Ar-PlaqSegm1: First Argentine database of B-mode ultrasound images for atherosclerotic plaque segmentation

This is an Argentinian B-mode ultrasound image database designed for the automatic segmentation of atherosclerotic plaques. This database is intended to promote reproducibility and collaborative research, particularly in the field of deep learning, by providing a set of images that represent real-world clinical scenarios.
The database consists of 541 pairs of ultrasound images (each measuring 800x800 pixels). Each image is accompanied by its corresponding binary segmentation mask, which has been validated by medical specialists. The pairs were split into a Test set (48 pairs, 9%) and a Training set (493 pairs, 91%)
The collection’s key feature is its diversity, which aims to overcome limitations found in other datasets:
Content: 201 pairs contain no visible plaques, while 340 pairs show the presence of one or more plaques.
Size and Location: The images have large dimensions and include cases with multiple plaques and plaques that are not centrally located.
Format: The final output of the preprocessing pipeline provides monochrome input images and binary output images (masks), both at 800x800 pixels.

Methods and Protocols
The study protocol was approved by the Institutional Independent Ethics Committee (IIEC) of the National University of Córdoba and the Rusculleda Foundation.
Images were collected by two physicians experienced in the manual delineation of plaques at the Centro Privado Blossom DMO in Córdoba, Argentina.
The study included a middle-aged to elderly cohort (57% male, 43% female) presenting with mild, moderate, or severe atherosclerosis within the context of both primary and secondary prevention. The clinical profile of the participants included hypertension in 35% of the cases (mean blood pressure ~124/72 mmHg), hypercholesterolemia in 27% (with overall dyslipidemia reaching up to 84.6%), and diabetes in 14% (mean HbA1c: 5.8–6.1%). Renal function varied, with estimated glomerular filtration rates (eGFR) ranging from 20 to 100. Furthermore, the cohort exhibited a Framingham risk score > 15%, indicative of subclinical atherosclerosis risk.
Inclusion criteria comprised ultrasound images exhibiting either the absence or presence of arterosclerotic plaques, where the latter was defined as having awhere the vascular wall thickness greater than 1 mm.
Ground truth for plaque boundaries was established through manual delineation by the expert physicians on the original images (using yellow dots and green stripes).

Instruments: Two B-mode ultrasound devices from Esaote and Vinno were used for image acquisition.
Initial Data Acquired:
26 image pairs in DICOM format (initial size 800x600 pixels).
515 image pairs in PNG format (initial size 1200x900 pixels).