cs.CVOct 4, 2026

EchoDino: A pediatric foundation model for transferable echocardiographic analysis across the lifespan

Authors: Sheng Cheng, Donnchadh M. O'Sullivan, Daniel J. Penny, Craig G. Rusin, Minh B. Nguyen, Devika Subramanian

Organizations: Department of Computer Science, Rice University, Houston, TX, USA · Division of Pediatric Cardiology, Baylor College of Medicine, Houston, TX, USA · Texas Children’s Hospital, Houston, TX, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Echocardiography is the most widely used cardiac imaging modality, yet interpretation demands integrating visual evidence across global anatomy, localized structures and dynamic cardiac motion. Machine-learning models have automated individual tasks, but they are typically built for a single purpose and depend on expensively labeled datasets - a barrier particularly acute in pediatric care, where data are scarce and anatomy changes with age. Here we present EchoDino, a self-supervised foundation model for echocardiography, created by adapting the DINOv3 framework to 3.7 million frames from 1.7 million unlabeled pediatric echocardiography videos. With its encoder frozen, EchoDino produces representations that capture global context, local anatomy, and dense spatial detail. We introduce Motion-biased Entropy Maximization Sampling (MEMS) to select the most informative frames for video-level analysis. Across nine pediatric and adult datasets, EchoDino outperformed strong baseline models, raising view-classification accuracy from 0.609 to 0.889 and the area under the receiver operating characteristic curve for structural-heart-disease detection from 0.811 to 0.872, while also cutting age-estimation error from 3.857 to 1.389 years, achieving the best segmentation accuracy and lowering ejection-fraction errors. By generalizing from label-free pediatric data to adult echocardiography, EchoDino offers a versatile foundation for cardiac image analysis across the lifespan.

