WorksComputer Science
Drift-Conditioned Structural Replay for Continual Cardiac Ultrasound Segmentation
Abstract
Conventional experience replay typically applies fixed preservation constraints to historical exemplars, overlooking variations in how learned representations, predictions, and anatomical structures change during sequential training. To address this limitation, we propose Drift-Conditioned Structural Replay Loss (DCSRL), an exemplar-adaptive supervision mechanism that regulates anatomical preservation according to source-referenced knowledge drift. DCSRL operates within a reliability-guided replay framework that selects informative historical exemplars using anatomical content and boundary characteristics while adaptively regulating replay during training. For each retained exemplar, DCSRL constructs a frozen source-task reference signature and subsequently measures changes in feature representations, class probabilities, boundary responses, and predicted anatomical structures. These complementary drift measurements are aggregated and normalized to generate bounded, exemplar-specific coefficients that dynamically adjust boundary and structural replay losses without modifying the underlying memory-selection policy. We evaluate DCSRL under sequential learning across CAMUS, CardiacNet, and EchoNet-Dynamic, comparing its performance with regularization-, distillation-, and replay-based continual-learning approaches. Experimental findings indicate that DCSRL improves overall final segmentation performance and anatomical boundary preservation, particularly under heterogeneous imaging conditions, although trade-offs remain in knowledge transfer across tasks. These findings suggest that source-referenced multidimensional drift conditioning offers a promising direction for adaptive anatomical knowledge preservation in continual cardiac ultrasound segmentation.
Introduction
Echocardiographic measurements of cardiac structure and function depend on accurate delineation of structures such as the left ventricular (LV) cavity, myocardium, and left atrium. Maintaining segmentation accuracy across datasets is challenging because both image appearance and the available annotations can change [1, 2, 3]. In the sequence studied here, CAMUS provides annotations of the LV cavity, myocardium, and left atrium in two- and four-chamber views, CardiacNet provides chamber annotations, and EchoNet-Dynamic provides LV-only supervision. During the final task, the model must therefore adapt to a new image distribution while retaining its ability to delineate structures for which it no longer receives direct supervision. Figure 1 illustrates the variation in image appearance across the three datasets.

Figure 1. Representative echocardiographic images from CAMUS, CardiacNet, and EchoNet-Dynamic.
Continual learning seeks to preserve performance on earlier tasks during sequential training [4, 5, 6]. Elastic Weight Consolidation penalizes changes to parameters important for earlier tasks [7], while Learning without Forgetting constrains changes in predictions through distillation [8]. Experience replay revisits a bounded set of historical examples during training. In medical imaging, replay methods have addressed how to populate this memory by retaining diverse cases [9], atypical examples [10], or examples with informative anatomical boundaries [11]. Once an exemplar is sampled, a further question arises: how much anatomical supervision does it require at the current stage of training?
Consider a CAMUS exemplar whose predicted myocardial wall develops a small break after training on CardiacNet. The break may affect few pixels and leave regional overlap high, yet it changes the connectivity of the predicted myocardium. Another exemplar may retain an accurate contour despite a change in its latent representation.
Such cases motivate measuring changes at several levels and using boundary information alongside region-based supervision [12]. We use drift to denote a change in the model’s response to the same exemplar relative to its source-task state. Figure 2 illustrates how these responses evolve over successive tasks.

Figure 2. Source-referenced knowledge drift across sequential echocardiographic tasks. The same CAMUS exemplar is evaluated after training on CAMUS, CardiacNet, and EchoNet-Dynamic. Panels (A)-(D) illustrate changes in bottleneck features, class probabilities, boundary responses, and predicted anatomical structures, respectively. All drift measurements are referenced to the model state after CAMUS training.
We propose Drift-Conditioned Structural Replay Loss (DCSRL), which adjusts anatomical supervision for each replayed exemplar. Its underlying reliability-guided replay baseline considers anatomical class content and boundary complexity when retaining exemplars and adjusts the overall weight of replay during training. DCSRL uses the same selection, replacement, and sampling rules. For each retained exemplar, it records a frozen source-task signature comprising a normalized bottleneck representation, probability statistics for the supervised classes, boundary-response statistics, and a predicted anatomical mask. When the exemplar is sampled again, DCSRL compares its current responses with this signature. It combines and normalizes the measured changes into bounded weights that increase boundary and structural supervision as drift grows. This requires additional storage and computation during training and introduces no additional trainable parameters or inference operations. The comparison with reliability-guided replay isolates the contribution of drift-conditioned supervision under a common memory policy. The experiments address three questions.
Does drift conditioning improve final segmentation? DCSRL achieved the highest final average accuracy, defined as mean final Dice across the three datasets (, compared with for reliability-guided replay). It improved the mean result in 14 of 15 dataset–metric combinations spanning regional overlap, boundary agreement, and surface distance.
How do the final gains relate to transfer? DCSRL achieved lower backward and forward transfer scores than reliability-guided replay. Reporting both measures distinguishes final segmentation performance from the effects of learning across tasks.
Which components contribute to anatomical preservation? We ablated each of the four drift signals and each of the two conditioning pathways. The complete objective yielded the best mean result among the tested variants on all five EchoNet-Dynamic metrics, while several ablated variants achieved lower surface-distance errors on CAMUS.
Related Work
This section reviews four lines of work that bear on DCSRL: echocardiographic segmentation and the datasets that define the present task sequence; continual learning and its relation to drift in non-stationary data; replay-based continual learning, with emphasis on how memory is populated, retrieved, and supervised; and loss functions that preserve boundary, shape, and topology. The section closes by positioning DCSRL relative to these methods.
Deep Learning for Echocardiographic Segmentation
Accurate cardiac segmentation requires both global context to identify a chamber and local detail to delineate its contour. U-Net [13] combines these sources of information through skip connections between its encoder and decoder, and nnU-Net [14] showed that a carefully configured U-Net remains a strong baseline across many medical segmentation benchmarks. Attention U-Net [15] uses attention gates to emphasize spatially relevant features, while TransUNet [16] and Swin-Unet [17] incorporate transformer-based contextual modeling. Large pretrained segmenters such as the Segment Anything Model [18] and its medical adaptation MedSAM [19] have broadened the available backbones, and MemSAM [2] adapted this family to echocardiography videos through a temporal memory mechanism.
In echocardiography, CAMUS [1] established a benchmark for segmenting the left ventricular (LV) cavity, myocardium, and left atrium in two- and four-chamber views, and LU-Net [20] improved robustness on this benchmark with a multistage attention design. EchoNet-Dynamic [3] demonstrated video-based LV segmentation and estimation of cardiac function at a larger scale, and CardiacNet [21] introduced echocardiographic data for cardiac disease assessment. Several studies have also shown that overlap-optimized networks can produce anatomically implausible masks, such as a myocardium with a gap or a cavity split into several regions. Anatomically constrained networks [22] regularize predictions with a learned shape representation, and Painchaud et al. enforced anatomical plausibility [23] and temporal consistency [24] in cardiac segmentation. These studies treat anatomical plausibility as a property that overlap metrics can miss, which motivates the boundary and surface measures used in this work.
These datasets differ in annotation scope, which matters when they are encountered sequentially. An LV tracing provides a target for updating the ventricular contour, but the effect of that update on a previously learned myocardial boundary can be assessed only on an earlier task. Existing echocardiographic methods are trained and evaluated on a fixed dataset; they do not address how a model’s anatomy changes when datasets with different label spaces arrive in sequence.
Continual Learning and Drift in Non-Stationary Data
Continual learning addresses training on a sequence of data distributions while limiting catastrophic forgetting of earlier ones [4, 5, 25]. Clinical imaging data have this character: scans accumulate over time and across institutions, scanners, and annotation protocols, so a deployed model is repeatedly exposed to new data rather than to a single fixed corpus [26]. The data-stream literature describes such changes as concept drift, a change in the joint distribution of inputs and targets over time, and proposes methods to detect and adapt to it [27, 28]. DCSRL uses drift in a narrower, model-centered sense: a change in the model’s own response to a fixed stored exemplar relative to the model state at which that exemplar was learned. The quantity is measured per exemplar rather than per data stream, and it is used to set supervision strength rather than to trigger model adaptation.
Continual-learning methods differ in how they preserve earlier knowledge during model updates. Regularization methods estimate the importance of each parameter and penalize changes to important ones: Elastic Weight Consolidation (EWC) [7] uses Fisher information, Synaptic Intelligence [29] accumulates each parameter’s contribution to the loss along the training trajectory, and Memory Aware Synapses [30] uses the sensitivity of the learned function. Learning without Forgetting (LwF) [8] instead uses predictions from an earlier model as distillation targets on incoming data. For dense prediction, the choice of what to preserve also has a spatial dimension. Continual semantic segmentation methods distill intermediate features [31], regularize latent representations [32], model the shifting meaning of the background class [33], or apply multi-scale pooled distillation with pseudo-labeling [34]. A parameter-importance estimate does not locate a displaced contour, and agreement on incoming images provides only indirect evidence of performance on earlier anatomy.
In medical image segmentation, early studies showed that sequential fine-tuning across sites or anatomies causes substantial forgetting and evaluated adaptations of EWC and distillation [35, 36, 37]. Lifelong nnU-Net [38] standardized the training and evaluation of continual segmentation, including a cardiac use case, and surveys have organized the growing literature on continual medical image analysis [6]. Class-incremental methods for CT have used pseudo-labels and class-specific output heads to add organs and tumors without access to earlier data [39, 40]. Continual learning has also been applied to cardiac signals across diseases, time, modalities, and institutions [41] and to domain shifts in chest X-ray classification [42]. These studies establish that clinical models must accommodate both new domains and new label spaces, the combination encountered in the present CAMUS CardiacNet EchoNet-Dynamic sequence.
Several methods incorporate shape or morphology explicitly. Shape and Semantics-based Selective Regularization (R) uses shape and semantic information to identify parameters to protect during subsequent training [43]. CLMS [44] addresses source-free adaptation through multi-scale reconstruction and style alignment to preserve morphology as image appearance changes. Work on online continual learning [45] has also documented abrupt changes in learned representations when new classes are introduced. For segmentation, however, feature displacement and anatomical error are distinct quantities: a representation can change while its decoded contour remains stable, and a contour can break while overall overlap remains high. This observation motivates measuring drift at several levels rather than relying on a single representation-level signal.
