DyL-UNet for Echocardiographic Video Segmentation

• The paper introduces DyL-UNet, a novel framework that integrates dynamic learning via an Echo-Dynamics Graph to enhance temporal consistency in echocardiographic segmentation.
• DyL-UNet combines a Swin-Transformer encoder-decoder with a Cardiac Phase-Dynamics Attention module to incorporate both spatial details and cardiac phase information.
• Empirical results on CAMUS and EchoNet-Dynamic show improved Dice scores and a significant reduction in Temporal Consistency of Dice (TCD), underscoring its robustness against motion artifacts.

Searching arXiv for DyL-UNet and closely related U-Net variants to ground the article in current papers. arxiv_search(query="DyL-UNet echocardiographic segmentation temporal consistency", max_results=10) DyL-UNet is a dynamic learning-based temporal consistency U-Net segmentation architecture for echocardiographic video analysis. It was proposed to address a specific failure mode of many cardiac ultrasound segmentation systems: even when single-frame masks are accurate, predicted contours can fluctuate from frame to frame because of speckle noise, deformation, weak boundaries, and complex cardiac motion. The framework combines an Echo-Dynamics Graph (EDG) derived from sequence-level dynamic learning, a multi-branch Swin-Transformer encoder-decoder for framewise spatial segmentation, and Cardiac Phase-Dynamics Attention (CPDA) modules on skip connections so that segmentation is conditioned on both cardiac phase and learned motion dynamics. On CAMUS and EchoNet-Dynamic, it is reported to maintain segmentation accuracy comparable to existing methods while achieving superior temporal consistency (Qu et al., 23 Sep 2025).

1. Name, scope, and disambiguation

The official model name is DyL-UNet, introduced in “A DyL-Unet framework based on dynamic learning for Temporally Consistent Echocardiographic Segmentation” (Qu et al., 23 Sep 2025). The architecture is specific to echocardiographic video segmentation and should not be conflated with several similarly named U-Net variants in other domains.

Model	Domain	Relation to DyL-UNet
DyL-UNet (Qu et al., 23 Sep 2025)	Echocardiographic video segmentation	Official model name
Dynamic U-Net (Yang et al., 2024)	Abdominal multi-organ CT segmentation	Different model; uses DCC, DCD, and DCU
DLUNet (Lai et al., 2022)	Semi-supervised abdominal CT segmentation	Different model; “Dual-Light UNet”
DDUNet (Li et al., 26 Jan 2025)	Ground-based cloud segmentation	Different model; “Dual Dynamic U-Net”
Dilated-UNet (Saadati et al., 2023)	2D medical image segmentation	Different model; dilation-based transformer U-Net

This distinction matters because the term “dynamic” is used in materially different ways across these papers. In DyL-UNet, dynamic learning refers to explicit modeling of cardiac sequence dynamics through EDG and the injection of those dynamics into skip-connection attention. In Dynamic U-Net, the term refers to adaptive calibration of convolution, downsampling, and upsampling operators for abdominal CT (Yang et al., 2024). In DLUNet, “dual-light” denotes a semi-supervised two-network training setup with one lightweight UNet used at inference (Lai et al., 2022). In DDUNet, “dual dynamic” refers to two dynamic mechanisms inside a lightweight cloud-segmentation U-Net (Li et al., 26 Jan 2025). In Dilated-UNet, the central mechanism is dilated neighborhood attention rather than temporal dynamics (Saadati et al., 2023).

2. Clinical problem and motivation

DyL-UNet targets the problem of temporally stable and precise echocardiographic segmentation. The motivating observation is that segmentation quality in ultrasound cannot be assessed solely frame by frame. Frame-to-frame jitter weakens clinical interpretability and can destabilize downstream functional measurements such as chamber volume curves or ejection fraction estimates. The paper explicitly attributes this instability to ultrasound speckle noise, deformation, weak boundaries, and complex cardiac motion (Qu et al., 23 Sep 2025).

