EchoDiST: Self-distillation-based joint learning for diffusion-conditioned echocardiographic myocardial motion estimation

Feiyue Qi; Xingyue Wei; Jianwen Luo

cs.CVOct 8, 2026

EchoDiST: Self-distillation-based joint learning for diffusion-conditioned echocardiographic myocardial motion estimation

Authors: Feiyue Qi, Xingyue Wei, Jianwen Luo

Organizations: School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, 100084, China

Abstract

Motion estimation in echocardiography is essential for quantitative assessment of cardiac function and myocardial mechanics, but remains challenging due to image artifacts, limited image information, speckle decorrelation, and the scarcity of ground-truth displacement fields. Anatomy-guided approaches can provide structural information, yet often rely on expert-labeled myocardial segmentations. We propose EchoDiST, a framework for unsupervised echocardiographic myocardial motion estimation that integrates self-distillation-based joint learning with a diffusion-conditioned motion estimation network. Here, unsupervised motion estimation refers to learning without ground-truth displacement fields. The self-distillation strategy jointly optimizes anatomical segmentation and myocardial motion estimation under limited anatomical annotations. Diffusion-based conditioning is used during training with stochastic perturbations, while inference requires only a single deterministic forward pass without iterative reverse-diffusion sampling. EchoDiST was evaluated on three echocardiographic datasets, including two external test datasets under cross-view and cross-dataset settings. Compared with seven representative learning-based methods, EchoDiST consistently improved anatomical alignment, myocardial strain assessment, and motion-derived functional and cardiac-phase assessment. These gains were statistically significant across the evaluated tasks and datasets. Overall, EchoDiST provides an effective approach for reliable myocardial motion estimation under limited anatomical supervision and supports downstream quantitative assessment of cardiac function.

Figures & tables

Figure 1: Overview of the proposed EchoDiST framework for echocardiographic motion estimation. (a) Stage 1 pretrains an EchoFM-based segmentation network using a limited set of manually annotated images to generate anatomical pseudo masks. (b) Stage 2 pretrains the diffusion-conditioned motion-estimation network to estimate the displacement field, with anatomical pseudo masks providing structural supervision. (c) Stage 3 jointly optimizes anatomical segmentation and motion estimation through the EMA-based self-distillation strategy.

Figure 2: Architecture of the proposed diffusion-conditioned motion-estimation network. The moving image $I_{m}$ , fixed image $I_{f}$ , and perturbed fixed image $I_{f}^{t}$ form the diffusion-conditioned input $X_{t}$ . The network contains a diffusion stream for noise prediction and a motion-estimation stream for motion estimation, together with coarse-to-fine feedback refinement. The diffusion timestep embedding is injected into the residual encoding blocks, and frozen EchoFM features are fused with the bottleneck feature through cross-attention.

Figure 3: The proposed self-distillation-based joint learning strategy. Student and Teacher segmentation networks are employed, with the Teacher updated using EMA. Reliability-aware self-distillation and bidirectional anatomy–motion consistency are introduced in EchoDiST. Anatomical consistency is imposed in both moving-to-fixed and fixed-to-moving directions, while only the moving-to-fixed branch is shown for clarity. The Student prediction is warped to the fixed frame using the estimated displacement field and constrained by an EMA Teacher-derived anatomical target. A sample-level reliability weight modulates the combined region-level Dice and pixel-wise MSE consistency losses, enabling joint optimization of anatomical segmentations and motion estimation.

Figure 4: Reference myocardial point configuration in the CAMUS simulation dataset, comprising five transmural layers with 36 points distributed along each layer. The enlarged region illustrates the local radial-tangential basis, where the radial direction $\mathbf{e}_{r}$ is defined from the corresponding inner-layer point toward the outer-layer point, and the tangential direction $\mathbf{e}_{t}$ is defined as the in-plane direction orthogonal to $\mathbf{e}_{r}$ .