Replay-Based Continual Learning
Experience replay [46, 47] interleaves historical examples with incoming data, and even small episodic memories can support retention. Replay has also been used as a constraint rather than a training signal: Gradient Episodic Memory [48] and A-GEM [49] project updates so that the loss on stored examples does not increase. Generative replay [50] replaces stored images with samples from a generative model, at the cost of training and maintaining that model.
With a limited memory, both buffer composition and retrieval determine which observations contribute to training. iCaRL [51] selects exemplars that approximate class means in feature space and combines rehearsal with distillation. Gradient-based sample selection [52] favors stored examples with diverse gradients, Rainbow Memory [53] favors diverse and uncertain samples, and GDumb [54] showed that a class-balanced memory alone forms a strong baseline. For dense prediction, Zhu et al. [55] learned an automatic memory-sample selection policy. Medical imaging methods have used acquisition-style diversity [9], sample atypicality [10], and the combination of boundary information and class balance in BBR-Net [11] to guide retention. Atlas Replay [56] takes a different approach to memory representation, using anatomical prototypes and registration for continual prostate MRI segmentation.
A separate group of methods changes what a retrieved exemplar contributes. Maximally Interfered Retrieval (MIR) [57] estimates which stored examples would incur the largest loss increase after a prospective update and gives those examples retrieval priority. Dark Experience Replay (DER) [58] stores historical logits and penalizes discrepancies between those logits and current outputs, and DER++ adds supervised rehearsal on stored labels. X-DER [59] revises stored logits as training proceeds, and CLS-ER [60] distills from slowly and rapidly updated model averages. MIR uses a prospective change to guide retrieval, whereas DER and its extensions use a historical or consolidated prediction as a training target. These precedents distinguish the decision to retrieve an exemplar from the choice of supervision it receives once retrieved. DCRSL addresses the second decision: it leaves retrieval unchanged and uses a measured change in the exemplar’s response to set the strength of anatomical supervision, while the stored annotation, not a stored prediction, supplies the target.
Anatomical and Structural Knowledge Preservation
Two masks with similar Dice scores can differ in contour placement or in the continuity of a thin myocardial wall. Region-based and distribution-based losses weight these errors differently [61]. Boundary loss [12] incorporates distance to the reference contour into the training objective. Hausdorff-inspired losses give greater weight to large spatial discrepancies [62], while boundary-wise objectives provide additional supervision at anatomical interfaces [63]. Kang et al. [64] incorporated neighborhood-based label smoothing and reverse boundary attention into Diffusion Transformer Segmentation to improve morphological representation and boundary delineation.
Topology-aware objectives address connectivity directly. Hu et al. [65] matched the persistent homology of predicted and reference likelihood maps, Clough et al. [66] used persistent homology to impose prior topological constraints, and clDice [67] preserves the connectivity of tubular structures through a skeleton-based overlap measure. These methods establish ways to penalize spatial and topological errors in a static training setting. In a continual setting, a further design choice is how strongly to apply such penalties to a particular replayed image. A uniform weight applies the same supervision to an exemplar whose myocardial wall has broken and to one whose prediction is unchanged.
Positioning of DCSRL
DCSRL addresses this choice within a reliability-guided replay baseline that considers anatomical class content and boundary complexity when retaining exemplars and adjusts the overall replay weight during training. After an exemplar is sampled, DCSRL compares its current features, supervised-class probabilities, boundary responses, and predicted mask with a frozen signature recorded during its source task. The measured changes determine bounded, exemplar-specific weights for the boundary and structural replay losses. Using the same memory selection, replacement, and sampling rules in both methods allows the comparison to isolate the effect of conditioning anatomical supervision on the exemplar’s observed drift.
Table 1 summarizes representative continual-learning strategies relevant to medical image segmentation and their differences from DCSRL. Existing approaches primarily preserve important parameters, distill historical predictions or features, or improve the construction and retrieval of replay memory. Static boundary and topology losses, in turn, apply a fixed weight to every training image. Although these strategies mitigate forgetting or improve contour quality, they do not regulate anatomical preservation according to multiple source-referenced drift measurements for each replayed exemplar. DCSRL addresses this complementary problem by measuring feature, probability, boundary, and structural changes and using their normalized discrepancies to adapt the strength of boundary and anatomical supervision. Because this conditioning operates after exemplar selection, it can be combined with the memory-selection and retrieval strategies reviewed above rather than replacing them.
| Category | Method | Relation to DCSRL |
| Regularization | EWC [7] | Protects important parameters; DCSRL adapts anatomical replay losses using measured drift. |
| Distillation | LwF [8] | Preserves previous predictions on incoming data; DCSRL measures feature, probability, boundary, and structural drift on stored exemplars. |
| Segmentation distillation | PLOP [34] | Distills multi-scale features on incoming images; DCSRL conditions supervision on replayed exemplars. |
| Benchmark framework | Lifelong nnU-Net [38] | Standardizes continual segmentation with uniform replay and regularization; DCSRL adds exemplar-specific supervision weights. |
| Exemplar replay | iCaRL [51] | Selects representative exemplars; DCSRL adjusts supervision after memory retrieval. |
| Interference replay | MIR [57] | Prioritizes samples with predicted interference; DCSRL conditions replay losses on observed drift. |
| Logit replay | DER/DER++ [58] | Preserves stored logits; DCSRL keeps the annotation as the target and adapts boundary and anatomical supervision. |
| Dynamic memory | Perkonigg et al. [9] | Maintains diverse historical samples; DCSRL measures changes in retained knowledge. |
| Anatomy-aware replay | BBR-Net [11] | Selects samples using class and boundary information; DCSRL adapts structural preservation strength. |
| Reliability-guided replay | Reliability-Guided Replay | Uses reliability-based selection and adaptive replay; DCSRL adds exemplar-specific drift conditioning. |
| Drift-conditioned replay | DCSRL (Ours) | Combines four source-referenced drift measures to adapt boundary and structural replay losses without changing memory selection. |
Table 1. Comparison of representative continual-learning approaches relevant to medical image segmentation.
Proposed Method
Drift-Conditioned Structural Replay Loss (DCSRL) adjusts anatomical supervision for each replayed exemplar according to how the model’s response has changed since the exemplar was learned. Each retained image has a frozen source-task signature. During replay, DCSRL compares the current response with this signature and uses the measured drift to weight boundary and structural losses. Figure 3 shows where these weights enter the training objective.
Figure 3. DCSRL training procedure. Current-task images receive segmentation and boundary supervision, while historical exemplars are retrieved using a reliability-guided replay policy. For each replayed exemplar, feature, probability, boundary, and structural drift are measured relative to its frozen source-task reference signature. These measurements determine exemplar-specific boundary and structural loss weights without modifying the underlying memory-selection policy.
Sequential Segmentation and Network Outputs
We train a single segmentation network on three successive tasks,
Task provides a dataset , where is an image, is its target mask, and identifies the annotation mode used for training. Earlier training images are accessible through a bounded replay memory . We denote the foreground channels used for exemplar by .
The shared output space comprises background (BG), left ventricular cavity (LV), myocardium (MYO), and left atrium (LA). CAMUS provides annotations for all three foreground structures [1]. CardiacNet provides chamber annotations in its native label space. EchoNet-Dynamic provides LV tracings [3]. For LV-only supervision, the target distinguishes LV from its complement without assigning a specific anatomical class to non-LV pixels.
The backbone is a U-Net [13] with four encoder stages, a bottleneck, and four decoder stages connected to the encoder through skip connections (Figure 4A). For image , the network returns
Figure 4. Network architecture and source-task reference construction. (A) The U-Net backbone produces segmentation predictions, auxiliary boundary responses, and bottleneck features. (B) For each retained exemplar, a source-task reference signature is constructed from its normalized bottleneck representation, spatially averaged class probabilities, boundary-response statistics, and predicted segmentation mask. The stored reference remains frozen during subsequent continual-learning tasks.
where contains the segmentation logits, contains the auxiliary boundary logits, and is the globally pooled bottleneck representation. We write for the class probabilities and for the boundary response.
Reliability-Guided Replay and Source References
Reliability-guided replay baseline
We use reliability-guided replay as the reference configuration for DCSRL. The baseline maintains a bounded memory of historical images and their segmentation masks. A priority score governs both exemplar retention and replay sampling. During training, an epoch-level controller adjusts the overall replay weight, and a structural loss provides additional supervision on eligible replay masks (§ 3.6).
Reliability-score definitions
Let . For an image normalized to , the image reliability score is
where is converted to an unsigned 8-bit image and is the OpenCV Laplacian with its default kernel.
Let be the number of foreground pixels and the number of boundary pixels. A pixel is marked as a boundary pixel if any horizontal or vertical neighbor has a different label. This rule includes interfaces between foreground classes and marks both sides of a transition. For an mask, . We set when . Otherwise,
where if and otherwise.
For EchoNet-Dynamic, temporal scores are computed from traced LV areas in the original frame coordinates. A video with one traced frame receives a score of 0.5. For a video with multiple traced frames, a frame with area at most 10 pixels receives 0.1. For the remaining frames, the score is 1.0 when the ratio of the minimum to maximum traced area lies in , and 0.5 otherwise.
Exemplar priority.
For candidate image and its training mask , we compute
Here, is the fraction of the three foreground label indices present in , and is the fraction of pixels adjacent to a horizontal or vertical label transition. The image score combines contrast, sharpness, and visibility. The anatomical score combines foreground area, compactness, and boundary smoothness. The temporal score is fixed at 0.5 for CAMUS and CardiacNet and is derived from traced LV areas for EchoNet-Dynamic. The component definitions are given in Appendix § 3.2.2.
Memory construction.
After each task, we score candidates from its training loader, build a task-specific buffer, and merge its retained entries into the global buffer. Both buffers have a capacity of 1800 exemplars. Candidates are appended while space is available. For a full buffer, the target task fractions are proportional to 0.25, 0.50, and 0.25 for CAMUS, CardiacNet, and EchoNet-Dynamic, respectively, and are normalized over the tasks represented at that stage.
