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

Motivation

• Artifacts, limited image information, and speckle decorrelation make echocardiographic motion estimation difficult, particularly during rapid or rotational motion.Clinical scarcity of ground-truth displacement fields also limits fully supervised estimation.Source: §1 · p. 1
• Anatomy-guided approaches can improve motion estimation, but often depend on expert-labeled segmentations that are difficult and labor-intensive to obtain at scale.Inaccurate segmentation may also lead to unreliable motion estimation.Source: §1 · p. 1

Approach

• EchoDiST jointly optimizes anatomical segmentation and unsupervised myocardial motion estimation through self-distillation under limited anatomical annotations.Diffusion-based conditioning is integrated into its motion-estimation network.Source: §1 · pp. 1–2

Evaluation

• EchoDiST consistently improves motion estimation across anatomical alignment, myocardial strain, and motion-derived functional and cardiac-phase assessment on three echocardiographic datasets, including two external test datasets.The joint-learning strategy reduces reliance on extensive manual segmentation annotations.Source: §1 · p. 2

2.1. Echocardiographic Motion Estimation Methods

• Conventional tracking can be computationally intensive and rely on prior knowledge or hand-crafted features, while learned unsupervised approaches may lack explicit anatomical supervision.Deep learning methods also include supervised approaches using synthetic reference displacement fields and unsupervised approaches using inter-frame spatiotemporal information.Source: §2.1 · p. 2
• Manual segmentation can guide motion estimation, and segmentation and motion can be complementary when jointly optimized, but scaling this approach requires costly expert annotations.EchoDiST learns segmentations from a limited labeled subset and uses predictions to guide motion estimation over more training pairs.Source: §2.1 · p. 2

2.2. Diffusion-Based Motion Estimation Methods

• Prior medical-image diffusion methods use diffusion-derived features or scores, or progressively denoise displacement fields, but diffusion-conditioned myocardial motion estimation remains underexplored.EchoDiST applies diffusion-conditioned learning to echocardiographic motion estimation while retaining deterministic single-pass inference.Source: §2.2 · pp. 2–3

2.3. Joint Learning of Segmentation and Registration with Limited Annotations

• Limited-label registration research extends anatomical supervision through predicted segmentations, few-shot reverse teaching, self-training, or foundation-model masks.EchoDiST uses EMA-based self-distillation to jointly learn anatomical segmentation and unsupervised motion estimation with limited annotations.Source: §2.3 · p. 3

3.3. Self-distillation-based Joint Learning

• The final stage jointly optimizes segmentation and motion using EMA-based self-distillation and reliability-weighted bidirectional anatomical consistency.Teacher-derived targets are reliability-weighted, and the consistency loss can update motion because the displacement field remains attached during warping.Source: §3 · p. 3, §3.3 · pp. 6–7

3.1. Segmentation Pretraining

• Stage 1 pretrains an EchoFM-based segmentation network on limited manual annotations, then uses its predictions as pseudo masks and initializes the Stage 3 Student.The segmentation network predicts background, LV cavity, myocardium, and LA, with auxiliary supervision and a boundary branch.Source: §3.1 · p. 3, Figure 1 · p. 4

3.2.1. Diffusion-Conditioned Motion Estimation Network

• The motion network takes moving, fixed, and diffusion-perturbed fixed images, then uses separate noise-prediction and motion-estimation streams to predict displacement.Frozen EchoFM features are fused with bottleneck features through cross-attention; diffusion perturbations are used during training, while inference uses one deterministic forward pass.Source: §3.2.1 · p. 4, Figure 2 · p. 5

3.2.2. Feedback Refinement

• At each decoder scale, coarse displacement and residual-misalignment feedback guide a preliminary flow, which is refined using warped features and propagated to finer scales.The feedback map represents the average residual-misalignment response at its scale.Source: §3.2.2 · p. 5

3.2.3. Motion-Estimation Objective

• The motion-estimation objective combines noise prediction, image similarity, displacement regularization, and anatomical correspondence between warped and fixed pseudo masks.Predicted noise also provides a spatial weighting signal for local image similarity.Source: §3.2.3 · p. 5

4.1. Datasets and Preprocessing

• CAMUS-4CH was used for development and in-domain evaluation, while CAMUS-2CH, the CAMUS simulation dataset, and EchoNet-Dynamic supported cross-view or external cross-dataset evaluation without fine-tuning.Images were resized to 256 × 256 pixels; motion estimation generally ran from ED to ES, with ED as the reference for full-cycle analyses.Source: §4.1 · p. 7
• CAMUS-4CH provided 350 training, 50 validation, and 100 held-out test subjects; 50 training subjects had manual annotations, and CAMUS-2CH supplied 100 cross-view test subjects.All 350 CAMUS-4CH training subjects were used for motion-estimation training.Source: §4.1 · p. 7
• EchoNet-Dynamic provided a fixed 110-case subset with complete videos and expert LV tracings, used to form cavity masks and identify ED and ES frames.The subset was defined independently of model predictions and downstream evaluation metrics.Source: §4.1 · p. 7
• The CAMUS simulation dataset contains 100 simulated apical 4CH sequences with reference motion derived from meshes of 180 myocardial points across five layers.Each layer has 36 circumferentially distributed points; corresponding positions across layers share a circumferential index.Source: §4.1 · p. 7, Figure 4 · p. 7

