EchoCare: Ultrasound AI Foundation & Guidance

• EchoCare is a versatile ultrasound AI framework that serves as a general foundation model, a transfer-learning backbone for fetal cardiac tasks, and a point-of-care echocardiography guidance system.
• It leverages self-supervised learning on the large-scale EchoCareData corpus (over 4.5M images) to improve generalization and performance across diverse ultrasound clinical applications.
• The system demonstrates practical benefits in disease diagnosis, segmentation, and real-time LVEF estimation via cascaded pipelines and multimodal alignment techniques.

EchoCare is a name used in recent ultrasound AI literature in two closely related but non-identical senses. Its primary usage denotes a fully open, ultrasound-specific foundation model for generalist clinical use, developed via self-supervised learning on the large-scale EchoCareData corpus and intended to support disease diagnosis, segmentation, detection, regression, enhancement, and report generation. In later work, the same pretrained backbone is reused as a component in semi-supervised fetal cardiac ultrasound systems and in an ultrasound vision-LLM, while a separate point-of-care transthoracic echocardiography study uses “EchoCare” to denote a cascaded assistance system for apical 4-chamber view acquisition and automated left ventricular ejection fraction estimation (Zhang et al., 15 Sep 2025, Wang et al., 18 Mar 2026, Lyu et al., 4 May 2026, Zhuang et al., 19 May 2026, Guo et al., 28 Mar 2026).

1. Scope, naming, and research setting

In the 2025 paper “A Fully Open and Generalizable Foundation Model for Ultrasound Clinical Applications,” EchoCare is presented as “a novel ultrasound foundation model for generalist clinical use,” trained on EchoCareData and benchmarked across ten representative ultrasound tasks. The same paper frames EchoCare as a response to a recurring limitation in ultrasound AI: many prior models are trained on small, private datasets or are narrowly specialized to a single organ or task, and therefore do not generalize well across real-world ultrasound workflows (Zhang et al., 15 Sep 2025).

Subsequent papers use EchoCare as an upstream pretrained backbone rather than as a standalone end application. In FM-DACL, EchoCare is “a pretrained ultrasound foundation model” and the “primary network” in a heterogeneous semi-supervised pair for fetal heart ultrasound segmentation and congenital heart disease classification. In the later semi-supervised fetal cardiac framework built with SAM-Med2D and DINOv3, EchoCare is the shared multi-task backbone with a ViT encoder, a UNETR-style decoder, and classification heads for CHD and fetal cardiac view recognition (Wang et al., 18 Mar 2026, Zhuang et al., 19 May 2026).

A separate naming usage appears in “Follow Your Heart: Landmark-Guided Transducer Pose Scoring for Point-of-Care Echocardiography.” There, EchoCare is described as a practical point-of-care transthoracic echocardiography assistance system centered on the apical 4-chamber view, with a cascaded pipeline for landmark detection, probe-pose scoring, and LVEF estimation from diagnostically suitable clips (Guo et al., 28 Mar 2026).

This distribution of usages suggests that “EchoCare” functions less as a single immutable architecture than as a research lineage: a general ultrasound foundation model, several derivative transfer-learning systems, and one separately named point-of-care guidance application.

2. EchoCare as a general ultrasound foundation model

EchoCare is built on EchoCareData, which the paper describes as the largest publicly available ultrasound image dataset to date. The dataset contains over 4.5 million ultrasound images. The narrative states that these images were assembled from 138 datasets collected from over 20 countries across 5 continents, while the abstract states “over 23 countries across 5 continents.” EchoCareData is explicitly multi-center, multi-device, and multi-ethnic, and spans multiple imaging modalities, including B-mode, CEUS, Doppler, M-mode, and elastography. It covers 9 major body regions and 52 anatomical organs/types in the narrative, while the methods section mentions 32 representative organs in one place. The curation pipeline draws from open repositories and public platforms such as Zenodo, Mendeley, Stanford AIMI, Figshare, Kaggle, GitHub, and Grand Challenge, with manual inspection, deduplication, and removal of non-ultrasound images, sensitive content, and textual noise (Zhang et al., 15 Sep 2025).