The method is positioned against three broad classes of prior echocardiographic segmentation approaches. First, single-frame methods focus on end-diastolic (ED), end-systolic (ES), or otherwise independent frames and therefore ignore temporal continuity. Second, some video methods still infer each frame separately and consequently do not impose meaningful temporal coupling. Third, temporal methods based on recurrent units, local memory, or $2D+t/3D+t$ convolutions capture temporal dependence only locally or over fixed windows. The authors argue that such designs are limited when cardiac motion is nonlinear, irregular, or influenced by arrhythmia. DyL-UNet instead introduces a separate dynamic-learning mechanism inspired by deterministic learning for nonlinear systems, with the aim of encoding intrinsic cardiac-cycle dynamics and feeding them directly into segmentation (Qu et al., 23 Sep 2025).

A plausible implication is that DyL-UNet should be understood less as a generic temporal smoothing model and more as a dynamics-conditioned segmentation framework. Its central claim is not merely that adjacent frames should look similar, but that segmentation should remain consistent with a learned representation of cardiac-cycle behavior.

3. Network organization and data flow

The input to DyL-UNet is an echocardiographic sequence

$$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$

The architecture then separates into two interacting streams. One stream performs dynamic learning over the full sequence and produces an EDG representation together with a low-dimensional dynamic descriptor $P_{EDG}$. The other stream performs framewise spatial encoding and decoding using a Swin-Transformer-based multi-branch encoder-decoder. These streams meet at the skip connections, where CPDA injects temporal dynamics and phase cues into the spatial features before decoder fusion (Qu et al., 23 Sep 2025).

The spatial backbone follows a hierarchical Swin-Unet-style design. Each frame $I_t$ is independently passed through a Swin-Transformer encoder, producing multi-scale features

$\{X_{Spatial}^{(t,l)}\}_{l=1}^{4},$

with four encoder scales. Downsampling is implemented through Patch Merging, while the symmetric decoder uses Patch Expanding for upsampling. The model therefore does not begin with a heavy joint spatiotemporal 3D encoder. Instead, it preserves a strong 2D spatial segmentation pathway and introduces temporal reasoning primarily where encoder and decoder information are fused.

For each scale $l$, the framewise spatial features are temporally stacked into $X_{Spatial}^{(l)}$. These temporally organized skip features are then modulated by CPDA using cardiac phase information $\varphi$ and the dynamic descriptor $P_{EDG}$. The decoder therefore receives skip information that is no longer purely spatial: it is spatial detail already conditioned on sequence-level motion structure and phase position within the cardiac cycle. This design differentiates DyL-UNet from recurrent and fixed-window temporal models, because temporal information is introduced as a structured prior on multi-scale feature fusion rather than as a hidden state propagated through the network (Qu et al., 23 Sep 2025).

4. Dynamic learning, EDG, and Cardiac Phase-Dynamics Attention

The most distinctive component of DyL-UNet is the Echo-Dynamics Graph. The paper adapts deterministic learning ideas, previously used for nonlinear dynamic systems and ECG analysis, to echocardiographic sequences. Consecutive frames are related by optical flow,

$Flow_{t\rightarrow t+1},$

and a polar pooling strategy is used to obtain physiology-aware motion descriptors. Taking the image center as the pole, the image is divided into $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$0 annular sectors. For each sector $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$1, the raw descriptor $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$2 includes radial optical-flow component, tangential optical-flow component, grayscale information, and statistical features of these quantities. These sectorwise descriptors are standardized and reduced by PCA:

$$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$3

The resulting sequence $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$4 is modeled as a nonlinear dynamic system,

$$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$5

with temporal increments approximated by an RBF neural network:

$$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$6

Here, $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$7 are RBF centers obtained by K-means clustering, $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$8 is a Gaussian kernel, and $$\{I_t\}_{t=1}^{T}\in \mathbb{R}^{T\times H\times W\times C}.$$9 are learned weights. After fitting the local dynamics, the model computes a residual-weighted basis response

$P_{EDG}$0

which acts as a dynamic energy map. This energy representation is remapped onto the sectorized image domain to form the EDG, and a second dimensionality reduction yields the low-dimensional feature $P_{EDG}$1 (Qu et al., 23 Sep 2025).