4.2. Implementation Details

• Training followed three stages: segmentation pretraining, motion-estimation optimization, and joint Student segmentation optimization with Teacher updates by EMA.The segmentation pretraining and motion-estimation stages each ran for 100 epochs; the Teacher used an EMA decay factor of 0.995 in Stage 3.Source: §4.2 · p. 7
• The EchoFM encoder in the motion-estimation network remained frozen during Stages 2 and 3.Stage 2 used 1000 forward-diffusion timesteps with a linear variance schedule from 1 × 10−6 to 1 × 10−2.Source: §4.2 · pp. 7–8
• At inference, motion estimation required a single deterministic forward pass without reverse-diffusion sampling.Training and evaluation ran on an RTX A6000-equipped server, with network code implemented in PyTorch.Source: §4.2 · pp. 7–8

4.3.1. Anatomical Alignment

• Anatomical alignment was evaluated with Dice, HD95, and ASSD; myocardial point-motion accuracy on the simulation dataset was measured by EPE in millimeters.CAMUS metrics averaged over LV cavity, myocardium, and LA; simulation and EchoNet-Dynamic metrics covered myocardium and LV cavity, respectively.Source: §4.3.1 · p. 8

4.3.2. Myocardial Strain Assessment

• Local radial and tangential strain errors were measured by RMSE at 180 myocardial points, while full-cycle global tangential and longitudinal strain trajectories were assessed with curve RMSE and Pearson correlation.Local strain used the ED-defined radial-tangential coordinate system; the shear component was excluded from reported metrics.Source: §4.3.2 · p. 8

4.3.3. Motion-Derived Functional and Cardiac-Phase Assessment

• EchoNet-Dynamic functional assessment compared motion-derived Area-EF with reference Area-EF, while cardiac-phase assessment used ES localization error and LV-area monotonic consistency.Predicted Area-EF came from the ED cavity and its motion-propagated ES area; ED and ES were identified as maximum- and minimum-area frames.Source: §4.3.3 · p. 8

4.3.4. Statistical Analysis

• Continuous paired endpoints were tested with two-sided Wilcoxon signed-rank tests and Holm correction; dependent EchoNet-Dynamic correlations were compared by paired case-level bootstrap.Adjusted p-values below 0.05 were considered statistically significant.Source: §4.3.4 · p. 8

4.4.1. Comparison Methods

• The comparison included DdC-AC-DLIR, NICE-Trans, NICE-Net, CorrMLP, SDHNet, TransMorph2D, and VoxelMorph; retrained methods used the CAMUS-4CH training data.Their original method-specific objectives and optimization settings were retained.Source: §4.4.1 · pp. 8–9

4.4.2. Anatomical Alignment and Generalization

• EchoDiST achieved the highest Dice and lowest HD95 and ASSD across all four evaluation settings, significantly outperforming every comparison method on all three metrics (all Holm-adjusted p < 0.001).The settings were CAMUS-4CH, CAMUS-2CH, EchoNet-Dynamic, and the CAMUS simulation dataset.Source: §4.4.2 · p. 9, Table 1 · p. 9

4.4.3. Myocardial Motion and Strain Assessment

• EchoDiST achieved the lowest point-motion, radial-strain, and tangential-strain errors on the CAMUS simulation dataset, with significant improvements in EPE (p < 0.001) and tangential strain (p < 0.05).Qualitative comparisons also showed closer agreement with reference motion direction, magnitude, and strain-map patterns.Source: §4.4.3 · p. 9, Table 2 · p. 11, Figure 7 · p. 11
• EchoDiST achieved the lowest curve RMSE and highest Pearson correlation for both global tangential and longitudinal strain, significantly outperforming all comparison methods (p < 0.001).Its estimated trajectories closely followed the reference curves in magnitude and cardiac-phase evolution.Source: §4.4.3 · pp. 9–10, Table 3 · p. 11, Figure 8 · p. 12

4.4.2. Anatomical Alignment and Generalization

• Qualitative alignment figures compare warped images and displacement grids, with Dice values shown for representative ED-to-ES pairs.Figure 5 covers CAMUS; Figure 6 covers EchoNet-Dynamic and the CAMUS simulation dataset.Source: §4.4.2 · p. 9, Figure 5 · p. 10, Figure 6 · p. 10

4.4.4. Motion-Derived Functional and Cardiac-Phase Assessment

• On EchoNet-Dynamic, EchoDiST achieved the lowest Area-EF RMSE and highest reference correlation, with Area-EF RMSE significantly lower than for all comparison methods (all Holm-adjusted p < 0.05).It also had the lowest ES localization error and highest monotonic consistency; its LV-area reduction was closer to the reference trajectory.Source: §4.4.4 · pp. 10, 12, Table 4 · p. 12