If the incoming task is below its target, we consider the stored task with the largest excess over its target for replacement. Otherwise, we consider entries from the incoming task itself. The incoming exemplar replaces the lowest-priority entry in the selected task only if its priority is strictly higher. Thus, the final targets of 450, 900, and 450 guide replacement, while the realized counts depend on candidate availability and priority.
Replay sampling.
Each replay batch contains eight draws with replacement from the entire global buffer. For buffer , the probability of drawing exemplar is
Both methods use this distribution. DCSRL additionally stores a source-task signature and uses it to adjust supervision after an exemplar has been sampled.
Source-task reference signatures. For each exemplar retained in memory, DCSRL records a frozen source-task reference signature comprising its normalized bottleneck representation, class-probability statistics, boundary-response statistics, and predicted segmentation mask:
The superscript 0 denotes the source state of that particular exemplar. For either the source state () or the current state (), the signature quantities are
Here, is a numerical stabilizer. The stored signature remains fixed throughout subsequent tasks. Its mask records the model’s source prediction. The exemplar’s annotation supplies the target for the supervised replay losses.
Exemplar-Level Drift
For each sampled exemplar, we measure changes in its representation, class probabilities, boundary response, and predicted structure (Figure 5).
Figure 5. Multidimensional drift estimation and adaptive replay supervision. For each replayed exemplar, feature, probability, boundary, and structural discrepancies are measured relative to its frozen source-task reference signature. The four drift measurements are combined into an aggregate score and normalized to derive exemplar-specific boundary and structural loss weights. These weights regulate anatomical supervision during continual training.
Feature and probability drift.
Feature drift measures the angular displacement of the bottleneck representation. Probability drift measures the mean absolute change in spatially averaged probabilities over the exemplar’s supervised foreground channels:
For an LV-only exemplar, .
Boundary-response drift.
We measure changes in the mean and variability of the auxiliary boundary response:
These statistics summarize the distribution of boundary responses. The structural term below measures where the predicted contours change.
Structural drift.
Let be the binary mask for class extracted from , let denote its contour, and let count its 8-connected components. We compute contour disagreement and component-count change as
where denotes symmetric difference. Let contain the classes in that appear in either the source or current prediction. For a nonempty , structural drift is
Excluding classes absent from both predictions prevents agreement on empty masks from diluting changes in the represented anatomy.
Drift-Conditioned Replay Weights
We aggregate the four discrepancies into an exemplar-level score:
The same bounded normalization is applied to the aggregate, boundary, and structural scores:
Writing , , and , the branch weights are
Normalization gives and . The structural branch assigns more weight to changes in the predicted mask, while the boundary branch gives equal weight to aggregate drift and boundary-response drift.
Segmentation and Anatomical Losses
For a target fully labeled in the shared output space, the regional loss combines multiclass cross-entropy and Dice loss. For an LV-only target, binary cross-entropy and Dice loss are evaluated using the LV probability and binary LV mask:
For exemplar , the boundary loss is pixel-averaged binary cross-entropy between and the boundary target extracted from its annotation. The base loss on a batch is
where is evaluated according to the exemplar’s annotation mode.
For replay targets with MYO and LA supervision, the structural loss uses a class-specific Tversky loss:
Here , , and are soft counts computed from the class probabilities and target masks. The larger false-negative coefficient penalizes missed anatomy. For LV-only exemplars, .
Epoch-Level Replay Regulation and Training Objective
Both the baseline and DCSRL adjust the global replay coefficient using validation Dice on earlier tasks. Let be the best previously observed validation Dice for task , and let be its validation Dice after epoch of the current task. For the nonempty set of earlier tasks, the controller computes
The replay coefficient is 1.0 during the first three epochs. Following warm-up, the coefficient for the next epoch is
Let and denote the current and replay batches, respectively, and let . For , the complete objective is
The boundary term uses because the base replay loss already includes boundary supervision. The factor combines the baseline structural term and its drift-dependent increment. Setting for all replayed exemplars recovers the reliability-guided replay objective. On the first task, training uses only . Reference storage and drift estimation add training costs, while inference uses the shared segmentation network without additional trainable parameters.
Algorithm 1 summarizes the complete training procedure of the proposed DCSRL framework. The method follows a sequential learning strategy in which current-task optimization is supplemented by historical exemplars maintained through a reliability-guided replay policy. For each replayed exemplar, DCSRL measures feature, probability, boundary, and structural drift relative to its frozen source-task reference signature. The resulting drift measurements determine exemplar-specific boundary and structural loss weights, allowing anatomical supervision to respond to changes in previously learned representations and predictions. Drift measurements are computed without gradient tracking and serve exclusively as exemplar-specific supervision coefficients rather than differentiable optimization targets. The segmentation network operates in training mode during parameter updates. Following each task, the replay memory is updated according to the underlying selection and replacement policy, while newly retained exemplars are associated with frozen reference signatures for subsequent tasks.
Experimental Setup
Datasets and Continual-Learning Protocol
All methods followed the sequence
At each stage, the model trained on the current dataset. Replay-based methods could also access a bounded memory of earlier training examples. After each stage, we evaluated the model across tasks to measure final performance, retention, and generalization to datasets encountered later in the sequence.
CAMUS.
The CAMUS dataset [1] contains two-dimensional echocardiographic examinations from 500 patients. Each examination includes apical two-chamber (2CH) and four-chamber (4CH) views, with expert annotations provided at end-diastole (ED) and end-systole (ES). The dataset contains patients with different image qualities and cardiac conditions, making it a useful benchmark for multi-structure echocardiographic segmentation. We followed the official patient-level partition: 400 patients were used for training, 50 for validation, and the remaining 50 for testing. Keeping the partition at the patient level ensures that images originating from the same patient cannot appear across training and evaluation subsets.
For our experiments, the segmentation label space was standardized to four classes: BG, LV, MYO, and LA, represented by labels 0, 1, 2, and 3, respectively. Only the annotated ED and ES acquisitions were used. Echocardiographic images and their corresponding masks were resized to pixels. Image intensities were normalized to the range [0, 1], while nearest-neighbor interpolation was used for segmentation masks to preserve discrete anatomical labels.
CardiacNet
CardiacNet was introduced by Yang et al. [2] for echocardiographic cardiac-disease assessment. Its inclusion as the second task introduces a distinct echocardiographic distribution after CAMUS rather than another partition of the same source dataset.
Algorithm 1 Training Procedure of Drift-Conditioned Structural Replay Loss (DCSRL)
Require: Sequential datasets ; segmentation model ; memory capacity ; drift weights Ensure: Continually trained model
1: Initialize model parameters 2: Initialize replay memory 3: for do  4: Initialize Adam optimizer and replay controller  5: Initialize early-stopping counter  6: for do   7: Set to training mode   8: for each current-task batch do    9: Compute current-task loss    10:    11: if then     12: Sample replay batch using baseline priorities     13: Compute base replay loss     14: Compute baseline structural replay loss     Source-referenced drift estimation:     15: Evaluate replay responses without gradient tracking     16: for each exemplar do      17: Retrieve frozen source reference      18: Compute feature and probability drift ,      19: Compute boundary and structural drift ,      20: Aggregate drift      21: Normalize , , using Eq. (19)      22: Compute multipliers , using Eqs. (20)–(21)     23: end for     24: Restore training mode     25: Compute using Eq. (30)    26: end if    27: Update using Adam and mixed-precision training   28: end for   29: Evaluate all previously learned tasks and the current task   30: Update replay coefficient for the next epoch using Eqs. (27)–(29)   31: Update early-stopping counter using current-task validation Dice   32: if early-stopping patience is reached then    33: break   34: end if   35: end for  36: Evaluate and save the end-of-task model checkpoint  37: Score current-task training exemplars using reliability-guided priorities  38: Construct task-specific candidate memory  39: Compute frozen source-task reference signatures for candidates  40: Merge candidates into using the baseline allocation and replacement policy  41: Retain reference signatures for admitted exemplars 42: end for 43: return
For the present segmentation experiments, the available NIfTI image-label pairs were processed slice-wise and harmonized with the common segmentation space
Following the label convention used throughout the experimental pipeline, CardiacNet label 4 was mapped to background,
Slices containing no foreground cardiac structure after label harmonization were excluded from training. CardiacNet therefore evaluates whether the model can adapt to a second echocardiographic distribution while retaining the multiclass anatomical representation acquired from CAMUS.
EchoNet-Dynamic.
EchoNet-Dynamic [3] is a large-scale echocardiographic video dataset containing 10,030 apical four-chamber videos acquired during routine clinical practice. In contrast to CAMUS and CardiacNet, EchoNet-Dynamic provides expert LV tracings rather than complete LV-MYO-LA segmentation annotations. Expert contour coordinates were converted into binary LV masks, and the corresponding annotated frames were extracted from the original videos. Frames and masks were subsequently resized to pixels. Because MYO and LA GT annotations are unavailable in EchoNet-Dynamic, these structures were not treated as background during optimization. Instead, EchoNet-Dynamic samples were handled using LV-only partial supervision. Evaluation on this dataset was likewise restricted to the LV. This distinction is particularly relevant to the CL setting: the final task changes not only the image domain but also the available anatomical supervision. For EchoNet-Dynamic, we follow the original dataset partition.
Continual-Learning Protocol
All methods followed the task sequence in Eq. (31). Let denote the Dice performance on task after completing training on task . At stage , training was performed primarily on the current dataset . Replay-based methods were allowed to access only their bounded replay memory from previous tasks; complete historical datasets were not revisited. This constraint distinguishes the experimental protocol from joint retraining and ensures that all replay-based methods operate under a common historical-data budget.
For the CAMUS dataset, training and validation subsets are established using either the official data partitioning protocol or an 80%–20% patient-level split to prevent data leakage. Similarly, CardiacNet is divided into internal training and validation subsets when used for model development, whereas the entire dataset serves as an independent external test set for assessing cross-dataset generalization.
EchoNet-Dynamic retained its predefined TRAIN and VAL partitions. The entire continual sequence was independently repeated for
For all replay-based comparisons, the total replay-memory capacity was fixed at
examples. The current-task mini-batch size was 16 and the replay mini-batch size was 8. Maintaining a common memory budget and replay batch size prevents performance differences from being attributed simply to unequal access to historical observations.
