An open B-mode ultrasound database for deep learning-based atherosclerotic plaque segmentation.
Authors
Affiliations (5)
Affiliations (5)
- Facultad de Ciencias Exactas, Físicas y Naturales, Universidad Nacional de Córdoba, Córdoba, Argentina. [email protected].
- Facultad de Ciencias de la Salud, Universidad Católica de Córdoba, Libertad 1255, X5004, Córdoba, Argentina.
- Facultad de Ciencias Exactas, Físicas y Naturales, Universidad Nacional de Córdoba, Córdoba, Argentina.
- Posgrado en Robótica e Inteligencia Artificial, Universidad Tecnológica del Uruguay, Rivera, Uruguay.
- Instituto de Investigaciones en Ciencias de la Salud, CONICET, Córdoba, Argentina.
Abstract
Cardiovascular events, predominantly ischemic, account for approximately 32% of global mortality and are expected to increase approximately 30% by 2030. A substantial proportion of these events is preventable through the control of established risk factors. Atherosclerosis is defined by the progressive development of arterial plaques and remains the main cause of ischemic cardiovascular disease. In this context, plaque detection in medical images is required for early diagnosis and longitudinal assessment, which is commonly based on total plaque area. Manual plaque annotation in B-mode ultrasound images requires specialized expertise and is affected by inter- and intra-operator variability. Automatic methods trained on expert-annotated data are therefore required to improve standardization and reproducibility. However, the research community faces a limited availability of open access databases designed for image segmentation and benchmarking under traditional and heterogeneous imaging conditions inherent to B mode ultrasound data. This work presents an open-access B-mode ultrasound image database designed for atherosclerotic plaque segmentation. The dataset includes 541 ultrasound images with a resolution of 800 × 800 pixels, each paired with a binary segmentation mask validated by clinical specialists. The database captures heterogeneous imaging conditions, including plaque-free cases, single and multiple plaques, and non-centralized plaque locations. To assess its suitability for algorithm development, the dataset was used to train a U-Net ensemble. The evaluation provided a median error of 0.35 mm<sup>2</sup> on plaque free test images and a mean Dice coefficient of 0.62 on test images containing plaques.