Ablation design

• The ablation study selectively enabled or disabled joint anatomy-motion estimation and feedback refinement while holding other components and training settings unchanged.The four configurations were w/o Joint & Feedback, w/o Joint, w/o Feedback, and Full.Source: §4.5 · pp. 12–13

Anatomical alignment

• The Full model significantly outperformed w/o Joint & Feedback and w/o Feedback on all three anatomical-alignment metrics across all four datasets.Feedback refinement alone recovered much of the alignment performance, particularly on CAMUS-4CH, while joint learning alone had less uniform effects; Table 5 reports alignment results.Source: §4.5 · p. 13, Table 5 · p. 13

Myocardial motion and strain

• Combining both components yielded the lowest EPE and local-strain errors, lowest curve RMSE, and highest correlation for global tangential and longitudinal strain.The Full model significantly outperformed all three ablation variants on both global strain measures (all Holm-adjusted p<0.001); the qualitative comparisons include point motion, local strain maps, and full-cycle trajectories.Source: §4.5 · pp. 13–14, Table 6 · p. 13, Table 7 · p. 13

Functional and cardiac-phase assessment

• On EchoNet-Dynamic, enabling both components produced the strongest overall functional and cardiac-phase performance and significantly improved Area-EF RMSE over every ablation variant.The LV-area trajectory was smoother and more consistent from ED to ES, with fewer frame-to-frame fluctuations; Table 8 reports functional and cardiac-phase metrics.Source: §4.5 · p. 14, Table 8 · p. 14

Overall ablation findings

• Combining joint anatomy-motion estimation and feedback refinement provided the most consistent performance across the evaluated tasks.The Full model significantly outperformed selected ablation variants on anatomical alignment and all three global-strain measures; it also achieved the strongest overall functional and cardiac-phase performance.Source: §4.5 · pp. 13–14

Main findings

• EchoDiST’s reported gains extend across echocardiographic views and external datasets, from anatomical alignment to myocardial motion, mechanical assessment, and motion-derived functional and cardiac-phase assessment.The discussion identifies these as the study’s main findings.Source: §5 · pp. 14–15

Joint anatomy-motion learning

• Joint learning uses anatomy to complement intensity-based correspondence, while propagated segmentations and target-anatomy agreement reciprocally inform motion estimation.The authors suggest this interaction may reduce local motion inconsistencies that affect strain assessment more than segmentation accuracy; feedback refinement corrects residual mismatches during coarse-to-fine estimation.Source: §5 · p. 15

Reference-data limitations

• CAMUS simulation trajectories support controlled motion and strain evaluation but are not physiological ground truth, and EchoNet-Dynamic lacks intermediate-frame annotations for direct motion validation.The discussion also notes that the simulation’s reverberation model does not capture the full complexity of clinical echocardiographic artifacts.Source: §5 · p. 15

Scope and computational cost

• Model development was restricted to CAMUS-4CH, and broader generalization across institutions, scanners, acquisition protocols, image quality, and patient populations remains unestablished.The multi-stage training and diffusion-conditioned estimation increase development cost; inference uses a single deterministic forward pass.Source: §5 · p. 16

Motion-estimation scope

• The current framework estimates two-dimensional in-plane motion and cannot capture out-of-plane myocardial displacement.The authors identify larger, more heterogeneous cohorts and multi-view or three-dimensional echocardiography as future validation directions.Source: §5 · p. 16

Framework

• EchoDiST combines a diffusion-conditioned motion-estimation network using diffusion perturbations as auxiliary conditioning with EMA-based self-distillation for joint segmentation and motion learning.The joint learning uses limited annotations without extensive manual labeling.Source: §6 · p. 17

Learning strategy

• The self-distillation strategy jointly optimizes anatomical segmentation and motion estimation using limited annotations, avoiding the need for extensive manual labeling.Source: §6 · p. 17

Experimental conclusion

• Experiments across echocardiographic views and datasets showed consistent improvements in myocardial motion estimation and mechanical, functional, and cardiac-phase assessment.The authors state these findings support EchoDiST’s potential for quantitative analysis of cardiac motion and function from echocardiographic images.Source: §6 · p. 17

Feiyue Qi

• Feiyue Qi contributed to writing, visualization, software, methodology, formal analysis, data curation, and conceptualization.Source: CRediT authorship contribution statement · p. 17

Xingyue Wei

• Xingyue Wei contributed to writing, methodology, data curation, investigation, and formal analysis.Source: CRediT authorship contribution statement · p. 17

Jianwen Luo

• Jianwen Luo contributed to writing, funding acquisition, methodology, supervision, and resources.Source: CRediT authorship contribution statement · p. 17

• CAMUS, EchoNet-Dynamic, and the CAMUS simulation dataset are publicly available from their respective repositories and were used under their data-access requirements.Source: Data Availability · p. 17
• No new patient data were collected for the study.Source: Data Availability · p. 17
• EchoDiST source code was not publicly available at submission and is to be released upon publication.Source: Data Availability · p. 17