	CAMUS-4CH				CAMUS-2CH
Method	Dice $\uparrow$	HD95 $\downarrow$	ASSD $\downarrow$	$p$ -value	Dice $\uparrow$	HD95 $\downarrow$	ASSD $\downarrow$	$p$ -value
DdC-AC-DLIR	0.886 $\pm$ 0.033	7.47 $\pm$ 2.72	2.88 $\pm$ 0.95	$<0.001$	0.872 $\pm$ 0.039	8.66 $\pm$ 2.98	3.44 $\pm$ 1.16	$<0.001$
NICE-Trans	0.863 $\pm$ 0.049	9.25 $\pm$ 3.89	3.46 $\pm$ 1.55	$<0.001$	0.848 $\pm$ 0.052	10.43 $\pm$ 4.18	4.07 $\pm$ 1.65	$<0.001$
NICE-Net	0.849 $\pm$ 0.048	9.92 $\pm$ 3.85	3.70 $\pm$ 1.43	$<0.001$	0.839 $\pm$ 0.052	11.05 $\pm$ 4.28	4.22 $\pm$ 1.56	$<0.001$
CorrMLP	0.848 $\pm$ 0.053	10.22 $\pm$ 4.17	3.83 $\pm$ 1.65	$<0.001$	0.835 $\pm$ 0.057	10.96 $\pm$ 4.08	4.37 $\pm$ 1.76	$<0.001$
SDHNet	0.868 $\pm$ 0.047	9.73 $\pm$ 4.40	3.42 $\pm$ 1.28	$<0.001$	0.861 $\pm$ 0.046	10.23 $\pm$ 4.27	3.79 $\pm$ 1.38	$<0.001$

Table 1: Quantitative comparison of anatomical alignment on CAMUS-4CH, CAMUS-2CH, EchoNet-Dynamic, and the CAMUS simulation dataset. All metrics are reported as mean

\pm

standard deviation over 100, 100, 110, and 100 test cases for the four datasets, respectively, with one ED-to-ES image pair evaluated per case.

p

-values indicate paired statistical comparisons of Dice between EchoDiST and each comparison method.

Figure 5: Qualitative comparison of different methods on the CAMUS dataset. The first column shows the ES fixed image and the ED moving image with anatomical contours, and the remaining columns show the results of the different methods. The dataset corresponding to each two-row group is indicated along the left margin. Within each group, the first row presents the warped moving image with the fixed anatomy in green and the warped moving anatomy in red, while the second row presents the corresponding displacement grid. The numbers displayed on the upper-left corner of each warped-image panel indicate the Dice score for the representative ED-to-ES image pair shown.

Figure 6: Qualitative comparison of different methods on EchoNet-Dynamic and the CAMUS simulation dataset. The first column shows the ES fixed image and the ED moving image with anatomical contours, and the remaining columns show the results of the different methods. The dataset corresponding to each two-row group is indicated along the left margin. Within each group, the first row presents the warped moving image with the fixed anatomy in green and the warped moving anatomy in red, while the second row presents the corresponding displacement grid. The numbers displayed on the upper-left corner of each warped-image panel indicate the Dice score for the representative ED-to-ES image pair shown.

Method	EPE (mm) $\downarrow$	Radial RMSE (%) $\downarrow$	Tangential RMSE (%) $\downarrow$	$p$ -value
DdC-AC-DLIR	7.10 $\pm$ 2.31	17.46 $\pm$ 5.52	18.43 $\pm$ 5.83	$<0.001$
NICE-Trans	6.33 $\pm$ 2.45	24.56 $\pm$ 6.11	19.78 $\pm$ 5.65	$<0.001$
NICE-Net	7.31 $\pm$ 2.20	21.64 $\pm$ 5.49	22.61 $\pm$ 5.73	$<0.001$
CorrMLP	6.81 $\pm$ 2.51	27.77 $\pm$ 6.49	21.45 $\pm$ 5.59	$<0.001$
SDHNet	9.04 $\pm$ 2.23	25.61 $\pm$ 6.53	25.02 $\pm$ 6.32	$<0.001$
TransMorph2D	6.74 $\pm$ 2.42	20.40 $\pm$ 5.26	18.67 $\pm$ 5.91	$<0.001$

Table 2: Quantitative comparison of myocardial point motion and local strain estimation on the CAMUS simulation dataset. All metrics are reported as mean

\pm

standard deviation over 100 test cases, with 180 myocardial points evaluated per case.

p

-values indicate paired statistical comparisons of EPE between EchoDiST and each comparison method.

