Search a curated directory of ultrasound datasets for research, benchmarking, and tool development. Filter by modality, clinical application, or licence, then download the spreadsheet if you want to work offline.
Curated benchmark derived from TCIA Breast-Lesions-USG with lesion masks, BIRADS features, and radiologist annotations.
1,875 images, 722 benign, 342 malignant cases, biopsy-proven tumors, semantic segmentation with DeepLabV3+
A comprehensive collection of 37,736 graded cardiac ultrasound images from scans of a CAE Blue Phantom, covering different heart views and quality levels (A4C, SC, PL, PSAV, PSMV, Random). Images are classified and grade...
Clinical data available
Technical benchmark set with conversion factors, 3 manual segmentations, and 7 computerized segmentations.
Segmentations: abdominal aorta, intrahepatic umbilical vein, fetal stomach, liver region; quality control excluded images with calipers/acoustic shadows
464 transorbital sonography images for optic nerve sheath analysis with multiple-rater annotations.
Paired breast ultrasound images and videos with expert diagnosis and biopsy results; official Zenodo description reports 257 patients.
266 segmented lesions; BIRADS feature labels; radiologist-annotated
TCIA collection used to build the BrEaST benchmark; 266 segmented lesions with BIRADS descriptors.
992 videos (516 from Site G, 476 from Site R), 4 cardiac views per patient, 100+ frames per video
B-mode muscle ultrasound dataset for neuromuscular echogenicity / Heckmatt grading; includes healthy and FSHD participants.
Histopathologically proven dataset with 123 malignant and 109 benign lesions; radiologist-defined masks.
Brain resection multimodal imaging database with pre-op MRI
ULMShare is an open-access dataset designed to accelerate algorithm development, validation, and benchmarking in Ultrasound Localization Microscopy (ULM). The data spans three years of experimental procedures, including...
A large, publicly accessible ultrasound challenge database comprising 1,833 distinct ultrasound data from 13 different ultrasound image anomalies. Derived from publicly available data on the Internet.
Aixplorer Ultimate system; radiologist-defined masks
256×256 pixels; radiologist-annotated; from ultrasoundcases.info
1,875 images with biopsy-proven tumors; benign/malignant labels, BIRADS, and segmentation masks.
683 images from 38 patients: 174 benign, 90 malignant, 419 normal; RGB masks and expert annotations.
780 images: 437 benign, 210 malignant, 133 normal; PNG format 500x500 pixels
927 breast ultrasound images collected 2020-2022 at Renmin Hospital of Wuhan University; tumor regions provided.
2,200 PNG files (1,100 original + 1,100 masks), 709x749x3 resolution, Mindary UMT-500Plus ultrasound
The dataset consists of aquistified ultrasound images of the common carotid artery.
Contrast transthoracic echocardiography (cTTE), rest and Valsalva maneuver videos, 4 sonographers annotation
Transperineal ultrasound (TPU); image dimensions 256×256; pixel-level labels (background, pubic symphysis, fetal head); restricted test set access
Commercial COVID-19 lung ultrasound phantom (CAE Healthcare); 564 phantom images; segmentation of rib, pleural line, A-line, B-line, B-line confluence
Micro-ultrasound scans and human prostate
3.6 GB; 29-MHz micro-ultrasound (3-4x higher resolution than traditional US); multiple annotator expertise levels; micro-ultrasound-specific modality
Multi-modality: 2D ultrasound and contrast-enhanced ultrasound (CEUS); pixel-wise semantic annotations; designed for unsupervised cross-domain segmentation
Dataset for multi-modality cross-domain semantic segmentation
Biceps brachii, tibialis anterior, and gastrocnemius medialis images with manual segmentations.
A PSFHS dataset comprising 1358 images, meticulously annotated at the pixel level
De-identified; acquired via LOGIQ P7 system
Includes electromagnetic tracking data for registration evaluation
Muscles: biceps brachii, tibialis anterior, gastrocnemius medialis
A reliable automated labeling method and large-scale dataset for ultrasound-based bone surface extraction
Access-controlled multimodal breast imaging dataset hosted on Zenodo.
TCIA CPTAC pancreatic ductal adenocarcinoma collection with multimodal imaging and linked proteogenomic resources.
TCIA CPTAC endometrial carcinoma collection with multimodal imaging plus linked pathology and proteogenomic data.
Limited details; last modified Aug 2025; check TCIA for DICOM
De-identified data in NIFTI format; multiple annotators for segmentation quality
A pre clinical ultrasound dataset of rat organs (brain brain tumor kidney) with an ultrafast device.
Tracked Artemis biopsy sessions with MRI/US fusion
A pre clinical ultrasound dataset of rat brain with an ultrafast device including RF data.