Figures & tables

Figure 1: Overview of EchoDino. (a) EchoDino adapts a DINOv3 ViT-L backbone using 3,747,848 unlabeled pediatric echocardiography frames. Self-supervised training aligns image-level representations across augmented views to capture global acquisition context, predicts teacher features for masked patches to retain localized anatomy, and preserves dense relationships among patch tokens during extended training. Training continued for 300,000 steps, and the final checkpoint was used for all downstream evaluations. (b) Frame-based tasks use EchoDino’s global class tokens, local patch tokens, or multiscale intermediate features with a frozen encoder. (c) For video-based tasks, a lightweight temporal module aggregates frozen frame representations to produce video-level predictions. (d) EchoDino consistently outperforms baseline models across frame- and video-based tasks. (e) EchoDino accurately performs segmentation outputs across views. Dashed contours indicate ground truth, and filled regions show model predictions. (f) EchoDino predicts age across pediatric age groups in an expected non-linear fashion. (g) Uniform and MEMS frame selections from the same video. The progress bar above each frame indicates its relative position between systole (S) and diastole (D). Compared with uniform frame sampling, the proposed Motion-biased Entropy Maximization Sampling prioritizes feature-diverse frames with substantial motion.
DatasetPatientsStudiesInputsTask(s)
Internal pediatric pretraining
TCH-Complex13,32317,9843,747,848Self-supervised pretraining [F]
Internal pediatric evaluation
TCH-SHD12,51716,4271,015,795Binary SHD classification [V]
TCH-View29830110,52252-class view classification [F]
TCH-EF2,6373,86413,809EF regression [V]
Table 1: Cohorts used for EchoDino pretraining and downstream evaluation. TCH denotes Texas Children’s Hospital, and all TCH-prefixed cohorts are internal pediatric datasets. TCH-SHD is a subset of TCH-Complex, consisting of 56.8% DICOM files and 57.4% studies. The remaining TCH downstream cohorts contained no studies included in TCH-Complex. Input counts follow the native unit used to define each cohort: pretraining frames for TCH-Complex, retained frames for TCH-Age and CAMUS, and retained videos for the remaining evaluation cohorts. F and V denote frame- and video-level tasks, respectively; EF, ejection fraction; LA, left atrium; LV, left ventricle; SHD, structural heart disease. Detailed cohort definitions and target distributions are provided in Sec. 5.1 .
MethodKNNLinear probe
k=10k=10k=20k=20k=30k=30
EchoPrime0.224 (0.206–0.244) [ p<0.001p<0.001 ]0.229 (0.210–0.248) [ p<0.001p<0.001 ]0.228 (0.210–0.248) [ p<0.001p<0.001 ]0.274 (0.246–0.302) [ p<0.001p<0.001 ]
PanEcho0.504 (0.482–0.528) [ p<0.001p<0.001 ]0.496 (0.475–0.518) [ p<0.001p<0.001 ]0.483 (0.462–0.506) [ p<0.001p<0.001 ]0.609 (0.579–0.640) [ p<0.001p<0.001 ]
DINOv30.369 (0.347–0.389) [ p<0.001p<0.001 ]0.366 (0.345–0.387) [ p<0.001p<0.001 ]0.359 (0.337–0.380) [ p<0.001p<0.001 ]0.464 (0.433–0.495) [ p<0.001p<0.001 ]
EchoDino0.850 (0.834–0.866)0.848 (0.833–0.864)0.848 (0.833–0.865)0.889 (0.868–0.908)
Table 2: View-classification accuracy on the TCH-View dataset. Values are bootstrap medians (95% CIs). P values compare each baseline with EchoDino.
 EchoNet-DynamicEchoNet-LVH
MethodEDV ↓\downarrowESV ↓\downarrowLVIDs ↓\downarrowLVIDd ↓\downarrowLVPWd ↓\downarrowIVSd ↓\downarrow
EchoPrime0.333 (0.314–0.355) [ p<0.001p<0.001 ]0.225 (0.212–0.239) [ p<0.001p<0.001 ]0.505 (0.463–0.552) [ p<0.001p<0.001 ]0.488 (0.445–0.532) [ p<0.001p<0.001 ]0.222 (0.201–0.243) [ p<0.001p<0.001 ]0.299 (0.265–0.335) [ p<0.001p<0.001 ]
PanEcho0.236 (0.224–0.250) [ p<0.001p<0.001 ]0.162 (0.153–0.171) [ p<0.001p<0.001 ]0.369 (0.339–0.401) [ p<0.001p<0.001 ]0.383 (0.349–0.415) [ p<0.001p<0.001 ]0.165 (0.149–0.179) [ p<0.001p<0.001 ]0.204 (0.180–0.230) [ p<0.001p<0.001 ]
DINOv3 (class token)0.286 (0.271–0.304) [ p<0.001p<0.001 ]0.214 (0.203–0.227) [ p<0.001p<0.001 ]0.466 (0.423–0.516) [ p<0.001p<0.001 ]0.482 (0.441–0.526) [ p<0.001p<0.001 ]0.218 (0.197–0.239) [ p<0.001p<0.001 ]0.294 (0.262–0.329) [ p<0.001p<0.001 ]