Architecturally, EchoCare is a self-supervised masked image modeling framework extended with a second branch for hierarchical anatomy classification. The paper describes a modular system with an image encoder, an image decoder, and a meta-object classifier. The masked image modeling branch takes an image $x \in \mathrm{R}^{H \times W \times C}$, masks 50% of the patches, and reconstructs the missing content with a mean absolute error objective:

$$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$

The second branch introduces a three-level anatomy ontology consisting of body parts, organs, and anatomical structures. The paper defines a hierarchy-coherent probability update as

$$\begin{cases} p_{H} = \min(s_u) & \text{if } \hat{l} = 1, \ 1 - p_{H} = \min(1-s_u) = 1 - \max(s_u) & \text{if } \hat{l} = 0, \end{cases}$$

and optimizes the hierarchical loss

$$\mathcal{L}_{\mathrm{HIE}} = \sum -\hat{l}\log(p_H) - (1 - \hat{l})\log(1 - p_H).$$

The stated motivation is that joint reconstruction-plus-hierarchy learning allows EchoCare to capture both local pixel-level ultrasound appearance and global anatomical context. The total pretraining objective is described as $$\mathcal{L}_{\mathrm{MIM}} + \mathcal{L}_{\mathrm{HIE}}$$ (Zhang et al., 15 Sep 2025).

The pretraining pipeline uses random flips, crops, grayscale conversion, and weak color jitter, and runs for one million steps with batch size 256 on $256 \times 256$ images using AdamW, cosine learning-rate decay, 10,000-step warmup, weight decay 0.05, and stochastic depth 0.1. The paper explicitly positions this as an ultrasound-specific alternative to off-the-shelf vision foundation model architectures.

3. Benchmark coverage, transfer behavior, and openness

EchoCare is evaluated on 10 external validation tasks spanning 11 clinical applications and 8 task types. The benchmark coverage includes disease diagnosis/classification, lesion segmentation, organ detection, landmark prediction, quantitative regression, image enhancement, and report generation. The specific downstream tasks named in the paper include thyroid nodule benign/malignant classification, breast BI-RADS grading, focal liver lesion classification, thyroid nodule segmentation, artery/vein segmentation, abdominal multi-organ segmentation, fetal thorax/cardiac detection, fetal brain landmark prediction, left ventricular ejection fraction regression, low-quality ultrasound enhancement, and ultrasound report generation (Zhang et al., 15 Sep 2025).

Several reported results are concrete. For classification, EchoCare attains AUC 86.48% and F1 87.45% for thyroid nodule benign-vs-malignant classification, 70.36% accuracy and 65.38% macro-F1 for breast BI-RADS grading, and 87.12% accuracy and 83.44% macro-F1 for focal liver lesion classification. For segmentation, it is reported to improve thyroid nodule segmentation by 2.09% DSC and 2.26% NSD, vessel segmentation by 1.36% mDSC and 1.03% mNSD, and abdominal multi-organ segmentation by 2.10% mDSC and 4.36% mNSD. On FOCUS, EchoCare achieves 97.26% AP for thoracic detection and 96.11% AP for cardiac detection, with a 5.42% mAP margin over the top ImageNet-based detector, and 94.42% accuracy on CTR measurement. On BrainBenchmark, it achieves average MSE 7.71. On CAMUS, EchoCare achieves MAE 3.91 for LVEF prediction, reported as a 19% reduction versus USFM and a 43% reduction versus EchoMEM. On the USData Liver report-generation benchmark, the paper states that an EchoCare-based encoder-decoder achieves the best performance on all ten language and classification metrics (Zhang et al., 15 Sep 2025).

The paper also emphasizes label efficiency and adaptation efficiency. It states that EchoCare remains strong when only a fraction of downstream labels are available, can outperform other models in thyroid segmentation using only 80% of the labeled training data, and can reduce downstream training time by about 20% to 40%. EchoCare and EchoCareData are both publicly released, with code, pretrained weights, example datasets, installation instructions, and downstream evaluation scripts made available. The paper therefore frames openness as a substantive part of the contribution, not merely a distribution detail.

Important limitations are also explicit. Pretraining currently uses only image data; the model treats videos as static frames rather than exploiting temporal dynamics; and the paper notes that real-world clinical adoption will require rigorous validation and integration into clinical decision-support systems. This suggests that the strongest current evidence concerns broad transferability across image-centric ultrasound tasks, while spatiotemporal and multimodal extensions remain open directions.

4. EchoCare in semi-supervised fetal cardiac ultrasound analysis

In FM-DACL, EchoCare is reused as a pretrained ultrasound foundation model in a semi-supervised framework for fetal heart ultrasound segmentation and CHD diagnosis. The method pairs EchoCare, denoted $f^1_\theta(\cdot)$, with a U-Net, denoted $f^2_\phi(\cdot)$, and combines supervised learning with cross pseudo supervision, interpolation consistency learning, and dual agreement consistency. The paper explicitly describes EchoCare as the “primary network,” states that it is loaded from the official foundation model checkpoint, and notes that it is the only network used at inference time for final predictions (Wang et al., 18 Mar 2026).