Figure 7: Qualitative comparison of selected representative methods for myocardial point motion and local strain estimation on the CAMUS simulation dataset. The first row shows ED-to-ES myocardial point-motion vectors, with reference and predicted motions shown in green and red, respectively. The second and third rows show the corresponding radial and tangential strain maps. EPE and strain RMSE are computed over the 180 myocardial points of the representative case and displayed on the upper-left corner of the corresponding motion and strain panels, respectively.

	Global Tangential Strain			Global Longitudinal Strain
Method	Curve RMSE (%) $\downarrow$	Pearson correlation $r\uparrow$	$p$ -value	Curve RMSE (%) $\downarrow$	Pearson correlation $r\uparrow$	$p$ -value
DdC-AC-DLIR	6.67 $\pm$ 2.68	0.666 $\pm$ 0.363	$<0.001$	8.74 $\pm$ 3.39	0.647 $\pm$ 0.361	$<0.001$
NICE-Trans	5.57 $\pm$ 2.67	0.663 $\pm$ 0.360	$<0.001$	8.24 $\pm$ 3.42	0.537 $\pm$ 0.422	$<0.001$
NICE-Net	7.10 $\pm$ 2.76	0.605 $\pm$ 0.449	$<0.001$	9.76 $\pm$ 3.50	0.463 $\pm$ 0.511	$<0.001$
CorrMLP	6.74 $\pm$ 2.95	0.441 $\pm$ 0.441	$<0.001$	9.33 $\pm$ 3.55	0.364 $\pm$ 0.479	$<0.001$
SDHNet	10.04 $\pm$ 3.38	$-0.377\pm 0.442$	$<0.001$	12.66 $\pm$ 4.17	$-0.407\pm 0.490$	$<0.001$

Table 3: Quantitative comparison of full-cycle global myocardial strain estimation on the CAMUS simulation dataset. All metrics are reported as mean

\pm

standard deviation over 100 test cases, with global tangential and longitudinal strain curves evaluated over the full cardiac cycle for each case.

p

-values indicate paired statistical comparisons of Curve RMSE between EchoDiST model and each comparison method.

Figure 8: Qualitative comparison of full-cycle global myocardial strain trajectories estimated by different methods on the CAMUS simulation dataset. (a) Global tangential strain and (b) Global longitudinal strain estimated by different methods. The reference trajectory is shown in black, and the ES frame is indicated by a vertical reference line.

Figure 9: Qualitative comparison of normalized left ventricular area trajectories estimated by different methods on EchoNet-Dynamic. Reference LV-area annotations at ED and ES are shown as discrete black markers.

	Area-based Functional Accuracy (Area-EF)			Temporal Consistency
Method	Area-EF RMSE (%) $\downarrow$	Area-EF $r\uparrow$	$p$ -value	ES Localization Error $\downarrow$	Monotonic Consistency $\uparrow$
DdC-AC-DLIR	26.90	0.383	$<0.001$	2.53 $\pm$ 4.34	0.830 $\pm$ 0.211
NICE-Trans	13.41	0.355	0.011	1.17 $\pm$ 0.57	0.846 $\pm$ 0.068
NICE-Net	33.93	0.212	$<0.001$	3.17 $\pm$ 5.54	0.768 $\pm$ 0.193
CorrMLP	18.57	0.472	$<0.001$	1.16 $\pm$ 0.34	0.858 $\pm$ 0.038
SDHNet	38.32	$-0.209$	$<0.001$	6.13 $\pm$ 7.94	0.626 $\pm$ 0.265

Table 4: Quantitative comparison of area-based functional accuracy and cardiac-phase consistency on EchoNet-Dynamic. All metrics are evaluated over 110 test cases. Area-based functional metrics are reported at the cohort level, and temporal consistency metrics as mean

\pm

standard deviation.

p

-values indicate case-level paired statistical comparisons for Area-EF between EchoDiST and each comparison method.

