Tri-correcting: Label noise correction via triple CNN ensemble for carotid plaque ultrasound image classification

• Published in: Biomedical signal processing and control 2024-05, Vol.91, p.105981, Article 105981
• Main Authors: Zhou, Ran, Gan, Weiyan, Wang, Furong, Yang, Zhi, Huang, Zhongwei, Gan, Haitao
• Format: Article
• Language: English
• Subjects: Carotid plaque; Ensemble classification; Label correction; Label noise; Tri-correcting
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2024.105981

In clinical practice, classifying carotid artery plaque is vital for assessing coronary artery disease (CAD) risk. Deep supervised learning has achieved significant success in carotid artery plaque ultrasound image classification, relying on large-scale accurately annotated datasets. However, label noise can significantly degrade model performance due to the professional knowledge and experience of annotators during data collection. Previous studies tended to collaboratively train two models to alleviate the above degradation, but they did not consider how to determine the final hypothesis when the two models are inconsistent. Inspired by Tri-Training, this study proposes a label noise correction method by using a triple convolutional neural networks (CNN) ensemble to improve the robust and accuracy of carotid plaque classification. Firstly, the noise filtering strategy is used to obtain a support dataset and a noisy dataset from the original dataset. Then, three CNNs models are trained using the support dataset, and the label noise dataset is corrected using the joint voting strategy among the three CNNs. The CNN models are then retrained using the instance allocation strategy before finally integrating and classifying with the three classifier models. Evaluated on 1270 carotid ultrasound images from 844 subjects followed in Zhongnan Hospital (Wuhan, China), our method improves the classification accuracy by 4%-5%, 3%, and 2%-3% respectively in the case of 10%, 20%, and 30% label noise rates compared to other advanced methods, which may help clinical doctors diagnose carotid artery plaque categories and evaluate the risk of CAD for patients. •Tri-Correcting is proposed for carotid plaque classification with noisy labels.•Joint voting and instance allocation strategies are proposed to correct noisy labels.•Tri-correcting’s high performance makes it useful in assessing CAD patient risk.