanced datasets in deep learning for stroke imaging.
• To reduce inter- and intra-rater variability in radiologist assessments.

Main Methods:

• Proposed CASCADE-FSL, a few-shot learning approach using a 2D ResNet-50 backbone.
• Trained the model on a small, unbalanced dataset, treating good and intermediate collaterals as normal classes.
• Identified poor collaterals as anomalies compared to the established normal classes.

Main Results:

• Achieved an overall accuracy of 0.88.
• Demonstrated high sensitivity (0.88) and specificity (0.89) in identifying poor collaterals.
• Effectively addressed the imbalanced dataset challenge.

Conclusions:

• CASCADE-FSL is effective in accurately identifying poor collateral circulation.
• The few-shot learning approach shows promise for stroke imaging with limited data.
• This method can aid in optimizing treatment strategies for ischemic stroke patients.