		CAMUS-4CH				CAMUS-2CH
Joint	Feedback	Dice $\uparrow$	HD95 $\downarrow$	ASSD $\downarrow$	$p$ -value	Dice $\uparrow$	HD95 $\downarrow$	ASSD $\downarrow$	$p$ -value
		0.888 $\pm$ 0.032	7.18 $\pm$ 2.23	2.85 $\pm$ 0.89	$<0.001$	0.869 $\pm$ 0.040	8.81 $\pm$ 3.01	3.54 $\pm$ 1.27	$<0.001$
	$\checkmark$	0.908 $\pm$ 0.024	6.10 $\pm$ 2.05	2.31 $\pm$ 0.73	0.479	0.885 $\pm$ 0.030	8.26 $\pm$ 2.79	3.19 $\pm$ 1.01	0.027
$\checkmark$		0.869 $\pm$ 0.041	8.15 $\pm$ 2.86	3.36 $\pm$ 1.22	$<0.001$	0.854 $\pm$ 0.043	9.47 $\pm$ 3.47	4.00 $\pm$ 1.46	$<0.001$
$\checkmark$	$\checkmark$	0.909 $\pm$ 0.022	5.96 $\pm$ 1.80	2.29 $\pm$ 0.68	ref.	0.892 $\pm$ 0.032	7.47 $\pm$ 2.65	2.89 $\pm$ 1.03	ref.
		EchoNet-Dynamic				CAMUS Simulation

Table 5: Quantitative ablation results of anatomical alignment on CAMUS-4CH, CAMUS-2CH, EchoNet-Dynamic, and the CAMUS simulation dataset. All metrics are reported as mean

\pm

standard deviation over 100, 100, 110, and 100 test cases for the four datasets, respectively, with one ED-to-ES image pair evaluated per case.

p

-values indicate paired statistical comparisons of Dice between the full EchoDiST model and each ablated variant.

Figure 10: Qualitative comparison of different model variants on the CAMUS dataset. The first column shows the ES fixed image and the ED moving image with anatomical contours, and the remaining columns show the alignment results of different model variants. The dataset corresponding to each two-row group is indicated along the left margin. Within each group, the first row presents the warped moving image with the fixed anatomy in green and the warped moving anatomy in red, while the second row presents the corresponding displacement grid. The numbers displayed on the upper-left corner of each warped-image panel indicate the Dice score for the representative ED-to-ES image pair shown.

Figure 11: Qualitative comparison of different model variants on EchoNet-Dynamic and the CAMUS simulation dataset. The first column shows the ES fixed image and the ED moving image with anatomical contours, and the remaining columns show the alignment results of different model variants. The dataset corresponding to each two-row group is indicated along the left margin. Within each group, the first row presents the warped moving image with the fixed anatomy in green and the warped moving anatomy in red, while the second row presents the corresponding displacement grid. The numbers displayed on the upper-left corner of each warped-image panel indicate the Dice score for the representative ED-to-ES image pair shown.

Joint	Feedback	EPE (mm) $\downarrow$	Radial RMSE (%) $\downarrow$	Tangential RMSE (%) $\downarrow$	$p$ -value
		8.10 $\pm$ 2.00	19.71 $\pm$ 5.80	21.00 $\pm$ 4.80	$<0.001$
	$\checkmark$	7.83 $\pm$ 2.03	26.21 $\pm$ 5.40	27.16 $\pm$ 5.21	$<0.001$
$\checkmark$		6.23 $\pm$ 1.74	18.82 $\pm$ 4.42	17.69 $\pm$ 3.92	$<0.001$
$\checkmark$	$\checkmark$	4.98 $\pm$ 1.58	16.59 $\pm$ 4.27	16.50 $\pm$ 5.33	ref.