In operational terms, the “graph” in EDG is not a conventional GNN graph with an explicit adjacency matrix. The paper does not provide node-edge equations or an adjacency tensor $P_{EDG}$2. Instead, EDG is better described as a structured dynamic map derived from sectorwise transition behavior. The text states that EDG captures physiologically meaningful spatiotemporal patterns, with localized inward-motion hotspots in early systole, larger arc-like high-energy bands in mid-systole, and diminished dynamic energy at ES as cavity volume reaches a minimum.

The Cardiac Phase-Dynamics Attention module is inserted on skip connections. For each spatial feature tensor $P_{EDG}$3, adaptive average pooling first yields a compact feature $P_{EDG}$4. Cardiac phase $P_{EDG}$5 and the dynamic feature $P_{EDG}$6, described as “derived from ED/ES frames via linear estimation,” are each encoded by MLPs to produce $P_{EDG}$7 and $P_{EDG}$8. These are concatenated into a fused temporal token $P_{EDG}$9, and temporal dependencies are modeled through multi-head self-attention:

$I_t$0

A linear projection followed by sigmoid generates the channel modulation factor $I_t$1, which modulates the spatial features by

$I_t$2

The final enhanced feature is

$I_t$3

This formulation shows that CPDA is not simply a temporal attention layer over appearance tokens. It combines pooled spatial features, cardiac phase, and the EDG-derived dynamic descriptor, then uses attention and a residual-style modulation rule to recalibrate skip features. The final $I_t$4 blend with $I_t$5 introduces local spatiotemporal refinement after channelwise modulation. The paper does not provide an explicit transformer $I_t$6 derivation, the numerical value of $I_t$7, or a full joint loss function, so DyL-UNet’s temporal consistency is best understood as being enforced architecturally rather than through a separately formalized temporal regularizer (Qu et al., 23 Sep 2025).

5. Datasets, metrics, and empirical findings

DyL-UNet is evaluated on CAMUS and EchoNet-Dynamic. CAMUS contains 500 patients with 2D echocardiography in the apical four-chamber (4CH) view, and segmentation masks are available for the left ventricle (LV), left ventricular myocardium (LVM), and left atrium (LA). EchoNet-Dynamic contains 10,030 apical four-chamber videos, but left ventricular area annotation is available only on ED and ES frames. For EchoNet-Dynamic, evaluation is therefore limited to ED/ES frames (Qu et al., 23 Sep 2025).

The paper reports three metrics. Dice Similarity Coefficient (Dice) measures segmentation overlap, with higher values better. Hausdorff Distance-95\% (HD95) measures a robust boundary distance, with lower values better. Temporal Consistency of Dice (TCD) is used as the principal temporal stability metric, and lower values indicate smoother temporal continuity and less interframe segmentation jitter. The manuscript does not provide the explicit TCD formula, but it treats TCD as the operational measure of frame-to-frame instability.

On CAMUS, DyL-UNet reports LV Dice 94.93, LVM Dice 88.55, LA Dice 93.04, Average Dice 92.17, HD95 3.90, and Average TCD 0.0062. The comparison table includes PKEcho-Net, DSA, BeU$I_t$8-Net, Echo-ODE, and NCM-Net. DyL-UNet achieves the best average Dice on CAMUS, but it does not achieve the best HD95; NCM-Net is better on that metric with HD95 3.27. Its clearest empirical advantage is temporal consistency, where TCD 0.0062 is the best reported value in the table. The paper states that this corresponds to about a 10.1\% TCD reduction relative to SOTA methods (Qu et al., 23 Sep 2025).