DINOv3-PATCH0.203 (0.193–0.215) [ p<0.001p<0.001 ]0.112 (0.104–0.119) [ p=0.006p=0.006 ]0.382 (0.347–0.417) [ p<0.001p<0.001 ]0.391 (0.356–0.426) [ p<0.001p<0.001 ]0.168 (0.152–0.183) [ p<0.001p<0.001 ]0.221 (0.196–0.247) [ p<0.001p<0.001 ]
EchoDino (class token)0.250 (0.237–0.264) [ p<0.001p<0.001 ]0.179 (0.170–0.189) [ p<0.001p<0.001 ]0.406 (0.374–0.442) [ p<0.001p<0.001 ]0.421 (0.385–0.456) [ p<0.001p<0.001 ]0.190 (0.173–0.207) [ p<0.001p<0.001 ]0.235 (0.209–0.262) [ p<0.001p<0.001 ]
Table 3: Left ventricular (LV) measurement performance. Values are bootstrap medians (95% CIs), and all metrics are mean absolute errors (MAE; ↓\downarrow ). EDV and ESV are reported in normalized 100-mL units (for example, 0.180 corresponds to 18.0 mL); LVIDs, LVIDd, LVPWd, and IVSd are reported in centimetres. Where reported, P values compare each method with EchoDino-PATCH.
 EchoNet-DynamicEchoNet-PediatricCAMUS
MethodLV ↑\uparrowLV ↑\uparrowLV ↑\uparrowMYO ↑\uparrowLA ↑\uparrow
EchoPrime0.864 (0.861–0.867) [ p<0.001p<0.001 ]0.809 (0.804–0.815) [ p<0.001p<0.001 ]0.907 (0.905–0.909) [ p<0.001p<0.001 ]0.787 (0.785–0.790) [ p<0.001p<0.001 ]0.847 (0.843–0.850) [ p<0.001p<0.001 ]
PanEcho0.887 (0.884–0.889) [ p<0.001p<0.001 ]0.860 (0.855–0.864) [ p<0.001p<0.001 ]0.919 (0.917–0.921) [ p<0.001p<0.001 ]0.840 (0.838–0.842) [ p<0.001p<0.001 ]0.868 (0.864–0.871) [ p<0.001p<0.001 ]
DINOv30.888 (0.886–0.890) [ p<0.001p<0.001 ]0.864 (0.860–0.868) [ p<0.001p<0.001 ]0.917 (0.915–0.918) [ p<0.001p<0.001 ]0.826 (0.824–0.828) [ p<0.001p<0.001 ]0.860 (0.856–0.863) [ p<0.001p<0.001 ]
EchoDino0.910 (0.909–0.912)0.903 (0.900–0.906)0.932 (0.930–0.933)0.869 (0.867–0.871)0.888 (0.885–0.891)
Table 4: Segmentation performance across clinical datasets. Values are bootstrap medians (95% CIs), and all metrics are Dice coefficients ( ↑\uparrow ). P values compare each baseline with EchoDino.
MethodMAE ↓\downarrowMSE ↓\downarrowAcc ↑\uparrowF1 ↑\uparrow
EchoPrime5.133 (4.854–5.428) [ p<0.001p<0.001 ]34.396 (31.585–37.353) [ p<0.001p<0.001 ]0.362 (0.316–0.408) [ p<0.001p<0.001 ]0.198 (0.155–0.243) [ p<0.001p<0.001 ]
PanEcho3.857 (3.624–4.089) [ p<0.001p<0.001 ]20.005 (17.929–21.970) [ p<0.001p<0.001 ]0.530 (0.481–0.578) [ p<0.001p<0.001 ]0.461 (0.411–0.511) [ p<0.001p<0.001 ]
DINOv34.178 (3.938–4.438) [ p<0.001p<0.001 ]23.568 (21.285–26.075) [ p<0.001p<0.001 ]0.397 (0.349–0.446) [ p<0.001p<0.001 ]0.269 (0.215–0.321) [ p<0.001p<0.001 ]
EchoDino (class token)2.747 (2.532–2.971) [ p<0.001p<0.001 ]11.687 (10.278–13.274) [ p<0.001p<0.001 ]0.630 (0.581–0.678) [ p<0.001p<0.001 ]0.626 (0.578–0.676) [ p<0.001p<0.001 ]
EchoDino-PATCH1.389 (1.266–1.524)3.606 (3.051–4.241)0.884 (0.851–0.916)0.884 (0.852–0.916)
Table 5: Age-prediction performance in years. Values are bootstrap medians (95% CIs), and frame-level predictions were aggregated for patient-level evaluation. P values compare each method with EchoDino-PATCH.
MethodAccuracy ↑\uparrowPrecision ↑\uparrowRecall ↑\uparrowF1 score ↑\uparrowAUC ↑\uparrow
EchoPrime0.621 (0.596–0.644) [ p<0.001p<0.001 ]0.612 (0.583–0.639) [ p<0.001p<0.001 ]0.582 (0.560–0.603) [ p<0.001p<0.001 ]0.571 (0.545–0.595) [ p<0.001p<0.001 ]0.644 (0.618–0.669) [ p<0.001p<0.001 ]
PanEcho0.738 (0.714–0.760) [ p<0.001p<0.001 ]0.735 (0.712–0.758) [ p<0.001p<0.001 ]0.721 (0.699–0.743) [ p<0.001p<0.001 ]0.724 (0.701–0.747) [ p<0.001p<0.001 ]0.811 (0.789–0.832) [ p<0.001p<0.001 ]
DINOv30.720 (0.696–0.742) [ p<0.001p<0.001 ]0.722 (0.698–0.746) [ p<0.001p<0.001 ]0.695 (0.673–0.717) [ p<0.001p<0.001 ]0.698 (0.674–0.722) [ p<0.001p<0.001 ]0.776 (0.751–0.798) [ p<0.001p<0.001 ]
EchoDino0.788 (0.765–0.806)0.788 (0.766–0.807)0.773 (0.752–0.793)0.778 (0.755–0.797)0.872 (0.852–0.888)
Table 6: TCH-SHD classification performance. Values are bootstrap medians (95% CIs), and study-level predictions were obtained by averaging logits from all videos associated with each study. P values compare each baseline with EchoDino.
 EchoNet-DynamicEchoNet-PediatricTCH-EF
MethodMAE ↓\downarrowMSE ↓\downarrowMAE ↓\downarrowMSE ↓\downarrowMAE ↓\downarrowMSE ↓\downarrow
EchoPrime5.291 (5.059–5.561) [ p<0.001p<0.001 ]48.433 (43.891–53.973) [ p<0.001p<0.001 ]4.801 (4.529–5.064) [ p=0.112p=0.112 ]40.242 (35.541–45.664) [ p=0.226p=0.226 ]9.112 (8.799–9.404) [ p<0.001p<0.001 ]130.915 (122.732–139.157) [ p<0.001p<0.001 ]