Table 6: Quantitative ablation results of myocardial point motion and local strain estimation on the CAMUS simulation dataset. All metrics are reported as mean

\pm

standard deviation over 100 test cases, with 180 myocardial points evaluated per case.

p

-values indicate paired statistical comparisons of EPE between the full EchoDiST model and each ablated variant.

		Global Tangential Strain			Global Longitudinal Strain
Joint	Feedback	Curve RMSE (%) $\downarrow$	Pearson correlation $r\uparrow$	$p$ -value	Curve RMSE (%) $\downarrow$	Pearson correlation $r\uparrow$	$p$ -value
		8.36 $\pm$ 2.40	0.571 $\pm$ 0.419	$<0.001$	10.69 $\pm$ 2.97	0.605 $\pm$ 0.387	$<0.001$
	$\checkmark$	8.00 $\pm$ 2.73	0.505 $\pm$ 0.481	$<0.001$	10.13 $\pm$ 3.23	0.585 $\pm$ 0.409	$<0.001$
$\checkmark$		5.71 $\pm$ 2.11	0.802 $\pm$ 0.228	$<0.001$	7.27 $\pm$ 2.76	0.829 $\pm$ 0.163	$<0.001$
$\checkmark$	$\checkmark$	2.48 $\pm$ 1.57	0.946 $\pm$ 0.095	ref.	2.98 $\pm$ 1.99	0.956 $\pm$ 0.060	ref.

Table 7: Quantitative ablation results of full-cycle global myocardial strain estimation on the CAMUS simulation dataset. All metrics are reported as mean

\pm

standard deviation over 100 test cases, with global tangential and longitudinal strain curves evaluated over the full cardiac cycle for each case.

p

-values indicate paired statistical comparisons of Curve RMSE between the full EchoDiST model and each ablated variant.

		Area-based Functional Accuracy (Area-EF)			Temporal Consistency
Joint	Feedback	Area-EF RMSE (%) $\downarrow$	Area-EF $r\uparrow$	$p$ -value	ES Localization Error $\downarrow$	Monotonic Consistency $\uparrow$
		19.43	0.620	$<0.001$	1.15 $\pm$ 0.65	0.791 $\pm$ 0.034
	$\checkmark$	17.42	0.683	$<0.001$	1.53 $\pm$ 1.87	0.864 $\pm$ 0.064
$\checkmark$		35.20	0.533	$<0.001$	5.86 $\pm$ 8.29	0.641 $\pm$ 0.271
$\checkmark$	$\checkmark$	10.39	0.686	ref.	1.08 $\pm$ 2.85	0.880 $\pm$ 0.124

Table 8: Quantitative ablation results of area-based functional accuracy and cardiac-phase consistency on EchoNet-Dynamic. All metrics are evaluated over 110 test cases. Area-based functional metrics are reported at the cohort level, and temporal consistency metrics as mean

\pm

standard deviation.

p

-values indicate case-level paired statistical comparisons for Area-EF between the full EchoDiST model and each ablated variant.

Figure 12: Qualitative comparison of different model variants for myocardial point motion and local strain estimation on the CAMUS simulation dataset. The first row shows ED-to-ES myocardial point-motion vectors, with reference and predicted motions shown in green and red, respectively. The second and third rows show the corresponding radial and tangential strain maps. EPE and strain RMSE are computed over the 180 myocardial points of the representative case and displayed on the upper-left corner of the corresponding motion and strain panels, respectively.