Preprocessing and Training
We converted images to grayscale, resized them to using area interpolation, and applied per-image min–max normalization:
Masks were resized using nearest-neighbor interpolation. Training augmentations comprised horizontal flipping and rotation by multiples of , each applied with probability 0.5. Each image and its mask received the same geometric transformation. Boundary targets were derived from transitions between neighboring mask labels.
All methods used the U-Net backbone described in § 3 [13]. The four encoder stages had 32, 64, 128, and 256 channels, followed by a 512-channel bottleneck. Each block contained two convolutions with batch normalization and ReLU activation. The encoder used max pooling, and the decoder used transposed convolutions and skip connections. The final decoder features supplied a four-channel segmentation head and a single-channel boundary head. DCSRL extracted its reference features by global average pooling of the bottleneck.
As illustrated in Table 2, we trained with Adam at an initial learning rate of and weight decay of . Each task ran for up to 20 epochs with automatic mixed precision and early stopping after eight epochs without validation improvement. Current-task batches contained 16 images. Replay-based methods used a capacity of 1800 exemplars and a replay batch size of eight. This budget fixes the number of retained examples; the storage required per exemplar depends on the method. The regional objective followed the annotation mode defined in § 3, with boundary-loss weight .
| Hyperparameter | Value |
| Segmentation backbone | U-Net with boundary head |
| Input resolution | |
| Input channels | 1 (grayscale) |
| Output classes | 4 |
| Initial feature channels | 32 |
| Optimizer | Adam |
| Initial learning rate | |
| Weight decay | |
| Maximum epochs per task | 20 |
| Early stopping patience | 8 epochs |
| Current-task batch size | 16 |
| Replay batch size | 8 |
| Mixed-precision training | Enabled |
| Random seeds | 42, 123, 2025 |
| Replay memory capacity | 1800 exemplars |
| Memory allocation (CAMUS/CardiacNet/EchoNet) | 25% / 50% / 25% |
| Initial replay coefficient | 1.0 |
| Adaptive replay range | [0.75, 1.75] |
| Replay-controller gain | 4.0 |
| Replay-controller EMA | 0.70 |
| Replay warm-up | 3 epochs |
| Boundary-loss weight | 0.30 |
| Structural replay weight | 0.30 |
| MYO structural weight | 1.25 |
| LA structural weight | 1.00 |
| Tversky coefficients | (0.30, 0.70) |
| Drift weights (feature/probability/boundary/structure) | 0.30/0.25/0.20/0.25 |
| Drift normalization bounds | [0.05, 0.50] |
| Boundary-conditioning gain | 0.75 |
| Structural-conditioning gain | 1.00 |
| Maximum boundary multiplier | 1.75 |
| Maximum structural multiplier | 2.00 |
| Numerical stabilizer |
Table 2. Training hyperparameters and configuration of the proposed DCSRL framework.
Comparison Methods
We compared DCSRL with 11 baselines using the same task order, backbone, preprocessing, and evaluation protocol.
Fine-tuning, regularization, and distillation.
Fine-tuning optimized the network sequentially on each current task. Elastic Weight Consolidation (EWC) used a diagonal Fisher estimate computed from up to 40 training batches per task and a regularization coefficient of 20 [7]. Learning without Forgetting (LwF) used the preceding model as a frozen teacher, with distillation weight 0.5 and temperature 2 [8]. Distillation used four-channel predictions for multiclass targets and LV-versus-non-LV predictions for EchoNet-Dynamic.
Memory construction and retrieval.
Plain Replay used first-in-first-out replacement and uniform retrieval. Random Replay used reservoir sampling for admission and uniform retrieval [47]. The iCaRL-style baseline used feature herding within class-containing subsets, followed by global herding to fill the remaining memory [51]. Maximally Interfered Retrieval (MIR) prioritized Note: EMA denotes exponential moving average. The three random seeds are 42, 123, and 2025. Replay-memory allocation percentages correspond to CAMUS, CardiacNet, and EchoNet-Dynamic, respectively.
stored examples expected to incur a larger loss after a current-task update [57]. We also included an AMSS-inspired exemplar-selection baseline adapted from continual semantic segmentation [55].
Historical-output replay.
Dark Experience Replay (DER) constrained current predictions using logits stored with historical examples. DER++ additionally applied supervised replay to the stored labels [58].
Reliability-guided replay.
We used the configuration defined in § 3.2.1 as the direct control for DCSRL. Both configurations share the exemplar priorities, replacement and sampling rules, epoch-level replay controller, and base structural loss. Their comparison evaluates the additional effect of exemplar-specific drift conditioning.
Evaluation Metrics
Segmentation Performance
After the final task, we evaluated regional overlap using Dice and intersection over union (IoU). For a predicted region and reference region ,
CAMUS scores were averaged over LV, MYO, and LA. CardiacNet scores were averaged over the three foreground channels retained by its experimental encoding. EchoNet-Dynamic was scored on LV alone.
For surface evaluation, we computed the distance from each predicted boundary point to its nearest reference boundary point and repeated the calculation in the reverse direction. The 95th-percentile Hausdorff distance (HD95) and average surface distance (ASD) were the 95th percentile and mean, respectively, of the pooled bidirectional distances. Boundary F1 was the harmonic mean of boundary precision and recall, using a matching tolerance of two pixels in each direction. Higher Dice, IoU, and boundary F1 indicate better agreement. Lower HD95 and ASD indicate smaller surface errors.
Continual-Learning Performance
Let denote the Dice score on task after training on task , and let . We report final average accuracy (FAA), backward transfer (BWT), and forward transfer (FWT). In this study, FAA is the unweighted mean final Dice across datasets. BWT measures the change on earlier tasks relative to their scores immediately after training [48]:
A positive BWT indicates that subsequent training improved the mean Dice on earlier tasks.
Our FWT measure is the mean Dice on future tasks before the model receives training data from those tasks:
This definition reports absolute pre-training performance on future tasks. Its numerical scale differs from FWT definitions that subtract a random-initialization reference.
Ablation and Computational Cost
We evaluated six ablations under the full model’s training and memory settings. Four variants individually removed feature, probability, boundary, or structural drift. Two further variants disabled boundary conditioning or structural conditioning while retaining the underlying supervised loss. We evaluated each variant on all three datasets using the five segmentation metrics. Reliability-guided replay provided the control with both conditioning branches disabled in the main comparison.
For DCSRL and reliability-guided replay, we recorded trainable parameter count, per-task training time, peak GPU memory, and persistent replay storage after each task. DCSRL storage included its source-reference signatures. For each task, training overhead was calculated as the percentage increase in mean training time relative to reliability-guided replay.
Statistical Analysis
We repeated each principal experiment with seeds 42, 123, and 2025 and report the mean and sample standard deviation across runs. For the direct comparison, we paired DCSRL and reliability-guided replay by seed. Let denote the difference for seed . We used a two-sided paired -test and calculated the paired standardized effect size , where and are the mean and sample standard deviation of the paired differences. Negative differences favor DCSRL for HD95 and ASD, and positive differences favor DCSRL for the overlap and boundary F1 scores. With three seed pairs, the analysis emphasizes the magnitude and direction of the observed differences. The reported standard deviations describe variation across training runs.
Experimental Results
Final Performance and Knowledge Transfer
DCSRL achieved the highest final average accuracy (FAA) across the three seeds, with a mean final Dice of 0.8784 0.0017, compared with 0.8718 0.0029 for reliability-guided replay (Table 3), the next-best method. This represents an absolute gain of 0.66 Dice percentage points under the same replay-memory policy.
| Method | FAA | BWT | FWT |
| Fine-tuning | |||
| EWC | |||
| LwF | |||
| Plain Replay | |||
| Random Replay | |||
| iCaRL Replay | |||
| DER | |||
| DER++ | |||
| MIR | |||
| AMSS-inspired | |||
| Reliability-Guided Replay | |||
| DCSRL (Proposed) |
Table 3. Continual-learning performance of the compared methods. Note: Values are reported as mean standard deviation over three random seeds (42, 123, and 2025). Higher FAA, BWT, and FWT indicate better performance. Bold indicates the best mean for each metric.
The transfer measures showed a different pattern. DCSRL’s backward transfer was close to zero (0.0003 0.0049), indicating little average change on earlier tasks relative to their scores immediately after training. Reliability-guided replay achieved higher backward transfer () and higher forward transfer ( versus ). Among all methods, DER++ achieved the highest backward transfer (), and the AMSS-inspired method achieved the highest forward transfer ().
Final Segmentation Performance
After the full training sequence, DCSRL improved the mean result in 14 of 15 dataset–metric combinations relative to reliability-guided replay (Table 4). CardiacNet showed the largest Dice gain, from 0.8096 to 0.8218, and the largest reductions in surface error. HD95 decreased from 16.0261 to 10.2148 (36.3%), and ASD decreased from 5.2089 to 3.7972 (27.1%). Boundary F1 also increased from 0.5765 to 0.5998.