PanEcho5.189 (4.931–5.460) [ p<0.001p<0.001 ]48.443 (42.989–54.830) [ p<0.001p<0.001 ]4.873 (4.538–5.219) [ p=0.038p=0.038 ]50.381 (40.918–62.726) [ p=0.170p=0.170 ]7.404 (7.154–7.661) [ p<0.001p<0.001 ]91.690 (85.663–98.400) [ p<0.001p<0.001 ]
DINOv35.503 (5.210–5.799) [ p<0.001p<0.001 ]55.670 (49.425–62.317) [ p<0.001p<0.001 ]5.417 (5.078–5.799) [ p<0.001p<0.001 ]61.603 (51.558–73.512) [ p<0.001p<0.001 ]7.711 (7.437–7.981) [ p<0.001p<0.001 ]99.624 (92.316–106.838) [ p<0.001p<0.001 ]
EchoDino (Uniform)5.042 (4.776–5.302) [ p<0.001p<0.001 ]47.045 (41.935–52.437) [ p<0.001p<0.001 ]4.800 (4.474–5.151) [ p=0.006p=0.006 ]49.933 (40.817–60.543) [ p=0.036p=0.036 ]6.410 (6.173–6.657) [ p<0.001p<0.001 ]70.234 (64.792–76.656) [ p<0.001p<0.001 ]
EchoDino-MEMS4.674 (4.456–4.894)39.353 (35.239–43.259)4.555 (4.235–4.877)45.197 (36.002–55.891)5.652 (5.444–5.854)56.148 (51.326–61.083)
Table 7: Ejection fraction (%) prediction performance across three clinical datasets. Values are bootstrap medians (95% CIs). P values compare each method with EchoDino-MEMS.
Figure 2: Age-stratified ejection-fraction prediction on the TCH-EF test set. Bootstrap median mean absolute error (MAE, EF percentage points; lower is better) is shown for videos from participants aged <1<1 year ( n=374n=374 ), 11 – <5<5 years ( n=190n=190 ), 55 – <11<11 years ( n=578n=578 ), and ≥11\geq 11 years ( n=938n=938 ). EchoDino using uniform sampling achieved lower MAE than the comparison methods in every age group, and EchoDino-MEMS further reduced MAE across all groups. Error bars indicate percentile 95% confidence intervals. Asterisks denote paired comparisons with EchoDino-MEMS ( ∗∗:P<0.01\ast\ast:P<0.01 , ∗∗∗:P<0.001\ast\ast\ast:P<0.001 ).
MethodAccuracy ↑\uparrowPrecision ↑\uparrowRecall ↑\uparrowF1 score ↑\uparrowAUC ↑\uparrow
EchoPrime0.674 (0.653–0.695) [ p<0.001p<0.001 ]0.712 (0.691–0.730) [ p<0.001p<0.001 ]0.674 (0.653–0.695) [ p<0.001p<0.001 ]0.686 (0.666–0.705) [ p<0.001p<0.001 ]0.839 (0.826–0.852) [ p<0.001p<0.001 ]
PanEcho0.944 (0.934–0.954) [ p<0.001p<0.001 ]0.942 (0.931–0.952) [ p<0.001p<0.001 ]0.944 (0.934–0.954) [ p<0.001p<0.001 ]0.940 (0.928–0.951) [ p<0.001p<0.001 ]0.987 (0.983–0.990) [ p=0.010p=0.010 ]
DINOv30.938 (0.927–0.948) [ p<0.001p<0.001 ]0.939 (0.927–0.949) [ p<0.001p<0.001 ]0.938 (0.927–0.948) [ p<0.001p<0.001 ]0.938 (0.927–0.949) [ p<0.001p<0.001 ]0.991 (0.988–0.993) [ p=0.914p=0.914 ]
EchoDino (Uniform)0.957 (0.948–0.965) [ p=0.316p=0.316 ]0.955 (0.946–0.964) [ p=0.196p=0.196 ]0.957 (0.948–0.965) [ p=0.316p=0.316 ]0.955 (0.946–0.963) [ p=0.174p=0.174 ]0.983 (0.978–0.988) [ p<0.001p<0.001 ]
EchoDino-MEMS0.961 (0.953–0.969)0.960 (0.951–0.969)0.961 (0.953–0.969)0.960 (0.951–0.969)0.991 (0.987–0.994)
Table 8: Sweep-mode classification on the TCH-Sweep dataset. Values are bootstrap medians (95% CIs) of support-weighted metrics. P values compare each method with EchoDino-MEMS.
Standardized clinical nameDisplay codeTrainingValidationTestTotal
Cohort size    
Retained videos, nn8,57796797810,522
Studies, nn2442928301
Patients, nn2422828298
Video-level class distribution, nn (%)
Two-dimensional parasternal short-axis reference view at the aortic-valve level2D-PSAX-AV-REF678 (7.9%)62 (6.4%)86 (8.8%)826 (7.9%)
Supplementary Table 1: Composition, video-level class distribution and nomenclature crosswalk for the TCH-View dataset. Cohort sizes report retained videos, unique studies and unique patients within each partition. Class percentages are calculated within each partition. The 52 retained display codes encode acquisition mode, view family and a reference designation, anatomical level or focus; designated reference acquisitions are identified by the suffix REF. Every retained class occurred in all three patient- and study-disjoint partitions.
CharacteristicTrainingValidationTestTotal
Cohort size
Retained videos, nn810,362103,008102,4251,015,795
Studies, nn13,1141,6761,63716,427
Patients, nn10,0131,2521,25212,517
Video-level label distribution, nn (%)
Normal484,465 (59.8%)62,478 (60.7%)62,709 (61.2%)609,652 (60.0%)