A pre clinical ultrasound dataset of transcranial awake mice brain with an ultrafast device including 4D data.
Video-based image dataset. Two of the four subdatasets were newly collected by experienced doctors.
We make public the ultrasound liver images, which were annotated to show the outlines, livers, and liver mass regions.
50,063 images and video frames; 435 malignant, 200 benign, 100 normal annotated
120 liver-mass ultrasound studies with B-mode and CEUS cine loops; reference standards include pathology and cross-sectional imaging.
1669 B-mode 2D ioUS images; dataset includes NIfTI images and NIfTI tumor masks
Image dimensions 800×540 pixels; landmarks: ellipse parameters (center, semi-axes, angle); pixel size range 0.052–0.326 mm
Prenatal frames by novices; 37.9 GB train set; Apr 2024
Needle trajectory tracking; bounding boxes and active contour masks; expert annotations
Apical 4-chamber and 2-chamber views, LVEndo, LVEpi, LA annotations, 450 training, 50 testing
7,643 labeled videos, 3,176 apical 4-chamber, 4,424 parasternal short axis, ages 0-18 years
PNG images; registration required for data access; dataset does not include video sequences (cines)
53 malignant, 110 benign images
Siemens ACUSON S2000; RGB masks (green=benign, red=malignant); expert radiologist annotations
86,640 DICOM images (2012-2021); AWS S3 public access
Train/test split should be video/patient-level, not frame-level. Includes COVID, healthy, bacterial pneumonia, viral pneumonia, and uninformative class in some workflows.
Endoscopic ultrasound; includes bounding boxes, categories, and anatomical locations; Query2 framework for detection
Multi-vendor dataset (Philips, GE, Acuson, Siemens, Toshiba, SonoSite); includes view type and kidney type (native vs. transplant) classification; real-world clinical settings
Prenatal Ultrasound Frames acquired by novice operators
Open-access raw ultrasonic signal dataset for 100 breast lesions with benign/malignant labels and BI-RADS.
Prostate masks, apex, base, lesions, zonal structures annotated
Thyroid gland occupies minimum 6% of image area; uses .jpg format
Single-class binary segmentation task
No overlap between training and test subjects; competition-based dataset
Multimodal breast ultrasound dataset used with TDF-Net for incomplete-modality diagnosis.
Cancer Moonshot Biobank colorectal collection with multimodal imaging and linked clinical / biobank data.
Cancer Moonshot Biobank lung cancer collection with multimodal imaging and linked clinical / biobank data.
Cancer Moonshot Biobank melanoma collection with longitudinal imaging and linked clinical / biobank data.
Challenge dataset for 3D ABUS tumor detection / segmentation / classification; access-controlled, with 100 training cases and challenge splits.
Requires citation; contact administrators for specific terms
Apical 4-chamber view videos, left ventricle semantic segmentation, frame-by-frame tracking
163 cases used for automated lesion detection; commonly cited split is 53 malignant and 110 benign.
Blind-sweep acquisition; low-resource settings (Africa/Europe); novice operators
188 ultrasound videos (113 malignant, 75 benign), totaling 25,272 frames with bounding-box annotations.
780 images: 437 benign, 210 malignant, 133 normal; PNG images around 500x500 pixels.
480-1,638 images (source variation), 637 with pixel-level labels, thyroiditis, goiters, nodules, cancer cases
ROI-free breast cancer diagnosis dataset released with HoVer-Trans; 2,405 images (886 benign, 1,519 malignant) from GDPH and SYSUCC.
Publicly available annotated image database for comparative studies.
Binary segmentation masks for AC and HC; training/validation splits in separate .txt files; Fast U-Net model weights available
Breast and axilla case archive on ultrasoundcases.info with lesion-type labels and mixed image modalities.
Public BUS dataset used in Rodtook et al.; 180 patients / 296 images across malignant, benign, and cyst cases.
4,554 total PNG format images; resolution range 335-719 pixels
2D ultrasound images with tumor segmentation masks
617 real US scans + 926 synthetic; organs segmented: liver, kidneys, pancreas, vessels, adrenals, gallbladder, bones, spleen
11 data formats supported; high annotation reliability (ICC 0.859-0.889)
README says it contains four classes: normal, cataract, glaucoma, retina disease. Version shown: 2, size 1.79 GB, 602 files.
812 ultrasound images with muscle bundle and fascia masks; B-mode ultrasound
External validation dataset; GE Voluson E10 system; benign/malignant not differentiated
A collection of files from across the Internet to help researchers and developers in the field of hip dysplasia.
Repository describes generation of 256x256 trans-cerebellum fetal ultrasound images using DSR-GAN and TB-GAN.
Includes biopsy-confirmed thyroid nodules, TI-RADS descriptors, histopathological diagnoses