The total loss is given as

$$\mathcal{L}_{\text{total}} = (\mathcal{L}_{\text{sup}}^{1}+\mathcal{L}_{\text{sup}}^{2}) + \lambda(\mathcal{L}_{\text{cps}}^{1}+\mathcal{L}_{\text{cps}}^{2}) + \tau \mathcal{L}_{\text{ict}} + \beta \mathcal{L}_{\text{dac}},$$

with supervised terms containing cross-entropy, Dice loss, and binary cross-entropy with logits for segmentation and multi-label CHD classification. The best EchoCare-based configuration reported in Table 1 is FM-DACL (EchoCare / ResUNet$\dagger$), with DSC 59.66, NSD 42.82, F1 30.62, and Score 36.84. The authors state that the method does not outperform the official baseline, but does show stable performance across backbone choices and demonstrates the feasibility of the semi-supervised framework (Wang et al., 18 Mar 2026).

A later fetal cardiac ultrasound paper uses EchoCare as a multi-task backbone with a Vision Transformer encoder, a UNETR-style decoder, and classification heads for 7-dimensional CHD prediction and 4-view classification. In that work, EchoCare outputs 15 segmentation categories, consisting of 1 background class and 14 cardiac structures. The framework adds SAM-Med2D for boundary refinement, DINOv3 for pseudo-label filtering, view-specific hard masking, and a two-stage optimization scheme with an EMA phase followed by Classification Fine-Tuning. The final reported FETUS 2026 leaderboard result is F1-score 41.20%, DSC 79.99%, NSD 61.62%, and Overall Score 56.00%. Compared with the baseline EchoCare result of F1 34.20, DSC 65.48, and NSD 45.55, the paper attributes the gains to boundary refinement, semantic pseudo-label filtering, anatomical masking, and staged optimization (Zhuang et al., 19 May 2026).

Taken together, these fetal applications indicate that EchoCare is being used as a transferable anatomical prior under severe annotation scarcity. A plausible implication is that the value of the foundation model is not restricted to generic ultrasound transfer, but extends to regimes where pseudo-label quality and feature complementarity dominate performance.

5. EchoCare-CLIP and ultrasound vision-language alignment

EchoCare-CLIP extends the EchoCare family into cross-modal representation learning. The paper introduces a CLIP-style dual-encoder framework intended to align ultrasound images with clinical text in a shared embedding space, motivated by the observation that prior ultrasound foundation models such as EchoCare and USFM are vision-only and therefore do not naturally support image-text matching or CLIP-style zero-shot classification (Lyu et al., 4 May 2026).

The image branch uses the EchoCare SwinTransformerV2 backbone pretrained on about 4.5M ultrasound images. Inputs are $$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$0 ultrasound images, and the image encoder produces a 2048-dimensional representation with a backbone size of about 120M parameters. Two text encoder families are evaluated: the OpenAI CLIP text encoder and BioClinicalBERT. Both image and text features are passed through a two-layer MLP projection head,

$$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$1

and mapped into a common 256-dimensional latent space with $$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$2 normalization (Lyu et al., 4 May 2026).

The model is trained with a symmetric InfoNCE / CLIP-style contrastive loss:

$$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$3

with cosine-similarity scores

$$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$4

and temperature initialized to $$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$5 and learned during training as $$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$6.

The training corpus contains 16,464 ultrasound image-text pairs from six public datasets spanning breast, liver, lung, and thyroid. Over 78% of captions derive from expert-annotated reports; the remaining samples are paired using a three-tier template-based captioning pipeline and an alternative LLM-based rewriting pipeline. The best internal paired alignment score is 0.682 for BioClinicalBERT plus full fine-tuning. On external held-out zero-shot classification, the strongest reported scores are 0.709 on BUSI and 0.626 on AULI, both from CLIP-based partially fine-tuned variants. The paper also reports that template-based captions match or outperform LLM-generated captions, despite the latter having substantially greater lexical diversity (Lyu et al., 4 May 2026).

