Ultrasound Datasets — Sonography & Doppler Data

Ultrasound (sonography) datasets capture real-time, non-ionizing imaging produced from high-frequency sound waves, and they are central to clinical AI because ultrasound is portable, inexpensive, and ubiquitous across radiology, obstetrics, emergency medicine, and the point of care. This category covers general body ultrasound rather than cardiac imaging, which lives under the separate Echocardiogram category; here the scope spans abdominal scans of the liver, kidneys, gallbladder, pancreas, and spleen; obstetric and fetal studies including biometry and second-trimester anomaly screening; gynecologic pelvic imaging; vascular and Doppler studies such as carotid stenosis assessment and lower-limb deep-vein-thrombosis compression exams; thyroid nodule characterization; breast lesion evaluation; musculoskeletal scanning of tendons, joints, and soft tissue; and point-of-care ultrasound (POCUS) protocols like FAST and lung assessment. Datasets are delivered as DICOM cine loops and still frames, with embedded metadata recording transducer type, preset, frequency, depth, gain, frame rate, and machine and probe identifiers, and they include several acquisition modes: grayscale B-mode, color and power Doppler, spectral pulsed-wave Doppler tracings, and shear-wave or strain elastography.

Clinically valuable cohorts carry expert annotations and structured findings: view and plane labels, frame-level lesion segmentation, organ and structure boundaries, fetal biometric measurements (biparietal diameter, head and abdominal circumference, femur length), Doppler velocity and resistive-index measurements, and standardized reporting schemas such as TI-RADS for thyroid nodules and the ultrasound BI-RADS lexicon for breast masses. Because ultrasound is strongly operator-dependent, image quality and view correctness vary widely, so robust datasets document acquisition quality, view and probe labels, and segmentation masks that enable automated view classification, quality scoring, and lesion localization. High-quality cohorts are demographically diverse, span multiple ultrasound vendors, transducers, and presets so models generalize beyond a single machine, and are rigorously de-identified to strip PHI from DICOM headers and, critically, to redact burned-in pixel text such as patient names, dates, and institution banners that ultrasound systems overlay on the image itself.

On GetDATA, researchers and medtech companies post ultrasound requests specifying anatomy and exam type, required modes (B-mode, Doppler, elastography), annotation type (view label, segmentation, or measurement), label taxonomy, scanner and probe diversity, and minimum case counts, and verified hospitals and labs fulfill them with compliant, quality-scored sonography data in DICOM. Such datasets also support pipelines for real-time guidance, standard-plane detection, automated biometry, and triage at the bedside, where consistent labeling and external validation across institutions are essential. Browse the open ultrasound requests below, or explore related imaging categories.