Supplementary Table 2: Composition and binary label distribution of TCH-SHD. Cohort sizes report retained videos, unique studies, and unique patients within each partition. Label counts are reported separately at the video, study, and patient levels, with percentages calculated within each partition and data level. Patients were mutually exclusive across partitions.
CharacteristicTrainingValidationTestTotal
Cohort size
Retained videos, nn9,4412,2882,08013,809
Studies, nn2,6146665843,864
Patients, nn1,7924493962,637
Video-level view distribution, nn (%)
A4C_ref_2D9,441 (100.0%)2,288 (100.0%)2,080 (100.0%)13,809 (100.0%)
Supplementary Table 3: Composition and label distribution of TCH-EF. Cohort sizes report retained videos, unique studies, and unique patients within each partition. Age and ejection fraction (EF) summaries are calculated at the study level. Partitions were generated at the patient level using ten equal-frequency bins of patient-level median EF; no patient or study occurred in more than one partition. IQR denotes interquartile range and SD denotes standard deviation.
CharacteristicTrainingValidationTestTotal
Cohort size
Retained frames, nn81,61910,0599,958101,636
Studies, nn3,0953853853,865
Patients, nn2,1963493702,915
Frame-level view distribution, nn (%)
2D_PSAX26,011 (31.9%)3,255 (32.4%)3,212 (32.3%)32,478 (32.0%)
Supplementary Table 4: Composition of TCH-Age. Cohort sizes report retained frames, unique studies, and unique patients within each partition. View distributions are frame-level counts, with percentages calculated within each partition. Age summaries are calculated at the study level. Studies and patients were mutually exclusive across partitions. IQR denotes interquartile range and SD denotes standard deviation.
CharacteristicTrainingValidationTestTotal
Cohort size
Videos, nn9,3882,0112,01113,410
Studies, nn2195047316
Patients, nn2194747313
Acquisition-mode distribution, nn (%)
Color Doppler4,345 (46.3%)932 (46.3%)932 (46.3%)6,209 (46.3%)
Supplementary Table 5: Composition of TCH-Sweep by data partition. Partitions were constructed at the patient level with stratification by acquisition mode; patients and studies were mutually exclusive across partitions. Acquisition-mode distributions are video-level counts, with percentages calculated within each partition.
DatasetVendor-associated studies
TCH-ComplexGE Vingmed Ultrasound, 16,386; Philips Medical Systems, 1,622; TomTec Imaging Systems GmbH, 700; TOMTEC, 126
TCH-SHDGE Vingmed Ultrasound, 15,007; Philips Medical Systems, 1,437; TomTec Imaging Systems GmbH, 189
TCH-ViewGE Vingmed Ultrasound, 186; Philips Medical Systems, 94; TOSHIBA_MEC_US, 23; SIEMENS, 7; GEMS Ultrasound, 5; GE Healthcare, 1
TCH-EFGE Vingmed Ultrasound, 3,053; Philips Medical Systems, 716; TOSHIBA_MEC_US, 196; SIEMENS, 33; GEMS Ultrasound, 28; GE Healthcare, 4
TCH-AgeGE Vingmed Ultrasound, 3,066; Philips Medical Systems, 719; TOSHIBA_MEC_US, 198; SIEMENS, 34; GEMS Ultrasound, 28; GE Healthcare, 4
TCH-SweepGE Vingmed Ultrasound, 186; Philips Medical Systems, 94; TOSHIBA_MEC_US, 23; SIEMENS, 6; GEMS Ultrasound, 5; GE Healthcare, 1
Supplementary Table 6: Scanner-vendor coverage across the TCH cohorts. Vendor counts denote unique studies associated with each metadata label. Counts need not sum to the cohort total because vendor metadata can be missing and one study can contain acquisitions from more than one system. The two TomTec spellings in TCH-Complex are preserved as separate source labels.
  EchoPrimePanEchoDINOv3EchoDino
Display codennPRF1PRF1PRF1PRF1
2D-A2C-REF240.143 (0.058–0.246) [ p<0.001p<0.001 ]0.333 (0.147–0.533) [ p<0.001p<0.001 ]0.200 (0.085–0.323) [ p<0.001p<0.001 ]0.731 (0.526–0.905) [ p=0.005p=0.005 ]0.667 (0.469–0.852) [ p=0.002p=0.002 ]0.696 (0.519–0.833) [ p<0.001p<0.001 ]0.306 (0.167–0.459) [ p<0.001p<0.001 ]0.500 (0.294–0.706) [ p<0.001p<0.001 ]0.379 (0.218–0.526) [ p<0.001p<0.001 ]1.000 (1.000–1.000) [-]0.962 (0.857–1.000) [-]0.980 (0.923–1.000) [-]
2D-A3C-REF110.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.100 (0.000–0.375) [ p<0.001p<0.001 ]0.083 (0.000–0.308) [ p<0.001p<0.001 ]0.095 (0.000–0.300) [ p<0.001p<0.001 ]0.222 (0.000–0.500) [ p<0.001p<0.001 ]0.267 (0.000–0.571) [ p<0.001p<0.001 ]0.240 (0.000–0.476) [ p<0.001p<0.001 ]1.000 (1.000–1.000) [-]1.000 (1.000–1.000) [-]1.000 (1.000–1.000) [-]
