Cross-Domain Carotid Artery Segmentation Using Folding Fan ResNet and Quadratic Mapping Loss

Vessel-wall volume and localized three-dimensional ultrasound (3DUS) metrics are sensitive indicators of carotid atherosclerosis changes in response to therapies. Deep-learning segmentation models have been proposed to segment the media-adventitia boundary (MAB) and lumen-intima boundary (LIB) necessary to obtain these measurements. However, the performance of these models declines substantially when applied to datasets with different properties from their training data, highlighting a need for more generalizable models. We introduce Folding Fan ResNet (FFRN), a deep-learning model designed to enhance cross-domain generalizability. FFRN incorporates a novel skip connection pathway, the Folding Fan Connection (FFC), which improves encoder-decoder feature fusion through multiple addition and convolutional operations. This design reduces parameter redundancy while maintaining efficient feature propagation, leading to increased generalizability compared to existing models. Additionally, the quadratic mapping loss (QML) function is proposed to improve model fine-tuning. QML addresses arterial mislocalization by deferring the optimization of incorrectly localized boundaries while prioritizing adjustments to correctly segmented regions. This strategy prevents the reinforcement of erroneous features and allows for better modeling of carotid structures. FFRN demonstrated higher generalizability than four state-of-the-art segmentation models, improving Dice similarity coefficients (DSCs) of 78.9±24.1% and 79.0±23.2% attained by the U-Net for LIB and MAB segmentation, respectively, to 84.0±18.2% and 84.6±17.1% in MAB and LIB segmentation, respectively. With QML fine-tuning, the DSCs increased to 88.9±8.2% (LIB) and 90.6±8.1% (MAB), outperforming all competing approaches, even after their performances were enhanced by QML. Notably, the fine-tuned model is the first to achieve a zero incorrect localization rate (ILR) in the target dataset. The improved generalizability afforded by the model enhances efficiency for carotid disease monitoring across clinical centers.