On EchoNet-Dynamic, the reported results are Dice 92.94 and HD95 3.59. These are competitive but not uniformly best. For comparison, the table reports DSA at Dice 93.40, HD95 3.32 and NCM-Net at Dice 93.18, HD95 2.66. The manuscript interprets DyL-UNet’s EchoNet-Dynamic performance as evidence of good generalization and a balance between segmentation accuracy and temporal stability, but temporal consistency is not reported on this dataset because evaluation is only on ED/ES frames (Qu et al., 23 Sep 2025).

The ablation study isolates the impact of dynamic features and phase information. The Baseline (w/o phase + EDG) yields Dice 88.26, HD95 5.84, and TCD 0.0096. DyL-UNet (w/o phase) improves to Dice 90.87, HD95 4.42, and TCD 0.0078. The full DyL-UNet reaches Dice 92.17, HD95 3.90, and TCD 0.0062. Adding EDG alone therefore yields +2.61 Dice, lower HD95, and improved TCD, while adding phase information on top of EDG yields a further +1.30 Dice and reduces TCD from 0.0078 to 0.0062. Relative to the baseline, the full model improves Dice by 3.91 points and TCD by 35.4\% according to the paper. These findings support the paper’s claim that dynamic features improve robustness to motion artifacts and noise, and that the combination of dynamics with phase information gives the strongest temporal stability (Qu et al., 23 Sep 2025).

6. Methodological significance, limitations, and relation to adjacent U-Net variants

DyL-UNet is methodologically distinctive because it does not treat temporal consistency as a post hoc smoothing problem. Instead, it injects a learned dynamics prior into the segmentation backbone. Temporal motion is summarized by EDG through optical flow and nonlinear dynamic modeling; temporal phase is encoded separately as $I_t$9; and temporal dependencies among frame-level spatial features are handled within CPDA by self-attention and a subsequent 3D convolution. This suggests a hybrid temporal modeling strategy: global sequence behavior is represented through EDG, physiological ordering is represented through phase, and local temporal refinement is represented through skip-level attention and convolution (Qu et al., 23 Sep 2025).

At the same time, the paper leaves several implementation and theoretical details underspecified. It does not provide an explicit segmentation loss, temporal consistency loss, or joint objective such as

$\{X_{Spatial}^{(t,l)}\}_{l=1}^{4},$0

It also omits most training hyperparameters, including optimizer type, learning rate, batch size, epoch count, and hardware, and it does not report parameter count, FLOPs, runtime, or GPU memory. The architecture is described at the level of four encoder scales, hierarchical Swin encoding, Patch Merging, Patch Expanding, CPDA, adaptive average pooling, MLP encoders for phase and EDG, multi-head self-attention, and 3D convolution, but it does not provide the number of attention heads, embedding dimensions, window sizes, or Swin stage depths. These omissions are significant for reproducibility and for assessing computational cost.

Several limitations are explicit in the evaluation scope. The paper does not provide a formal graph adjacency formulation despite the term “Echo-Dynamics Graph.” Temporal consistency on EchoNet-Dynamic cannot be evaluated as fully as on CAMUS because only ED and ES frames are annotated. The manuscript also notes that future work is needed for 3D and multi-view ultrasound. Consequently, DyL-UNet should be regarded as a specialized framework for 2D apical four-chamber echocardiographic video segmentation, rather than a general-purpose temporal U-Net family.

Within the broader U-Net literature, DyL-UNet occupies a different methodological niche from other recent variants. Dynamic U-Net for abdominal CT re-engineers convolution, downsampling, and skip-fusion alignment with DCC, DCD, and DCU (Yang et al., 2024). DLUNet focuses on semi-supervised learning with two lightweight UNets during training and one at inference (Lai et al., 2022). DDUNet uses dynamic multi-scale convolution and dynamic generation of classifier weights for lightweight cloud segmentation (Li et al., 26 Jan 2025). Dilated-UNet replaces convolution-centric processing with dilated neighborhood attention in a transformer U-Net for 2D medical segmentation (Saadati et al., 2023). DyL-UNet differs from all of these in centering the segmentation problem on cardiac dynamics and temporal consistency, rather than on multi-organ CT calibration, semi-supervision, cloud masking efficiency, or dilation-based token attention.