2D-A4C-LH18-0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.333 (0.000–1.000) [ p=0.258p=0.258 ]0.053 (0.000–0.188) [ p<0.001p<0.001 ]0.091 (0.000–0.286) [ p<0.001p<0.001 ]-0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.722 (0.455–0.933) [-]0.556 (0.316–0.789) [-]0.625 (0.400–0.800) [-]
2D-A4C-REF800.342 (0.272–0.415) [ p<0.001p<0.001 ]0.738 (0.638–0.831) [ p=0.048p=0.048 ]0.467 (0.388–0.542) [ p<0.001p<0.001 ]0.525 (0.434–0.615) [ p<0.001p<0.001 ]0.776 (0.679–0.863) [ p=0.149p=0.149 ]0.626 (0.542–0.701) [ p<0.001p<0.001 ]0.519 (0.432–0.605) [ p<0.001p<0.001 ]0.839 (0.753–0.914) [ p=0.712p=0.712 ]0.641 (0.560–0.713) [ p<0.001p<0.001 ]0.833 (0.747–0.908) [-]0.864 (0.783–0.934) [-]0.847 (0.783–0.902) [-]
2D-AP-RVC13-0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.667 (0.000–1.000) [ p=0.934p=0.934 ]0.143 (0.000–0.385) [ p=0.004p=0.004 ]0.235 (0.000–0.526) [ p=0.021p=0.021 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.000 (0.000–0.000) [ p<0.001p<0.001 ]0.647 (0.400–0.875) [-]0.857 (0.615–1.000) [-]0.733 (0.519–0.889) [-]
Supplementary Table 7: Bootstrap analysis of per-class linear-probe view classification on TCH-View. The display-code column uses the standardized codes defined in Supplementary Table 1 , which provides the corresponding clinical names. Each cell reports the median, 95% confidence interval, and bootstrap pp value from 10,000 resamples. EchoDino is the reference model for the reported pp values, and its pp -value fields are marked by [-]. Within each row and metric, the highest median is bolded and the second-highest distinct median is underlined; tied values receive the same formatting. A dash in a precision cell indicates that the model made no positive predictions for that class, such that the precision denominator was zero and the metric was undefined. P, R, and F1 denote precision, recall, and F1 score, respectively, and nn denotes test support.
Supplementary Fig. 1: Patient-level age-interval classification by EchoDino-PATCH. Continuous patient-level age estimates were assigned to the intervals [0,2][0,2] , (2,12](2,12] , and (12,18](12,18] years. Rows indicate chronological-age intervals, columns indicate predicted intervals, and cells show patient counts. Frame-level estimates were aggregated within each patient before interval assignment.
Metadata labelnnEchoPrimePanEchoDINOv3EchoDino
D-looped transposition of the great arteries170.7061.0000.9411.000
Common atrioventricular septal defect10.0001.0000.0001.000
Tricuspid atresia170.8241.0001.0001.000
Truncus arteriosus80.2501.0001.0001.000
Total anomalous pulmonary venous connection50.6000.8000.8000.800
Double inlet left ventricle120.4171.0000.9171.000
Supplementary Table 8: Study-level SHD-positive detection rate stratified by detailed TCH-SHD labels. Video-level logits were averaged within each study before binary SHD_any classification. The value for each label is the fraction of label-positive studies classified as SHD_any; Labels are non-mutually exclusive, and nn denotes the number of studies carrying each label. Best values are bold, and second-best values are underlined.
Supplementary Fig. 2: Predicted versus reference ejection fraction on TCH-EF. Panels compare EchoPrime, PanEcho, DINOv3, EchoDino with uniform sampling, and EchoDino-MEMS in reading order. Dashed lines indicate identity ( y=xy=x ). Panel annotations report mean absolute error (MAE) in EF percentage points.
Supplementary Fig. 3: Effect of temporal-anchor count and motion bias on ejection-fraction estimation. Mean absolute error (MAE; EF percentage points) on EchoNet-Dynamic is shown for a fixed 16-frame budget. CC denotes the number of uniformly spaced temporal anchors, and λ\lambda controls the motion bias applied during diversity sampling. Both anchor settings achieved their lowest MAE at λ=0.6\lambda=0.6 , while C=4C=4 yielded lower error than C=8C=8 at every tested value. The best configuration ( C=4C=4 , λ=0.6\lambda=0.6 ) achieved an MAE of 4.674. Lower values indicate more accurate EF estimates.