Figure 13: Qualitative comparison of full-cycle global myocardial strain trajectories estimated by different model variants on the CAMUS simulation dataset. (a) Global tangential strain and (b) Global longitudinal strain estimated by different model variants. The reference trajectory is shown in black, and the ES frame is indicated by a vertical reference line.

Figure 14: Qualitative comparison of normalized left ventricular area trajectories estimated by different model variants on EchoNet-Dynamic. Reference LV-area annotations at ED and ES are shown as discrete black markers.

Explore similar work

May 12, 2026cs.CV

EchoTracker2: Enhancing Myocardial Point Tracking by Modeling Local Motion

Myocardial point tracking (MPT) has recently emerged as a promising direction for motion estimation in echocardiography, driven by advances in general-purpose point tracking methods. However, myocardial motion fundamentally differs from motion encountered in natural videos, as it arises from physiologically constrained deformation that is spatially and temporally continuous throughout the cardiac cycle. Consequently, motion trajectories typically remain locally confined despite substantial tissue deformation. Motivated by these properties, we revisit the architectural design for MPT and find that coarse initialization in commonly used two-stage coarse-to-fine architectures may be unnecessary in this domain. In this work, we propose a fine-stage-only architecture, \textbf{EchoTracker2}, which enriches pixel-precise features with local spatiotemporal context and integrates them with long-range joint temporal reasoning for robust tracking. Experimental results across in-distribution, out-of-distribution (OOD), and public synthetic datasets show that our model improves position accuracy by

6.5\%

and reduces median trajectory error by

12.2\%

relative to a domain-specific state-of-the-art (SOTA) model. Compared to the best general-purpose point tracking method, the improvements are

2.0\%

and

5.3\%

, respectively. Moreover, EchoTracker2 shows better agreement with expert-derived global longitudinal strain (GLS) and enhances test-rest reproducibility. Source code will be available at: https://github.com/riponazad/ptecho.

Md Abulkalam Azad, Vegard Holmstrøm, John Nyberg +4

Norwegian University of Science and Technology, Trondheim, Norway · Clinic of Cardiology, St. Olavs Hospital, Trondheim, Norway · SINTEF Digital, Trondheim, Norway

Point Tracking Medical Image Analysis Echocardiography Ultrasound Imaging

Ask PDF

May 27, 2026eess.IV

Deep Learning Strain Estimation: Is Physics-Based Simulation the Solution?

Speckle tracking echocardiography (STE) is the clinical standard for myocardial strain estimation. Despite good performance on global strain (GLS), its accuracy for regional strain remains limited, even though this biomarker is highly relevant for early diagnosis and the characterization of subtle abnormalities. from clinical data. Deep learning is a promising alternative, but its development is constrained by the lack of reliable motion references. Existing solutions rely either on STE-derived labels or on simulations generated by physics-based models, but these synthetic sequences still have limited realism compared with clinical data.In this paper, we propose a novel simulation strategy that incorporates speckle decorrelation measures from real videos and uses an iterative refinement process to improve the motion realism in the simulations. We created an open-source photorealistic dataset of 1,478 videos with reference motion, which was used to train an echocardiographic motion estimation algorithm. The proposed method achieves unmatched performance on global and regional strain, notably reaching a GLS variability of 1.42% in an inter-expert setting compared to 1.78% for the clinical reference.

Thierry Judge, Nicolas Duchateau, Andreas Østvik +13

Dept. of Computer Science, University of Sherbrooke, Sherbrooke, Canada · INSA, Universit´e Lyon 1, CNRS UMR 5220, Inserm U1206, CREATIS, Villeurbanne, France · Institut Universitaire de France (IUF) +7

Medical Image Analysis Echocardiography Ultrasound Imaging Medical Image Synthesis Physics-Based Simulation

Ask PDF

Oct 4, 2026cs.CV

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

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.

Sheng Cheng, Donnchadh M. O'Sullivan, Daniel J. Penny +3

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

Medical Imaging Foundation Models Self-Supervised Learning Echocardiography Medical Image Representation Learning

Ask PDF