A central finding is that stronger cross-modal alignment does not guarantee better downstream transfer. The paper reports that full end-to-end fine-tuning can degrade zero-shot and few-shot generalization, while partial fine-tuning of one encoder often yields the best trade-off between domain adaptation and representational generalizability. This suggests that EchoCare’s visual prior remains useful in multimodal settings, but that preserving pretrained geometry is as important as increasing paired alignment scores.

6. EchoCare as a point-of-care echocardiography assistance system

In “Follow Your Heart: Landmark-Guided Transducer Pose Scoring for Point-of-Care Echocardiography,” EchoCare denotes a practical point-of-care transthoracic echocardiography assistance system centered on the apical 4-chamber view rather than the general ultrasound foundation model. The system is designed for settings in which novice users must obtain diagnostically useful A4CH views without external transducer tracking hardware. Its pipeline is cascaded: an uncertainty-aware landmark detector first estimates anatomy, a transducer pose scoring module then classifies pose quality, and a deep R(2+1)D network estimates LVEF only from clips deemed suitable for diagnosis (Guo et al., 28 Mar 2026).

The landmark detector is trained from EchoNet Dynamic, which the paper states contains 10,000 video clips and 20,000 annotated frames, with 42 LV contour landmarks per frame. The authors add about 3,000 frames with five additional A4CH landmarks—RV, RA, LA, mid-plane of the tricuspid valve, and tricuspid valve annulus—yielding 47 landmarks in total. The detector uses a modified ResNet34 encoder-decoder with five upsampling layers of channel dimensions 512, 256, 128, 64, and 64, followed by a $$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$7 convolution to produce 47 heatmap channels. The batch loss is a weighted negative log likelihood:

$$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$8

Probe guidance is formulated as a three-level “traffic light” scoring problem: on target, close to target, or far from target. The frame-level rubric deducts points for missing anatomy or off-axis structure appearance: LV free wall not visible: -1; RV free wall not visible: -0.5; LA entirely/partially out of view: -2 / -1; RA not visible: -0.5; aorta clearly visible (apical 5-chamber view): -1; other significant signal dropout: -0.5. The final label is assigned by total deduction: red if total deduction $$\mathcal{L}_{\mathrm{MIM}} = \frac{1}{M}\sum_{i\in M} |\hat{x}_i^p - x_i^p|.$$9, yellow if total deduction $$\begin{cases} p_{H} = \min(s_u) & \text{if } \hat{l} = 1, \ 1 - p_{H} = \min(1-s_u) = 1 - \max(s_u) & \text{if } \hat{l} = 0, \end{cases}$$0, else green. Two backbones are evaluated for pose scoring: a ResNet18 regression model and a multimodal LLaMA/LLaMA-Adapter model. Across subject-level 5-fold cross-validation, the mean test accuracy is 0.69 for the ResNet18 model with images plus landmarks and 0.71 for the LLaMA-based model with images plus landmarks; the paper also states that the overall approach achieves 71% accuracy averaged across red, yellow, and green frames, and up to 87% for red frames (Guo et al., 28 Mar 2026).

For landmark localization, the average mean Euclidean distance error across all visible landmarks is $$\begin{cases} p_{H} = \min(s_u) & \text{if } \hat{l} = 1, \ 1 - p_{H} = \min(1-s_u) = 1 - \max(s_u) & \text{if } \hat{l} = 0, \end{cases}$$1 pixels. For automated LVEF estimation, the downstream EchoNet model predicts an average LVEF of $$\begin{cases} p_{H} = \min(s_u) & \text{if } \hat{l} = 1, \ 1 - p_{H} = \min(1-s_u) = 1 - \max(s_u) & \text{if } \hat{l} = 0, \end{cases}$$2 across 10 subjects, compared with a ground-truth mean of $$\begin{cases} p_{H} = \min(s_u) & \text{if } \hat{l} = 1, \ 1 - p_{H} = \min(1-s_u) = 1 - \max(s_u) & \text{if } \hat{l} = 0, \end{cases}$$3 as visually assessed by a cardiac anesthesiologist. The real-time pipeline of landmark detection plus pose scoring with the ResNet18 backbone runs at about 14 frames per second on an Apple M3 Pro with 36 GB of memory (Guo et al., 28 Mar 2026).

This use of the name highlights an important distinction. Here EchoCare is not a large-scale reusable foundation encoder, but an end-to-end acquisition-and-function-assessment workflow for point-of-care scanning. The naming overlap is therefore substantive but not architectural identity: both usages target ultrasound decision support, yet one emphasizes generalist pretraining and downstream transfer, while the other emphasizes anatomy-aware acquisition guidance and immediate functional readout.