Explore similar work

CardsList

  1. Beyond Independent Frames: Latent Attention Masked Autoencoders for Multi-View Echocardiography

    Apr 16, 2026Simon Böhi, Irene Cannistraci, Sergio Muñoz Gonzalez +8Medical Imaging Foundation ModelsMasked Autoencoders

    Department of Biomedical Engineering, University of Basel, Switzerland · Department of Computer Science, ETH Zurich, Switzerland

  2. Evaluating self-supervised echocardiographic representations across downstream extraction strategies for left-ventricular segmentation and ejection fraction estimation

    Jun 22, 2026Sylwia Majchrowska, Philip TeareUltrasound Image SegmentationLeft Ventricular Ejection Fraction Estimation

    aR&D Data ScienceSylwiaSkills &Majchrowskaa,Partnership, DataPhilipScienceTeareb& AI, BioPharma R&D, AstraZeneca, Pepparedsleden 1, Mölndal, 431 83, Sweden · bCentre for AI, Data Science & AI, BioPharma R&D, AstraZeneca, BiomedicalJun Campus, 1 Francis Crick Ave, Trumpington, Cambridge, CB2 0AA, UK

  3. EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function

    May 6, 2026Elias Stenhede, Joanna Sulkowska, Eivind Bjørkan Orstad +2Medical ImagingEchocardiography

    Medical Technology & E-Health, Akershus University Hospital, Norway · Department of Technology Systems, University of Oslo, Norway · Department of Cardiology, Akershus University Hospital, Norway