| Dataset | Method | Dice ↑ | IoU ↑ | HD95 ↓ | ASD ↓ | Boundary F1 ↑ |
| CAMUS | Fine-tuning | 0.3268 ± 0.0289 | 0.2608 ± 0.0293 | 77.9952 ± 27.9890 | 38.0120 ± 17.1613 | 0.1100 ± 0.0195 |
| CAMUS | EWC | 0.4564 ± 0.0202 | 0.3610 ± 0.0254 | 31.8637 ± 4.3128 | 10.6272 ± 2.2729 | 0.1690 ± 0.0440 |
| CAMUS | LwF | 0.3767 ± 0.0269 | 0.3073 ± 0.0186 | 53.1436 ± 7.0249 | 19.4207 ± 5.7180 | 0.1728 ± 0.0392 |
| CAMUS | Plain Replay | 0.3392 ± 0.0745 | 0.2472 ± 0.0705 | 41.7560 ± 7.3061 | 17.4481 ± 3.4357 | 0.1417 ± 0.0385 |
| CAMUS | Random Replay | 0.8870 ± 0.0032 | 0.7986 ± 0.0051 | 8.3637 ± 0.4174 | 3.1037 ± 0.0779 | 0.5522 ± 0.0125 |
| CAMUS | iCaRL Replay | 0.8810 ± 0.0067 | 0.7891 ± 0.0105 | 9.1623 ± 0.6260 | 3.3030 ± 0.1932 | 0.5463 ± 0.0073 |
| CAMUS | DER | 0.8760 ± 0.0017 | 0.7814 ± 0.0030 | 21.1039 ± 21.3896 | 5.6569 ± 4.1085 | 0.5069 ± 0.0268 |
| CAMUS | DER++ | 0.8882 ± 0.0005 | 0.8005 ± 0.0008 | 7.9559 ± 0.2681 | 2.9995 ± 0.0437 | 0.5574 ± 0.0039 |
| CAMUS | MIR | 0.8880 ± 0.0059 | 0.8002 ± 0.0095 | 8.6204 ± 0.3489 | 3.0953 ± 0.1001 | 0.5525 ± 0.0078 |
| CAMUS | AMSS-inspired | 0.8910 ± 0.0052 | 0.8049 ± 0.0084 | 8.1574 ± 0.6152 | 2.9944 ± 0.2062 | 0.5669 ± 0.0170 |
| CAMUS | Reliability-Guided Replay | 0.8977 ± 0.0024 | 0.8114 ± 0.0044 | 8.2862 ± 0.5435 | 2.9359 ± 0.1164 | 0.5750 ± 0.0096 |
| CAMUS | DCSRL (Proposed) | 0.8986 ± 0.0030 | 0.8172 ± 0.0049 | 8.4169 ± 0.7366 | 2.8752 ± 0.1583 | 0.5900 ± 0.0033 |
| CardiacNet | Fine-tuning | 0.1823 ± 0.0108 | 0.1205 ± 0.0101 | 102.0346 ± 2.2724 | 59.2795 ± 2.4060 | 0.1253 ± 0.0114 |
| CardiacNet | EWC | 0.2104 ± 0.0078 | 0.1364 ± 0.0055 | 71.8151 ± 3.4866 | 36.4733 ± 8.2664 | 0.1202 ± 0.0195 |
| CardiacNet | LwF | 0.1242 ± 0.0118 | 0.0719 ± 0.0072 | 87.4117 ± 15.4822 | 39.5617 ± 9.8568 | 0.0972 ± 0.0100 |
| CardiacNet | Plain Replay | 0.7820 ± 0.0195 | 0.6439 ± 0.0259 | 20.9466 ± 6.9804 | 6.8668 ± 2.0048 | 0.5463 ± 0.0433 |
| CardiacNet | Random Replay | 0.7516 ± 0.0259 | 0.6053 ± 0.0321 | 21.5832 ± 4.2886 | 7.3905 ± 1.5881 | 0.5146 ± 0.0299 |
| CardiacNet | iCaRL Replay | 0.7280 ± 0.0349 | 0.5753 ± 0.0413 | 23.2788 ± 6.1884 | 7.8652 ± 2.0925 | 0.4850 ± 0.0480 |
| CardiacNet | DER | 0.6021 ± 0.1064 | 0.4447 ± 0.1071 | 49.6329 ± 27.5366 | 19.0059 ± 11.6101 | 0.3144 ± 0.0887 |
| CardiacNet | DER++ | 0.7519 ± 0.0196 | 0.6041 ± 0.0258 | 22.3016 ± 4.5647 | 7.5371 ± 1.1471 | 0.4882 ± 0.0186 |
| CardiacNet | MIR | 0.7646 ± 0.0186 | 0.6206 ± 0.0239 | 18.8856 ± 4.3442 | 6.5883 ± 1.3403 | 0.5252 ± 0.0324 |
| CardiacNet | AMSS-inspired | 0.7553 ± 0.0131 | 0.6089 ± 0.0172 | 21.3780 ± 2.8054 | 7.3317 ± 0.6696 | 0.4968 ± 0.0191 |
| CardiacNet | Reliability-Guided Replay | 0.8096 ± 0.0055 | 0.6777 ± 0.0106 | 16.0261 ± 0.9857 | 5.2089 ± 0.1841 | 0.5765 ± 0.0122 |
| CardiacNet | DCSRL (Proposed) | 0.8218 ± 0.0029 | 0.6978 ± 0.0041 | 10.2148 ± 0.4996 | 3.7972 ± 0.1541 | 0.5998 ± 0.0071 |
| EchoNet | Fine-tuning | 0.9109 ± 0.0006 | 0.8363 ± 0.0010 | 15.8577 ± 0.4639 | 4.3225 ± 0.0816 | 0.5137 ± 0.0017 |
| EchoNet | EWC | 0.9139 ± 0.0005 | 0.8414 ± 0.0009 | 15.4920 ± 0.4814 | 4.2132 ± 0.0819 | 0.5229 ± 0.0041 |
| EchoNet | LwF | 0.5215 ± 0.0935 | 0.3562 ± 0.0830 | 47.9719 ± 14.1903 | 15.5658 ± 2.2557 | 0.1662 ± 0.0256 |
| EchoNet | Plain Replay | 0.9061 ± 0.0014 | 0.8283 ± 0.0023 | 16.0567 ± 0.1107 | 4.4771 ± 0.0612 | 0.4927 ± 0.0083 |
| EchoNet | Random Replay | 0.9099 ± 0.0009 | 0.8347 ± 0.0015 | 15.8187 ± 0.1701 | 4.3632 ± 0.0297 | 0.5039 ± 0.0011 |
| EchoNet | iCaRL Replay | 0.9102 ± 0.0007 | 0.8352 ± 0.0011 | 15.6754 ± 0.1058 | 4.3406 ± 0.0512 | 0.5037 ± 0.0063 |
| EchoNet | DER | 0.9019 ± 0.0072 | 0.8213 ± 0.0119 | 41.9954 ± 44.3794 | 9.3054 ± 8.2795 | 0.4606 ± 0.0626 |
| EchoNet | DER++ | 0.9085 ± 0.0004 | 0.8324 ± 0.0008 | 15.7607 ± 0.0759 | 4.3708 ± 0.0147 | 0.5026 ± 0.0022 |
| EchoNet | MIR | 0.9078 ± 0.0014 | 0.8311 ± 0.0023 | 15.8751 ± 0.3035 | 4.3957 ± 0.0688 | 0.5060 ± 0.0008 |
| EchoNet | AMSS-inspired | 0.9101 ± 0.0016 | 0.8350 ± 0.0027 | 15.9973 ± 0.1409 | 4.3776 ± 0.0501 | 0.5076 ± 0.0052 |
| EchoNet | Reliability-Guided Replay | 0.9081 ± 0.0023 | 0.8331 ± 0.0026 | 15.3991 ± 0.0927 | 4.2779 ± 0.0449 | 0.5094 ± 0.0042 |
| EchoNet | DCSRL (Proposed) | 0.9148 ± 0.0005 | 0.8429 ± 0.0009 | 15.3023 ± 0.2800 | 4.1372 ± 0.0527 | 0.5283 ± 0.0085 |
Table 4. Final segmentation performance after sequential echocardiographic training.
On CAMUS, Dice changed only slightly, from 0.8977 to 0.8986, while boundary F1 increased from 0.5750 to 0.5900. CAMUS HD95 was the sole metric on which DCSRL performed worse than reliability-guided replay, increasing from 8.2862 to 8.4169. On EchoNet-Dynamic, DCSRL achieved the best mean value among the compared methods on all five metrics, including Dice () and boundary F1 ().
Component Ablation
The clearest ablation result was on EchoNet-Dynamic as illustrated in Table 5. Removing any one of the four drift signals or either conditioning branch reduced performance Note: Methods were trained sequentially on CAMUS, CardiacNet, and EchoNet-Dynamic. Values are reported as mean ± standard deviation over seeds 42, 123, and 2025. Higher Dice, IoU, and Boundary F1 are better, whereas lower HD95 and ASD indicate better performance. Bold indicates the best mean within each dataset and metric.
