An adaptively weighted ensemble of multiple CNNs for carotid ultrasound image segmentation

Abstract

Carotid atherosclerotic plaques cause stroke when plaques rupture and clog the blood vessels that deliver blood to brain. Ultrasound measurements (i.e. total-plaque-area and intima-media-thickness) are mainly used to monitor the progression and regression of plaques. Recently, deep learning has provided powerful tools for ultrasound carotid image segmentation. However, the performances of deep learning models vary on different network architectures. In this paper, we report on the development of an adaptively weighted ensemble of multiple convolutional neural networks (CNNs) for carotid ultrasound image segmentation, aiming at combining the advantages of different CNN models to achieve higher accuracy and better generalization performance. During the joint training of the ensemble networks, the model weights and sample weights were combined to improve the segmentation performance. This adaptively weighted ensemble algorithm was applied to three UNet++ models with different backbones (ResNet152, DenseNet169 and VGG19), and evaluated on 510 carotid ultrasound images from 144 subjects who were followed in the Stroke Prevention and Atherosclerosis Research Centre (SPARC, London, Canada). The experimental results show that our method increases the segmentation accuracy, and reduces the distance errors as compared to using a single classifier, three ensemble algorithms (average weighting, majority voting and SegNet-UNet+) and a published carotid segmentation algorithm. With high accuracy and low variance, the proposed adaptively weighted ensemble model could be used to measure carotid plaques in clinical practice and clinical trials.