FeaCL: Carotid plaque classification from ultrasound images using feature-level and instance-level contrast learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease Research

Background:

• Accurate classification of carotid plaques from ultrasound images is vital for predicting cardiovascular and cerebrovascular disease risks.
• Deep learning models show promise but are hindered by limited labeled carotid plaque image datasets.

Purpose of the Study:

• To introduce a novel self-supervised learning technique, FeaCL (FEature-level and instAnce-level contrast learning), to improve carotid plaque classification accuracy.
• To address the challenge of limited labeled data in training deep learning models for carotid plaque analysis.

Main Methods:

• FeaCL employs a triplet network with strong and weak augmentation in a pretext task to learn robust carotid plaque representations.
• The method promotes feature and instance similarity across different augmented views of the same image.
• The pre-trained encoder from the pretext task is fine-tuned on labeled ultrasound images for the downstream classification task.

Main Results:

• FeaCL achieved 83.4% classification accuracy using only 30% of the training data.
• This represents a significant improvement of 16.3% compared to models trained without the self-supervised pretext task.
• The method demonstrates effective learning of carotid plaque features even with limited labeled data.

Conclusions:

• FeaCL significantly enhances carotid plaque classification accuracy from ultrasound images, particularly in low-data regimes.
• The self-supervised approach effectively learns meaningful representations, improving diagnostic capabilities for clinicians.
• This technique offers a valuable tool for risk stratification and treatment planning in patients with carotid artery disease.