| Dataset | Variant | Dice | IoU | HD95 | ASD | Boundary F1 |
| CAMUS | Full DCSRL | 0.8986 0.0030 | 0.8172 0.0049 | 8.4169 0.7366 | 2.8752 0.1583 | 0.5900 0.0033 |
| CAMUS | w/o Feature Drift | 0.8972 0.0020 | 0.8150 0.0032 | 8.4092 0.3539 | 2.8947 0.0699 | 0.5864 0.0031 |
| CAMUS | w/o Probability Drift | 0.8981 0.0021 | 0.8165 0.0035 | 8.1014 0.5967 | 2.8402 0.1590 | 0.5861 0.0021 |
| CAMUS | w/o Boundary Drift | 0.8980 0.0013 | 0.8164 0.0022 | 8.3509 0.1885 | 2.8674 0.0482 | 0.5881 0.0019 |
| CAMUS | w/o Structural Drift | 0.8980 0.0019 | 0.8164 0.0031 | 8.1242 0.2546 | 2.8364 0.0850 | 0.5859 0.0065 |
| CAMUS | w/o Boundary Conditioning | 0.8985 0.0020 | 0.8172 0.0032 | 8.1326 0.5401 | 2.8307 0.0869 | 0.5889 0.0063 |
| CAMUS | w/o Structural Conditioning | 0.8982 0.0017 | 0.8166 0.0028 | 8.1778 0.2651 | 2.8473 0.0720 | 0.5872 0.0009 |
| CardiacNet | Full DCSRL | 0.8218 0.0029 | 0.6978 0.0041 | 10.2148 0.4996 | 3.7972 0.1541 | 0.5998 0.0071 |
| CardiacNet | w/o Feature Drift | 0.8217 0.0012 | 0.6977 0.0015 | 10.4531 0.9575 | 3.8113 0.1986 | 0.5983 0.0024 |
| CardiacNet | w/o Probability Drift | 0.8205 0.0009 | 0.6960 0.0014 | 10.5543 0.9134 | 3.8761 0.2159 | 0.5985 0.0038 |
| CardiacNet | w/o Boundary Drift | 0.8206 0.0020 | 0.6961 0.0029 | 10.6801 0.2106 | 3.9065 0.0598 | 0.5979 0.0042 |
| CardiacNet | w/o Structural Drift | 0.8211 0.0021 | 0.6968 0.0031 | 10.7225 0.8841 | 3.9071 0.2008 | 0.6000 0.0069 |
| CardiacNet | w/o Boundary Conditioning | 0.8211 0.0050 | 0.6969 0.0071 | 10.2307 1.1434 | 3.8003 0.2944 | 0.5964 0.0079 |
| CardiacNet | w/o Structural Conditioning | 0.8213 0.0034 | 0.6972 0.0046 | 10.6956 1.3333 | 3.8997 0.4195 | 0.5992 0.0078 |
| EchoNet-Dynamic | Full DCSRL | 0.9148 0.0005 | 0.8429 0.0009 | 15.3023 0.2800 | 4.1372 0.0527 | 0.5283 0.0085 |
| EchoNet-Dynamic | w/o Feature Drift | 0.9110 0.0002 | 0.8366 0.0004 | 15.8196 0.4126 | 4.3089 0.0726 | 0.5150 0.0013 |
| EchoNet-Dynamic | w/o Probability Drift | 0.9109 0.0004 | 0.8363 0.0007 | 15.7062 0.3316 | 4.3013 0.0610 | 0.5144 0.0004 |
| EchoNet-Dynamic | w/o Boundary Drift | 0.9108 0.0007 | 0.8363 0.0012 | 15.8187 0.2762 | 4.3191 0.0399 | 0.5131 0.0023 |
| EchoNet-Dynamic | w/o Structural Drift | 0.9110 0.0012 | 0.8366 0.0021 | 15.7303 0.5187 | 4.3086 0.1202 | 0.5140 0.0081 |
| EchoNet-Dynamic | w/o Boundary Conditioning | 0.9109 0.0007 | 0.8364 0.0012 | 15.8913 0.3425 | 4.3307 0.0521 | 0.5135 0.0003 |
| EchoNet-Dynamic | w/o Structural Conditioning | 0.9113 0.0009 | 0.8370 0.0015 | 15.5972 0.3332 | 4.2753 0.0463 | 0.5160 0.0008 |
Table 5. Component ablation results of the DCSRL framework. Note: Results follow sequential training on CAMUS, CardiacNet, and EchoNet-Dynamic. Values are mean sample standard deviation over three seeds (42, 123, and 2025). Higher Dice, IoU, and Boundary F1 are better; lower HD95 and ASD are better. Bold indicates the best mean within each dataset, including ties at the displayed precision.
on all five metrics. Across the six ablations, boundary F1 ranged from 0.5131 to 0.5160, compared with 0.5283 for full DCSRL.
On CardiacNet, the full model achieved the best mean Dice, IoU, HD95, and ASD. Removing structural drift yielded a marginally higher boundary F1 (0.6000 versus 0.5998). On CAMUS, removing probability drift reduced HD95 from 8.4169 to 8.1014, and disabling boundary conditioning reduced ASD from 2.8752 to 2.8307. These differences across the three-seed means show that the effect of conditioning varied by dataset and metric.
Computational and Storage Costs
DCSRL and reliability-guided replay each contained 7,762,597 trainable parameters. DCSRL increased mean training time by 2.82% on CAMUS, 1.13% on CardiacNet, and 15.36% on EchoNet-Dynamic. After the final task, storing the source-reference signatures increased the replay footprint from 1,350.0 to 1,464.3 MB, an additional 114.3 MB (8.47%).
Qualitative Results
Figure 6 shows a recurring pattern across the selected cases. In the multiclass examples, the predictions differ in whether they recover each labeled region and preserve its position relative to neighboring structures. In the LV-only examples, all methods recover the ventricular cavity, and the visible differences concern its outline.

Figure 6. Qualitative comparison of final segmentation predictions after continual echocardiographic training. Two examples from each dataset are shown in order: CAMUS, CardiacNet, and EchoNet-Dynamic. Columns display the input echocardiogram, ground-truth mask, predictions from Plain Replay, Random Replay, iCaRL-style replay, and DCSRL, followed by the DCSRL boundary-error map. White pixels in the final column indicate disagreement between predicted and reference contours.
On CAMUS, Plain Replay fragments the myocardial wall in the upper example and produces disconnected regions in the lower example. Random Replay, iCaRL-style replay, and DCSRL each recover a continuous myocardial band around the ventricular cavity, with the atrium at its base. On CardiacNet, Random Replay omits the upper blue region in the first example, while iCaRL-style replay adds green fragments alongside that region in the second. DCSRL recovers the upper region in both cases and preserves its position relative to the two lower regions. Its contour still differs locally from the reference, particularly along the left edge of the upper region in the second example.
On EchoNet-Dynamic, the methods produce similar LV masks. The DCSRL error maps show residual disagreement along the sloping ventricular contour in the first example and near the apex and base in the second. In both cases, the cavity remains a single connected region.
Ablation Studies
We tested two aspects of DCSRL: the signals used to estimate exemplar drift and the use of that estimate to weight replay losses. Four variants each removed one drift signal. Two further variants disabled drift conditioning for either the boundary or the structural loss while retaining the underlying supervised loss. All variants used the same task order, replay-memory policy, training settings, and random seeds. Reliability-guided replay served as the control with both conditioning branches disabled and is reported in Table 4.
CAMUS.
Full DCSRL achieved the highest mean Dice and boundary F1, although Dice varied by only 0.0014 across all variants (Table 5). The surface-distance measures favored different configurations. Removing probability drift yielded the lowest HD95 (8.1014), while disabling boundary conditioning yielded the lowest ASD (2.8307) and matched the full model’s IoU (0.8172) at the reported precision.
Note: Training time and peak GPU memory are mean sample standard deviation over seeds 42, 123, and 2025. Time overhead denotes the percentage increase relative to reliability-guided replay for each task. Replay storage is the persistent footprint after each task and includes source-task reference signatures for DCSRL.
CardiacNet.
Full DCSRL achieved the highest mean Dice and IoU and the lowest HD95 and ASD. Disabling structural conditioning had a larger effect on surface distance than on overlap: HD95 increased from 10.2148 to 10.6956 and ASD from 3.7972 to 3.8997, while Dice changed from 0.8218 to 0.8213. Removing the structural drift signal yielded the highest mean boundary F1 (0.6000), only 0.0002 above the full model, and increased HD95 to 10.7225.
EchoNet-Dynamic.
The final task showed the most consistent response to ablation. Removing any drift signal or disabling either conditioning branch worsened all five mean metrics. Boundary F1 fell from 0.5283 for full DCSRL to 0.5131–0.5160 across the six variants. Disabling boundary conditioning produced the largest increases in both HD95 and ASD. Thus, the full configuration was favored across all measured aspects of LV segmentation on EchoNet-Dynamic, while the preferred configuration on CAMUS depended on the measure of contour agreement.
Computational Efficiency and Memory Analysis
DCSRL and reliability-guided replay use the same segmentation network with 7,762,597 trainable parameters. They also retain the same exemplars under a common memory budget. DCSRL adds two training costs: storing a source-task signature for each retained exemplar and comparing that signature with the current model response during replay. At inference, both methods use the same segmentation forward pass. The replay buffer and reference signatures are needed only for continued training.
Training time and peak GPU memory.
Across the complete sequence, DCSRL increased mean training time from 6.60 to 6.95 hours, an additional 20.6 minutes (5.20%). EchoNet-Dynamic accounted for 17.2 minutes of this increase. CardiacNet was the longest training stage for both methods, but its relative time overhead was the smallest (1.13%). Peak GPU memory was nearly identical on CAMUS and increased by approximately 0.22 GB on each subsequent task. The highest mean peak for DCSRL was 2.53 GB on CardiacNet as shown in Table 6.
| Task | Method | Training time (s) | Time overhead | Peak GPU memory (GB) | Replay storage (MB) |
| CAMUS | Reliability-guided replay | – | 1200.00 | ||
| DCSRL | +2.82% | 1301.60 | |||
| CardiacNet | Reliability-guided replay | – | 1350.00 | ||
| DCSRL | +1.13% | 1464.30 | |||
| EchoNet-Dynamic | Reliability-guided replay | – | 1350.00 | ||
| DCSRL | +15.36% | 1464.30 |
Table 6. Computational cost and replay-storage requirements during sequential training.
Persistent replay storage.
The buffer held 1,600 exemplars after CAMUS and reached its capacity of 1800 after CardiacNet. At capacity, the source-task signatures added 114.30 MB to the baseline footprint of 1350.00 MB, an increase of 8.47%. The buffer remained at 1464.30 MB after EchoNet-Dynamic as new exemplars replaced existing entries.
Discussion
The comparison with reliability-guided replay shows that the supervision applied to a stored exemplar can influence final segmentation quality even when the network, memory capacity, and sampling policy are held fixed. DCSRL increased the mean final Dice across datasets from 0.8718 to 0.8784. Its clearest benefit appeared on CardiacNet, where the model was evaluated after subsequent training on EchoNet-Dynamic. Dice increased from 0.8096 to 0.8218, while HD95 and ASD decreased by 36.3% and 27.1%, respectively (Table 4). Thus, a modest change in regional overlap accompanied a substantial reduction in the distance between predicted and reference contours.
Final performance and transfer.
The higher final Dice and the lower transfer scores describe different aspects of the training trajectory. Backward transfer measures change relative to each method’s own performance immediately after learning a task. DCSRL’s near-zero backward transfer (0.0003) indicates that, on average, its final scores on earlier tasks remained close to those initial values. Reliability-guided replay showed a larger subsequent gain (0.0489), even though its final Dice was lower on each dataset. It also achieved higher forward transfer (0.4366 versus 0.4080), which in this study measures performance on a future dataset before training on it (Table 3). The evidence for DCSRL is therefore strongest for segmentation quality at the end of the evaluated sequence.
What the ablations reveal.
