A hybrid deep learning paradigm for carotid plaque tissue characterization and its validation in multicenter cohorts...

Area of Science:

• Cardiology
• Neurology
• Artificial Intelligence

Background:

• Stroke is a leading cause of death, necessitating early detection of carotid plaques.
• Artificial intelligence (AI) offers automated solutions for plaque characterization.
• Current solo deep learning (SDL) models do not leverage AI's hybrid capabilities.

Purpose of the Study:

• To explore hybrid deep learning (HDL) models for improved carotid plaque detection.
• To compare the performance of HDL models against SDL and transfer learning (TL) techniques.
• To introduce the Atheromatic™ 2.0_HDL system for multicenter plaque characterization.

Main Methods:

• Developed two types of HDL frameworks: SDL fusion and SDL with machine learning (ML) tandem models.
• Trained 11 HDL models using an augmentation framework and three loss functions.
• Benchmarked 11 HDL models against one SDL and five TL models, totaling 17 AI models.

Main Results:

• The best performing HDL model (CNN and decision tree) achieved 99.78% accuracy and 0.99 AUC.
• HDL models outperformed SDL and TL models by significant margins in accuracy, AUC, DOR, and Kappa statistics.
• The Atheromatic™ 2.0_HDL system provided results in under 2 seconds.

Conclusions:

• Hybrid deep learning (HDL) is a fast, reliable, and effective tool for carotid plaque characterization.
• HDL enables early stroke risk stratification.
• This study pioneers the use of HDL in a multicenter framework for plaque analysis.