EchoCareData: Multi-Center Ultrasound Repository

• EchoCareData is a comprehensive ultrasound data resource aggregating over 4.5 million images from diverse centers worldwide for multi-task clinical research.
• It employs a three-tier ontology to harmonize heterogeneous annotations across various modalities, anatomical regions, devices, and acquisition settings.
• The echo-specific workflow uses controlled video diffusion to synthesize apical two-chamber views, significantly enhancing ejection fraction estimation from limited data.

Searching arXiv for the specified papers to ground the article with current records. EchoCareData denotes a curated ultrasound data resource associated with two closely related research directions. In the broader formulation, it is the first publicly available, large-scale, multi-center, multi-device and multi-ethnic ultrasound repository, assembled to support self-supervised pre-training for general clinical ultrasound applications and used to develop the EchoCare foundation model (Zhang et al., 15 Sep 2025). In a narrower echocardiography-focused formulation, the name also refers to a corpus-building pipeline that combines real apical biplane videos with synthetic apical two-chamber views generated by a controlled video diffusion model, with the goal of improving ejection fraction estimation from limited point-of-care ultrasound acquisitions (Kondori et al., 25 Aug 2025). This suggests that “EchoCareData” functions both as a public foundation-model pre-training corpus and as an echo-specific data expansion recipe centered on view completion.

1. Definition and scope

EchoCareData, in its public repository form, comprises 4,500,000+ de-identified ultrasound image–ontology tuples drawn from 138 publicly accessible datasets collected across more than 23 countries on five continents: Asia, Europe, North America, South America, and Oceania (Zhang et al., 15 Sep 2025). The corpus includes B-mode, contrast-enhanced ultrasound (CEUS), Doppler, M-mode, and elastography examinations, and spans nine high-level anatomical regions: head, chest, abdomen, limbs, back, fetus, dorsum, pelvis, and “other” (Zhang et al., 15 Sep 2025).

The dataset is explicitly positioned as a pre-training resource for comprehensive clinical ultrasound applications rather than as a task-specific benchmark. Its design emphasizes heterogeneity across centers, devices, anatomies, and acquisition settings. No single center contributes more than 15% of the total, and representative scanners include GE Healthcare, Siemens, Toshiba, Samsung, SuperSonic Imaging, and handheld devices such as mSonics MU1 and SSUN, using linear, convex, and high-frequency transducers (Zhang et al., 15 Sep 2025).

In the echocardiography-specific workflow, EchoCareData is framed as a large-scale, high-quality echocardiography corpus built from paired apical four-chamber (A4C) and apical two-chamber (A2C) studies, with synthetic A2C generation used to compensate for the harder-to-acquire view (Kondori et al., 25 Aug 2025). That formulation is centered on ejection fraction (EF), a critical parameter traditionally measured from biplane apical views.

2. Internal structure, ontology, and statistical profile

A three-tier ontology organizes semantic content across the public corpus. Level 1 encodes clinical region with 9 classes; Level 2 encodes meta-object or organ with 52 classes, including examples such as “liver,” “fetal brain,” and “carotid vessel”; and Level 3 encodes specific anatomical structure with 56 classes, including examples such as “left cardiac ventricle” and “cerebellar perimeter landmark” (Zhang et al., 15 Sep 2025). This ontology is the principal mechanism for harmonizing labels across otherwise heterogeneous source datasets.

Where demographic metadata are available, EchoCareData covers age spans from early embryonic/fetal development at 10 weeks gestation through geriatric patients older than 80 years, includes a balanced mix of male and female subjects, and spans diverse clinical conditions including healthy volunteers, benign and malignant lesions, and congenital anomalies (Zhang et al., 15 Sep 2025). These attributes are not uniformly available for every source, but they define the intended breadth of the repository.

The modality distribution is dominated by B-mode, which constitutes approximately 90% of the corpus, followed by CEUS at approximately 5%, Doppler at approximately 3%, M-mode at approximately 1%, and elastography at approximately 1% (Zhang et al., 15 Sep 2025). The region-level totals reported in the dataset summary are as follows.