The contribution of drift conditioning depended on both the dataset and the measure of contour agreement. On CardiacNet, disabling structural conditioning worsened all five mean metrics. Removing the structural drift signal produced a different outcome: boundary F1 increased by only 0.0002, while HD95 and ASD both increased (Table 5). The distinction matters because removing one input to the drift estimate still allows the remaining signals to regulate structural supervision. On CAMUS, the full model achieved the highest mean boundary F1, yet several reduced variants had lower surface distances. Boundary F1 measures agreement within a fixed tolerance, whereas HD95 captures the upper tail of contour displacement. A gain in one can therefore coexist with a decline in the other. With three training seeds, small differences between variants warrant a descriptive interpretation.
EchoNet-Dynamic showed the most consistent ablation pattern: removing any drift signal or disabling either conditioning branch worsened all five mean metrics. Since EchoNet-Dynamic supplies LV-only supervision, the structural replay term acts on eligible historical exemplars during this stage. Its association with final LV performance suggests that the way earlier examples are supervised can also influence learning on the incoming domain.
Interpreting source-referenced drift.
A frozen signature records what the model predicted when an exemplar entered memory. A later departure from that prediction can reflect either deterioration or correction of an earlier error. DCSRL uses the departure to set the weight of the supervised replay loss, while the stored annotation supplies the target. This distinction also identifies a useful test of the mechanism: whether larger drift predicts a subsequent increase in annotation-based error for the same exemplar. Comparing DCSRL with a constant-weight control matched for average supervision strength would further clarify how much of the gain comes from assigning additional weight to particular exemplars.
Practical implications.
The additional cost arises during model updating. Across the full sequence, DCSRL increased training time by 5.20% and required 114.3 MB of additional storage at the 1800-exemplar capacity (Table 6). Both methods use the same segmentation network at inference. The observed contour improvements motivate a downstream evaluation of measurements such as ventricular volume, for which the location of a segmentation error may matter as much as its contribution to overall overlap.
Limitations and Future Work
Task order and annotation scope.
We evaluated three echocardiographic datasets in one order: CAMUS CardiacNet EchoNet-Dynamic. Because the LV-only task always appeared last, its position in the sequence was coupled to a change in annotation scope. The present design therefore cannot separate the effects of task order, image distribution, and available labels. Repeating the experiment across all six task permutations would test sensitivity to order. A separate experiment using the same CAMUS images with either full or LV-only annotations would isolate the effect of narrowing supervision while holding image appearance fixed.
Interpretation of drift.
A frozen source-task signature records the model’s response when an exemplar enters memory. A subsequent departure can reflect an emerging segmentation error or a correction to an inaccurate source prediction. The current drift score gives both changes a positive magnitude. Comparing exemplar-level drift with changes in error against the stored annotation would show when drift identifies deterioration that warrants stronger supervision. A constant-weight control matched to DCSRL’s mean boundary and structural weights would also test how much of the benefit comes from assigning supervision to particular exemplars.
Sensitivity of the conditioning rule.
The drift weights, normalization thresholds, and conditioning gains were fixed across tasks. The ablations in Table 5 test the effect of removing individual components, but leave open how sensitive performance is to their weights. The structural descriptor also summarizes each class through contour disagreement and connected-component count. A myocardial mask can retain its connectivity while its wall thickness changes locally. A useful extension would compare the current descriptor with a wall-thickness or signed-distance measure on cases where the two descriptors disagree.
Statistical and architectural scope.
Three training seeds characterize run-to-run variation under the chosen protocol. They provide limited evidence for ranking variants with nearly identical means, and they do not quantify uncertainty across patients or institutions. For example, removing structural drift changed mean CardiacNet boundary F1 from 0.5998 to 0.6000. Additional seeds and paired patient-level confidence intervals would help distinguish training variability from variation across cases. The experiments also used a single U-Net backbone. Since feature drift is measured at its pooled bottleneck, testing a transformer encoder under the same replay budget would show whether the current normalization transfers to a different feature representation.
Replay storage and scale.
DCSRL assumes continued access to stored images, annotations, and source-task signatures. At the same 1800-exemplar capacity as reliability-guided replay, the signatures increased persistent storage by 8.47%. This leaves two practical questions. First, would DCSRL retain its advantage under an equal storage budget in bytes, where signature storage competes with space for additional exemplars? Second, how would it perform over a longer sequence as a fixed buffer must represent more domains? Evaluating both conditions on additional institutions would test the method under tighter memory constraints and a broader range of acquisition shifts.
Conclusion
This study introduced Drift-Conditioned Structural Replay Loss (DCSRL) for continual echocardiographic segmentation under heterogeneous imaging distributions and annotation spaces. DCSRL builds upon a reliability-guided replay strategy while introducing an exemplar-adaptive supervision mechanism that responds to changes in previously acquired knowledge. By comparing current model representations, class probabilities, boundary responses, and predicted anatomical structures with frozen source-task references, DCSRL dynamically regulates boundary and structural preservation during sequential learning. Experimental evaluations across CAMUS, CardiacNet, and EchoNet-Dynamic demonstrated that drift-conditioned supervision improves final segmentation performance and anatomical boundary preservation compared with conventional continual-learning approaches. The findings also revealed trade-offs between final segmentation accuracy and knowledge transfer, emphasizing that stronger anatomical preservation does not necessarily improve every aspect of continual learning. Importantly, DCSRL achieves these improvements without introducing additional trainable parameters or modifying the inference-time architecture, although source-reference storage and drift estimation increase training-time computational requirements. Future research will investigate the robustness of DCSRL across alternative task orders, imaging institutions, and annotation configurations. Further work will also examine the relationship between measured knowledge drift and subsequent segmentation errors to determine when stronger anatomical supervision is beneficial. Extending drift-conditioned replay to other medical imaging modalities and investigating more efficient reference-storage strategies could further improve its applicability to continual medical image analysis.
Declaration of Generative AI and AI-assisted technologies in the writing process
During the preparation of this work, the authors used ChatGPT to assist with language editing and improve the readability of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
CRediT authorship contribution statement
Zahid Ullah: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing - original draft, Writing - review & editing. Jaeyoon Seo: Writing - original draft, Writing - review & editing. Jihie Kim: Formal analysis, Investigation, Supervision, Project administration, Project management.
Declaration of Competing Interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability
The datasets analyzed in this study were obtained from publicly available research resources. CAMUS is accessible at (https://www.creatis.insa-lyon.fr/Challenge/camus/). CardiacNet is available through (https://github.com/xmed-lab/CardiacNet), subject to the dataset provider’s access instructions. EchoNet-Dynamic is available at (https://echonet.github.io/dynamic/), subject to registration and the applicable research-use agreement. No new clinical imaging dataset was collected for this study.
Acknowledgments
This work was supported by the Dongguk University Research Fund of 2026. This research was supported by the MSIT(Ministry of Science and ICT), Korea, under the ITRC(Information Technology Research Center) support program(IITP-2026-RS-2020-II201789), and the Artificial Intelligence Convergence Innovation Human Resources Development(IITP-2026-RS-2023-00254592) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation).
References
- [1]Leclerc S, Smistad E, Pedrosa J, Østvik A, Cervanensky F, Espinosa F, et al. Deep learning for segmentation using an open large-scale dataset in 2D echocardiography. IEEE transactions on medical imaging. 2019;38(9):2198–2210.DOI
- [2]Deng X, Wu H, Zeng R, Qin J. Memsam: Taming segment anything model for echocardiography video segmentation. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2024. p. 9622–9631.DOI
- [3]Ouyang D, He B, Ghorbani A, Yuan N, Ebinger J, Langlotz CP, et al. Video-based AI for beat-to-beat assessment of cardiac function. Nature. 2020;580(7802):252–256.DOI
- [4]Parisi GI, Kemker R, Part JL, Kanan C, Wermter S. Continual lifelong learning with neural networks: A review. Neural networks. 2020;113:54–71.DOI
- [5]De Lange M, Aljundi R, Masana M, Parisot S, Jia X, Leonardis A, et al. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence. 2021;44(7):3366–3385.DOI
- [6]Wu X, Xu Z, Tong RKy. Continual long term forgetting in medical image analysis: A survey. Computers in Biology and Medicine. 2024;182:109206.
- [7]Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences. 2017;114(13):3521–3526.DOI
- [8]Li Z, Hoiem D. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence. 2017;40(12):2935–2947.
- [9]Perkonigg M, Hofmanninger J, Herold CJ, Brink JA, Pianyhk O, Prosch H, et al. Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging. Nature communications. 2021;12(1):5678.DOI
- [10]Bera S, Ummadi V, Sen D, Mandal S, Biswas PK. Memory replay for continual medical image segmentation through atypical sample selection. In: International conference on medical image computing and computer-assisted intervention. Springer; 2023. p. 513–522.DOI
- [11]Ullah Z, Choi S, Kim J. JBR-Net: Boundary-Balanced Replay for Continual Medical Image Segmentation. arXiv preprint arXiv:260614731. 2026;.DOI
- [12]Kervadec H, Bouchtiba J, Desrosiers C, Granger E, Dolz J, Ayed IB. Boundary loss for highly unbalanced segmentation. Medical image analysis. 2021;67:101851.
- [13]Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. Springer; 2015. p. 234–241.arxiv.org/abs/1505.04597
- [14]Isensee F, Jaeger PF, Kohl SA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods. 2021;18(2):203–211.DOI
- [15]Oktay O, Schlemper J, Folgoc LL, Lee M, Heinrich M, Misawa K, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:180403999. 2018;arxiv.org/abs/1804.03999
- [16]Chen J, Lu Y, Yu Q, Luo X, Adeli E, Wang Y, et al. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:210204306. 2021;arxiv.org/abs/2102.04306
- [17]Cao H, Wang Y, Chen J, Jiang D, Zhang X, Tian Q, et al. Swin-unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. Springer; 2022. p. 205–218.arxiv.org/abs/2105.05537
- [18]Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, et al. Segment anything. In: 2023 IEEE/CVF international conference on computer vision (ICCV). IEEE; 2023. p. 3992–4003.arxiv.org/abs/2304.02643
- [19]Ma J, He Y, Li F, Han L, You C, Wang B. Segment anything in medical images. Nature Communications. 2024;15:654.
