A balanced sample generation method for semi-supervised carotid ultrasound image classification via pseudo-label distribution and model confidence

Abstract

Abstract The rupture of carotid plaques dramatically increases the risk of stroke, making it essential to classify these plaques to promptly identify high-risk cases and assist in patient management. Deep learning-based plaque classification methods need large labeled datasets, which are often challenging to acquire. Semi-supervised learning offers a solution by enabling deep learning networks to perform well with fewer labeled images. However, this method filters pseudo-labels based on thresholds, resulting in imbalanced quantities and cognitive bias in the network. To tackle this issue, we present a Balanced Sample Generation based Pseudo-label algorithm (BSGP) for semi-supervised carotid ultrasound image classification. BSGP generates new samples based on pseudo-label distribution and model confidence in each category, reducing cognitive bias from imbalanced pseudo-label quantities. Evaluated on 1,270 ultrasound carotid plaque images, our BSGP outperformed other semi-supervised methods (FixMatch, FlexMatch, AdaMatch, FreeMatch) when using 10%, 30%, and 50% labeled data. This illustrates the efficacy of our method in leveraging unlabeled images for model training, potentially advancing the use of deep learning in ultrasound-based plaque classification and supporting clinical diagnoses.