AGTS-Net: Anatomically Guided Two-Stage Network for Actionable Stroke Classification in Complete NCCT Studies

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Emergency stroke assessment using non-contrast computed tomography (NCCT) is challenging due to subtle early ischemic signs and anatomical variations.
• Accurate and rapid stroke classification is critical for timely treatment and improved patient outcomes.

Purpose of the Study:

• To develop and evaluate AGTS-Net, an explainable deep learning framework for stroke classification from NCCT scans.
• To enhance the accuracy and interpretability of AI models for emergency stroke diagnosis.

Main Methods:

• Developed AGTS-Net, a two-stage deep learning network that first routes NCCT slices to anatomical regions and then applies region-specific classifiers.
• Trained and internally validated the model on a proprietary cohort of 3440 NCCT slices from 99 patients.
• Evaluated anatomical pre-classifier performance on a public Kaggle dataset.
• Utilized Grad-CAM for visualizing model predictions.

Main Results:

• Achieved 0.99 accuracy in anatomical routing.
• Demonstrated improved classification performance compared to global models through anatomical regionalization.
• Transfer learning experiments confirmed the benefit of the two-stage approach.
• External cross-domain evaluation validated the anatomical pre-classifier.

Conclusions:

• AGTS-Net provides an effective, explainable deep learning approach for stroke classification using NCCT.
• The framework's anatomical guidance enhances diagnostic accuracy and supports clinical decision-making in urgent stroke assessment.
• AGTS-Net serves as a valuable decision-support tool, offering actionable insights and visual guidance.