An open B-mode ultrasound database for deep learning-based atherosclerotic plaque segmentation

Area of Science:

• Medical imaging
• Cardiovascular disease research
• Artificial intelligence in healthcare

Background:

• Cardiovascular events cause significant global mortality, with atherosclerosis being a primary driver.
• Accurate detection of atherosclerotic plaques in medical images is crucial for early diagnosis and management.
• Manual plaque annotation in ultrasound images is time-consuming and prone to variability.

Purpose of the Study:

• To introduce a novel, open-access B-mode ultrasound image database for atherosclerotic plaque segmentation.
• To address the scarcity of annotated datasets for developing automated segmentation algorithms.
• To facilitate standardized and reproducible plaque analysis.

Main Methods:

• A dataset of 541 B-mode ultrasound images (800x800 pixels) with expert-validated binary segmentation masks was created.
• The dataset encompasses diverse imaging conditions, including plaque-free, single-plaque, and multiple-plaque scenarios.
• A U-Net ensemble model was trained and evaluated using the developed dataset.

Main Results:

• The U-Net ensemble achieved a median error of 0.35 mm² on plaque-free images.
• A mean Dice coefficient of 0.62 was obtained on images containing atherosclerotic plaques.
• The dataset demonstrated suitability for training and evaluating plaque segmentation algorithms.

Conclusions:

• The presented open-access database is valuable for developing and benchmarking automated atherosclerotic plaque segmentation tools.
• This resource can improve the standardization and reproducibility of plaque detection in B-mode ultrasound.
• The dataset supports advancements in early cardiovascular disease diagnosis and risk stratification.