Exploring the Impact of Learning Paradigms on Network Generalization: A Multi-Center IMT Study

The intima-media thickness (IMT) is an important parameter for evaluating cardiovascular disease risk and progression and can be extracted from B-mode longitudinal ultrasound images of the carotid artery. Despite its clinical significance, inter- and intra-operator variability in IMT measurement is a challenge due to subjective factors. Therefore, automatic and semi-automatic approaches based on heuristic methods and deep neural networks have been proposed to reduce the variability in IMT measurement. However, the inter- and intra- operator variability still remains an issue as it affects the quality and diversity of ground truth (GT) data used for training deep learning models. In this study, the authors evaluate the performance of different learning paradigms using different GTs on a multi-center IMT dataset. A recent segmentation network, ConvNeXt, is trained on a dataset of 2576 B-mode longitudinal ultrasound images of the carotid artery, using different GT annotations and learning paradigms. The method is then tested on an external dataset of 448 images from four different centers for which three manual segmentations were available. The results show how the use of different GT annotations and learning paradigms can enhance the generalization ability of deep learning models, demonstrating the importance of selecting appropriate GT data and learning strategies in achieving robust and reliable solutions. The study highlights the significance of incorporating heuristic methods in the training process of deep learning models to enhance the accuracy and consistency of IMT measurement, thus enabling more precise cardiovascular disease risk assessment.