Region	Total images	Brief profile
Head	500,000	B-mode dominant
Chest	1,000,000	B-mode dominant
Abdomen	1,725,000	Largest regional subset
Limbs	200,000	Smaller peripheral subset
Fetus	400,000	Obstetric and fetal imaging
Other	600,000	Miscellaneous ultrasound content

These counts come from the published table totaling 4,418,000 images; the reported numbers are rounded, and the remainder includes rare “back,” “pelvis,” and “dorsum” classes (Zhang et al., 15 Sep 2025). The discrepancy between 4.5 million+ overall tuples and 4,418,000 tabulated images reflects that rounding and the exclusion of rare classes from the table rather than a contradiction in dataset identity.

3. Acquisition, preprocessing, annotation, and governance

The public corpus was assembled through exhaustive screening of candidate repositories from Zenodo, Kaggle, Grand Challenge, GitHub, Mendeley, Stanford AIMI, and Figshare using the keyword “ultrasound” (Zhang et al., 15 Sep 2025). Initial filtering retained image and compressed archive formats including PNG, JPG, BMP, DICOM, ZIP, and TFRecord, and a GPT-4o-assisted parser was used to retrieve download links where required. Exact and hash-based deduplication were then applied, followed by manual review by imaging scientists to remove non-ultrasound scans, phantoms, and text-only submissions (Zhang et al., 15 Sep 2025).

Before inclusion, every image was screened for minimal content exceeding 1,000 non-zero pixels, for removal of protected health information in pixel-overlaid text, and for automated frame sampling with a 10-frame stride in video sequences (Zhang et al., 15 Sep 2025). All images were then rescaled to $256 \times 256$ pixels by bicubic interpolation, converted to single-channel greyscale, and stored as lossless PNG for pre-training. During self-supervised learning, on-the-fly augmentation includes random horizontal and vertical flips, random cropping and center-padding to $256 \times 256$, and intensity normalization

$I' = \frac{I - \mu_I}{\sigma_I},$

where $\mu_I$ and $\sigma_I$ are the per-image mean and standard deviation, together with random brightness, contrast, saturation, and hue jitter (Zhang et al., 15 Sep 2025).

Annotation is hierarchical rather than dense. Each image carries a class label at ontology Levels 1–3, assigned by automated mapping from source metadata to the unified ontology and then manually validated (Zhang et al., 15 Sep 2025). For each class, two board-certified sonographers independently reviewed a stratified random sample of 100 images and corrected misclassifications. Pairwise inter-rater agreement exceeded Cohen’s kappa $\kappa = 0.82$, and discrepancies were reconciled through consensus review (Zhang et al., 15 Sep 2025). No pixel-wise segmentation or landmark annotations are included in the pre-training corpus; the dataset is designed primarily for representation learning and hierarchical image classification. This directly counters the common assumption that a large ultrasound foundation-model corpus must also provide universal dense annotations.

EchoCareData is distributed as one monolithic corpus without predefined train/validation/test splits. Suggested downstream practice is to sample disjoint subsets, for example 80%/10%/10%, stratified by Level 1 region or Level 2 organ (Zhang et al., 15 Sep 2025). From an administrative standpoint, all source datasets were fully anonymized or stripped of patient identifiers prior to curation; DICOM tags and pixel-burned text were removed in compliance with HIPAA Safe Harbor and GDPR de-identification guidelines. The release license is CC BY-NC 4.0, and the corpus, ontology mapping tables, code, and metadata are made available through the EchoCare project website and GitHub repository (Zhang et al., 15 Sep 2025).

4. Echocardiography-specific synthetic expansion by controlled video diffusion

In the echocardiography-focused usage, EchoCareData is constructed from two complementary elements: assembly and preprocessing of real apical biplane videos, and synthetic completion of the harder-to-acquire A2C view using a controlled video diffusion model of the type introduced in ControlEchoSynth (Kondori et al., 25 Aug 2025). The underlying architecture consists of a 3D denoising U-Net with a ControlNet branch. The U-Net backbone is an encoder–decoder with ResNet blocks at each scale, temporal self-attention layers for motion modeling, and 3D up/down-sampling. The diffusion process uses a cosine-annealed, 1,000-step discrete schedule $\{\beta_t\}_{t=1}^T$ with $\alpha_t = 1 - \beta_t$, and the model is trained in the standard discrete denoising diffusion framework (Kondori et al., 25 Aug 2025).