- [20]Leclerc S, Smistad E, Østvik A, Cervanensky F, Espinosa F, Espeland T, et al. LU-Net: a multistage attention network to improve the robustness of segmentation of left ventricular structures in 2-D echocardiography. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control. 2020;67(12):2519–2530.DOI
- [21]Yang J, Lin Y, Pu B, Guo J, Xu X, Li X. Cardiacnet: Learning to reconstruct abnormalities for cardiac disease assessment from echocardiogram videos. In: European Conference on Computer Vision. Springer; 2024. p. 293–311.arxiv.org/abs/2410.20769
- [22]Oktay O, Ferrante E, Kamnitsas K, Heinrich M, Bai W, Caballero J, et al. Anatomically constrained neural networks (ACNNs): application to cardiac image enhancement and segmentation. IEEE transactions on medical imaging. 2017;37(2):384–395.DOI
- [23]Painchaud N, Skandarani Y, Judge T, Bernard O, Lalande A, Jodoin PM. Cardiac segmentation with strong anatomical guarantees. IEEE transactions on medical imaging. 2020;39(11):3703–3713.DOI
- [24]Painchaud N, Duchateau N, Bernard O, Jodoin PM. Echocardiography segmentation with enforced temporal consistency. IEEE transactions on medical imaging. 2022;41(10):2867–2878.DOI
- [25]Wang L, Zhang X, Su H, Zhu J. A comprehensive survey of continual learning: Theory, method and application. IEEE transactions on pattern analysis and machine intelligence. 2024;46(8):5362–5383.DOI
- [26]Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. Journal of big Data. 2021;8(1):53.DOI
- [27]Gama J, Žliobaitė I, Bifet A, Pechenizkiy M, Bouchachia A. A survey on concept drift adaptation. ACM computing surveys (CSUR). 2014;46(4):1–37.
- [28]Lu J, Liu A, Dong F, Gu F, Gama J, Zhang G. Learning under concept drift: A review. IEEE transactions on knowledge and data engineering. 2018;31(12):2346–2363.DOI
- [29]Zenke F, Poole B, Ganguli S. Continual learning through synaptic intelligence. In: International conference on machine learning. Pmlr; 2017. p. 3987–3995.arxiv.org/abs/1703.04200
- [30]Aljundi R, Babiloni F, Elhoseiny M, Rohrbach M, Tuytelaars T. Memory aware synapses: Learning what (not) to forget. In: European conference on computer vision. Springer; 2018. p. 144–161.arxiv.org/abs/1711.09601
- [31]Michieli U, Zanuttigh P. Incremental learning techniques for semantic segmentation. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). IEEE; 2019. p. 3205–3212.arxiv.org/abs/1907.13372
- [32]Michieli U, Zanuttigh P. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2021. p. 1114–1124.arxiv.org/abs/2103.06342
- [33]Cermelli F, Mancini M, Bulo SR, Ricci E, Caputo B. Modeling the background for incremental learning in semantic segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2020. p. 9230–9239.arxiv.org/abs/2002.00718
- [34]Douillard A, Chen Y, Dapogny A, Cord M. Plop: Learning without forgetting for continual semantic segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2021. p. 4039–4049.arxiv.org/abs/2011.11390
- [35]Baweja C, Glocker B, Kamnitsas K. Towards Continual Learning in Medical Imaging. arXiv preprint arXiv:181102496. 2018;.arxiv.org/abs/1811.02496
- [36]Ozdemir F, Fuerstahll P, Goksel O. Learn the new, keep the old: Extending pretrained models with new anatomy and images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer; 2018. p. 361–369.DOI
- [37]Gonzalez C, Lemke N, Ranem A, Sakas G, Mukhopadhyay A. What is wrong with continual learning in medical image segmentation? In: Proceedings of the International Workshop on Personalized Incremental Learning in Medicine; 2025. p. 25–34.arxiv.org/abs/2010.11008
- [38]González C, Ranem A, Pinto dos Santos D, Othman A, Mukhopadhyay A. Lifelong nnU-Net: a framework for standardized medical continual learning. Scientific Reports. 2023;13(1):9381.
- [39]Zhang Y, Li X, Chen H, Yuille AL, Liu Y, Zhou Z. Continual learning for abdominal multi-organ and tumor segmentation. In: International conference on medical image computing and computer-assisted intervention. Springer; 2023. p. 35–45.arxiv.org/abs/2306.00988
- [40]Ji Z, Guo D, Wang P, Yan K, Lu L, Xu M, et al. Continual segment: Towards a single, unified and non-forgetting continual segmentation model of 143 whole-body organs in ct scans. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE; 2023. p. 21083–21094.DOI
- [41]Kiyasseh D, Zhu T, Clifton D. A clinical deep learning framework for continually learning from cardiac signals across diseases, time, modalities, and institutions. Nature Communications. 2021;12(1):4221.DOI
- [42]Lenga M, Schulz H, Saalbach A. Continual learning for domain adaptation in chest x-ray classification. In: Medical Imaging with Deep Learning. PMLR; 2020. p. 413–423.arxiv.org/abs/2001.05922
- [43]Zhang J, Gu R, Xue P, Liu M, Zheng H, Zheng Y, et al. S 3 R: Shape and semantics-based selective regularization for explainable continual segmentation across multiple sites. IEEE Transactions on Medical Imaging. 2023;42(9):2539–2551.DOI
- [44]Li W, Zhang Y, Zhou H, Yang W, Xie Z, He Y. CLMS: Bridging domain gaps in medical imaging segmentation with source-free continual learning for robust knowledge transfer and adaptation. Medical Image Analysis. 2025;100:103044.DOI
- [45]Caccia L, Aljundi R, Asadi N, Tuytelaars T, Pineau J, Belilovsky E. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:210405025. 2021;.
- [46]Rolnick D, Ahuja A, Schwarz J, Lillicrap T, Wayne G. Experience replay for continual learning. Advances in neural information processing systems. 2019;32.
- [47]Chaudhry A, Rohrbach M, Elhoseiny M, Ajantán T, Dokania PK, Torr PH, et al. On tiny episodic memories in continual learning. arXiv preprint arXiv:190210486. 2019;.arxiv.org/abs/1902.10486
- [48]Lopez-Paz D, Ranzato M. Gradient episodic memory for continual learning. Advances in neural information processing systems. 2017;30.
- [49]Chaudhry A, Ranzato M, Rohrbach M, Elhoseiny M. Efficient lifelong learning with a-gem. arXiv preprint arXiv:18120420. 2018;.arxiv.org/abs/1812.00420
- [50]Shin H, Lee JK, Kim J, Kim J. Continual learning with deep generative replay. Advances in neural information processing systems. 2017;30.arxiv.org/abs/1705.08690
- [51]Rebuffi SA, Kolesnikov A, Sperl G, Lampert CH. icarl: Incremental classifier and representation learning. In: 2017 IEEE conference on computer vision and pattern recognition (CVPR). IEEE; 2017. p. 5533–5542.arxiv.org/abs/1611.07725
- [52]Aljundi R, Lin M, Goujaud B, Bengio Y. Gradient based sample selection for online continual learning. Advances in neural information processing systems. 2019;32.arxiv.org/abs/1903.08671
- [53]Bang J, Kim H, Yoo Y, Ha JW, Choi J. Rainbow memory: Continual learning with a memory of diverse samples. In: 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE; 2021. p. 8214–8223.arxiv.org/abs/2103.17230
- [54]Prabhu A, Torr PH, Dokania PK. Gdumb: A simple approach that questions our progress in continual learning. In: European conference on computer vision. Springer; 2020. p. 524–540.DOI
- [55]Zhu L, Chen T, Yin J, See S, Liu J. Continual semantic segmentation with automatic memory sample selection. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2023. p. 3082–3092.arxiv.org/abs/2304.05015
- [56]Ranem A, González C, dos Santos DP, Bucher AM, Othman AE, Mukhopadhyay A. Continual atlas-based segmentation of prostate MRI. In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE; 2024. p. 7548–7557.arxiv.org/abs/2311.00548
- [57]Aljundi R, Belilovsky E, Tuytelaars T, Charlin L, Caccia M, Lin M, et al. Online continual learning with maximal interfered retrieval. Advances in neural information processing systems. 2019;32.
- [58]Buzzega P, Boschini M, Porrello A, Abati D, Calderara S. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems. 2020;33:15920–15930.arxiv.org/abs/2004.07211
- [59]Boschini M, Bonicelli L, Buzzega P, Porrello A, Calderara S. Class-incremental continual learning into the extended der-vese. IEEE transactions on pattern analysis and machine intelligence. 2022;45(5):5497–5512.DOI
- [60]Arani E, Sarfraz F, Zonooz B. Learning fast, learning slow: A general continual learning method based on complementary learning system. arXiv preprint arXiv:220112604. 2022;.arxiv.org/abs/2201.12604
- [61]Ma J, Chen J, Ng M, Huang R, Li Y, Li C, et al. Loss odyssey in medical image segmentation. Medical image analysis. 2021;71:102035.DOI
- [62]Karimi D, Salcudean SE. Reducing the hausdorff distance in medical image segmentation with convolutional neural networks. IEEE Transactions on medical imaging. 2019;39(2):499–513.arxiv.org/abs/1904.10030
- [63]Lin Q, Chen X, Chen C, Garibaldi JM. Boundary-wise loss for medical image segmentation based on fuzzy rough sets. Information Sciences. 2024;661:120183.DOI
- [64]Kang S, Song J, Kim J. Advancing medical image segmentation: morphology-driven learning with diffusion transformer. arXiv preprint arXiv:240800347. 2024;.arxiv.org/abs/2408.00347
- [65]Hu X, Li F, Samaras D, Chen C. Topology-preserving deep image segmentation. Advances in neural information processing systems. 2019;32.arxiv.org/abs/1906.05404
- [66]Clough JR, Byrne N, Oksuz I, Zimmer VA, Schnabel JA, King AP. A topological loss function for deep-learning based image segmentation using persistent homology. IEEE transactions on pattern analysis and machine intelligence. 2020;44(12):8766–8778.DOI
- [67]Shit S, Paetzold JC, Sekubouinya A, Ezhov I, Unger A, Zhylka A, et al. clDice-a novel topology-preserving loss function for tubular structure segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2021. p. 16555–16564.DOI


