Robust slice-level stroke classification in non-contrast head CT via structural consistency regularization and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Stroke is a time-critical neurological emergency requiring rapid diagnosis.
• Non-contrast head CT is crucial for initial stroke assessment but faces interpretation challenges due to low contrast, blurred boundaries, and artifacts in single slices.

Purpose of the Study:

• To develop an advanced slice-level stroke CT classification framework.
• To enhance the reliability of stroke detection in challenging single-slice CT images.

Main Methods:

• Proposed a framework integrating structural consistency regularization and counterfactual suppression.
• Employed a multi-branch feature refinement strategy for improved evidence analysis.
• Ensured representation stability and attenuated spurious cues for robust classification.

Main Results:

• Achieved superior performance across Accuracy, Precision, Recall, F1, and AUC on two stroke CT datasets.
• Demonstrated high accuracy (e.g., 0.9823 on the primary dataset) and AUC (e.g., 0.9982).

Conclusions:

• The proposed method reliably identifies stroke-discriminative evidence in low-contrast, high-interference single-slice CT scans.
• Offers robust support for emergency stroke triage, preliminary diagnosis, and neurorehabilitation feedback.