Conditioning is performed through a parallel ControlNet branch whose encoder and middle weights are copied from the U-Net and topped by three Zero-3DConv layers. The control signal is the A4C video concatenated channel-wise with a motion mask, and this fused input is injected into each U-Net level through zero-initialized convolutions (Kondori et al., 25 Aug 2025). Real data are drawn from paired A4C–A2C studies, exemplified by CAMUS plus an internal 8,074-pair set. Temporal alignment resamples each video to 32 frames via a sliding window and then central-crops to 16 frames by nearest-neighbor selection; frames are resized to $64 \times 64$ by bilinear interpolation. Motion masks are formed by per-pixel absolute differences between consecutive A4C frames, smoothed with a small Gaussian kernel, and concatenated to the raw A4C stream to form a $16 \times 2 \times 64 \times 64$ control tensor (Kondori et al., 25 Aug 2025).

Synthetic dataset construction is explicitly selective rather than unconditional. For each real A4C clip, 18 synthetic A2C candidates are generated by ancestral sampling through the 1,000-step reverse diffusion process. A pretrained EF estimator, such as ResNet2+1D, is then applied to each candidate; the absolute EF error $256 \times 256$0 is computed, and the top $256 \times 256$1 candidates with the smallest $256 \times 256$2 are retained (Kondori et al., 25 Aug 2025). Applied to 8,524 real A4C clips, consisting of 450 CAMUS and 8,074 internal studies, this yields up to approximately 25,572 high-fidelity synthetic A2C videos. Each retained synthetic A2C inherits the ground-truth EF label from the corresponding A4C side after ranking (Kondori et al., 25 Aug 2025).

The implementation details are unusually explicit. The diffusion model uses $256 \times 256$3 timesteps with cosine-annealed $256 \times 256$4 from $256 \times 256$5 to $256 \times 256$6; the unconditional U-Net phase runs for 100,000 iterations with learning rate $256 \times 256$7 to $256 \times 256$8 under a cosine scheduler and batch size 8, while the conditional phase fine-tunes the U-Net and ControlNet for 80,000 iterations at learning rate $256 \times 256$9 with 10-step warmup (Kondori et al., 25 Aug 2025). Reproducibility recommendations include fixing all random seeds, freezing ControlNet’s zero-init layers until after the first epoch, and using mixed precision on a single NVIDIA V100 16 GB GPU.

5. Demonstrated model pipelines and empirical effects

The main public use of EchoCareData is pre-training the EchoCare foundation model, described as a self-supervised Masked AutoEncoder with an added hierarchical anatomy-classification branch (Zhang et al., 15 Sep 2025). The hierarchical classifier is introduced to enable joint learning of pixel-level and representation-level features, thereby capturing both global anatomical contexts and local ultrasound characteristics. Pre-training on 4.5 million images for 1 million steps produces representations that, when fine-tuned on ten independent benchmarks, achieve state-of-the-art performance across classification, segmentation, detection, landmark localization, regression, enhancement, and report generation tasks (Zhang et al., 15 Sep 2025).

The reported benchmark gains are quantitative rather than merely descriptive. Fine-tuned models show an AUC uplift of 3–5% for classification, Dice improvement of 2–4% for segmentation, mAP improvement of 5% for detection, standard detection rate improvement of 6–8% for landmark localization, MAE reduction of 19% for regression, FID reduction of 15% for enhancement, and BLEU-4 and F1 gains of 4.6% and 18.7%, respectively, for report generation (Zhang et al., 15 Sep 2025). Downstream performance is reported to scale logarithmically with pre-training set size,

$I' = \frac{I - \mu_I}{\sigma_I},$0

which is presented as evidence that corpus scale materially affects generalization.

In the echocardiography-specific pipeline, two off-the-shelf EF-estimation architectures are implemented: ResNet2+1D, described as a 2D ResNet front end with 1D temporal convolutions, and EchoCoTr-S, described as a spatiotemporal transformer built on UniFormer (Kondori et al., 25 Aug 2025). Inputs can be single-view clips of shape $I' = \frac{I - \mu_I}{\sigma_I},$1 or biplane pairs concatenated channel-wise as $I' = \frac{I - \mu_I}{\sigma_I},$2, and the output is a scalar EF estimate. Integration strategies include pre-training on synthetic A2C plus real A4C followed by fine-tuning on real pairs, minibatch data augmentation with real and synthetic pairs at ratios 1:1 or 1:3, and an “Augmented” mode that combines real A2C with synthetic A2C paired to the same A4C to expand the dataset by approximately 3× (Kondori et al., 25 Aug 2025).

Performance is evaluated on a held-out CAMUS test set of 50 patients using $I' = \frac{I - \mu_I}{\sigma_I},$3, MAE, and RMSE. For ResNet2+1D, A4C-only input yields $I' = \frac{I - \mu_I}{\sigma_I},$4, MAE $I' = \frac{I - \mu_I}{\sigma_I},$5, and RMSE $I' = \frac{I - \mu_I}{\sigma_I},$6, whereas A4C plus synthetic A2C yields $I' = \frac{I - \mu_I}{\sigma_I},$7, MAE $I' = \frac{I - \mu_I}{\sigma_I},$8, and RMSE $I' = \frac{I - \mu_I}{\sigma_I},$9 (Kondori et al., 25 Aug 2025). For EchoCoTr-S, A4C-only input yields $\mu_I$0, MAE $\mu_I$1, and RMSE $\mu_I$2, whereas A4C plus synthetic A2C yields $\mu_I$3, MAE $\mu_I$4, and RMSE $\mu_I$5 (Kondori et al., 25 Aug 2025). The published summary states that purely synthetic biplane training outperforms single-view or real-only biplane training, and that mixing real and synthetic data produces further gains for the transformer model.

6. Related generative context, evaluation discipline, and limitations

EchoCareData’s echocardiographic extension belongs to a wider line of work on controllable synthetic ultrasound. A closely related example is ECHOPulse, an ECG-conditioned echocardiography video generation model that uses VQ-VAE tokenization, masked visual token modeling, and ECG token conditioning rather than text prompts or expert annotations (Li et al., 2024). ECHOPulse is trained on public datasets including CAMUS and EchoNet-Dynamic, as well as a private set of 94,078 apical 2- and 4-view ECHO videos with simultaneously recorded 12-lead ECG, and reports a 64-frame sampling time of 6.4 s for ECHOPulse versus 146 s for EchoDiffusion (Li et al., 2024). On private A2C data, ECG conditioning improves generated video quality relative to text prompting, with FID $\mu_I$6, FVD $\mu_I$7, and SSIM $\mu_I$8, and EF alignment for generated videos is reported as $\mu_I$9, MAE $\sigma_I$0 percentage points, and RMSE $\sigma_I$1 (Li et al., 2024). This provides a distinct conditioning paradigm—physiological time-series control instead of view-conditioned diffusion—for the broader problem of synthetic echo generation.

The EchoCareData literature also defines practical safeguards against synthetic-data misuse. In the echocardiography setting, train/validation/test splitting is recommended at the patient level, for example 450/50/50 for CAMUS, and a small real-only test set should always be held out to detect over-fitting to synthetic artefacts (Kondori et al., 25 Aug 2025). Synthetic A2C clips are to be curated via EF-error ranking, and discarding the bottom half of candidates is recommended to avoid training on “hallucinated” anatomies. Mixing synthetic and real data at ratios greater than 3:1 is discouraged because it may cause over-specialization to synthetic statistics; additional monitoring should track EF bias across the EF range and include periodic side-by-side inspection of synthetic and real clips for motion consistency and left ventricular boundary realism (Kondori et al., 25 Aug 2025).

These constraints clarify what EchoCareData is and is not. It is not a universally annotated ultrasound benchmark, since the public corpus excludes pixel-wise segmentations and landmarks (Zhang et al., 15 Sep 2025). It is not a fixed split benchmark, since the public release is monolithic and downstream splitting is delegated to users (Zhang et al., 15 Sep 2025). And in its echo-specific form, it is not merely a raw repository of acquired videos, but a procedure for expanding apical biplane supervision through controlled generation and post hoc curation (Kondori et al., 25 Aug 2025). A plausible implication is that the name now spans two levels of abstraction: a globally aggregated ultrasound foundation-model corpus, and a specialized echocardiographic data-engineering workflow for improving EF estimation under limited-view acquisition constraints.