acceptodds
Under review as a conference paper at ICLR 2027

FreeEcho: Learning Prompt-Free Echocardiography Video Segmentation via Distillation

Abstract

Echocardiography video segmentation delineates cardiac structures across frames to support quantitative assessment. Prompt-free models enable automated segmentation but remain limited in accuracy and generalization across data sources, whereas prompt-based models achieve stronger performance by relying on external localization inputs that hinder scalable inference. We present FreeEcho, a knowledge distillation framework that converts existing prompt-based echocardiography video segmentation models into prompt-free counterparts. FreeEcho adapts the model prediction pipeline, combines ground-truth supervision with teacher-derived distillation signals, and employs a prompt removal curriculum during training. Experiments across multiple models and public datasets show that the distilled students consistently approach or achieve their teachers' performance while substantially outperforming direct prompt removal. They also outperform existing prompt-free models, with clear advantages in generalization across different data sources. Ablation studies clarify the effects of distillation design choices.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Echocardiography Video Segmentation, Prompt-Free Segmentation, Knowledge Distillation, Medical AI, Computer Vision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_K5xSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FreeEcho, a knowledge distillation framework for prompt-free echocardiography video segmentation. It transfers knowledge from prompt-based teacher models to students that do not require prompts during inference. The method combines multiple distillation losses and a progressive prompt-removal strategy. Experiments on four public datasets demonstrate improved segmentation performance and cross-dataset generalization.

Critical strengths
  1. The paper addresses a practical problem. Removing the need for manual prompts is important for automatic echocardiography analysis.

  2. The proposed framework can be applied to both per-frame and memory-based segmentation models, showing good flexibility.

  3. The experiments are relatively comprehensive, covering different models, multiple datasets, and cross-dataset evaluation.

  4. The ablation studies demonstrate the contributions of different distillation losses and the prompt-removal strategy.

Critical weaknesses
  1. Limited novelty in the distillation strategy. The main techniques, including knowledge distillation, feature alignment, and curriculum learning, have been widely studied. The authors should better explain what is technically new compared with existing prompt-free distillation methods.

  2. The benefits of different supervision signals need further analysis. Several distillation losses are used, but it is not fully clear how much each component contributes to the final performance. The authors should provide experiments with simpler combinations of losses and discuss the training complexity.

  3. Dependence on teacher models. The framework relies on well-performing prompt-based teachers. It is unclear whether the method remains effective when teacher predictions are inaccurate. The authors should evaluate students trained with different teacher quality levels or noisy teacher predictions.

  4. Limited analysis of temporal consistency. The paper focuses mainly on Dice scores, but temporal stability is also important for video segmentation. The authors should report temporal consistency metrics and provide examples of segmentation failures across video frames.

Questions for authors
  1. What are the main differences between FreeEcho and existing knowledge distillation methods?

  2. How does the performance change when using a weaker teacher model?

  3. Can the authors provide more analysis of temporal consistency in video segmentation?

3 comments

Q1. What are the main differences between FreeEcho and existing knowledge distillation methods?

FreeEcho organizes distillation around where prompt information enters the prediction pipeline and how it is propagated. Soft-mask supervision transfers the final prediction, while architecture-specific supervision transfers the intermediate localization or memory representations that support that prediction. This connects the prompt-free interface, the auxiliary objectives, and the prompt-removal curriculum in a single conversion framework.

On a per-frame model such as EchoONE, the prompt is a sparse token read independently on every frame. The student has the same prompt encoder and decoder, so the deployed interface is that decoder run with an empty point. The prompt’s useful content is the soft left-ventricle boundary and the coarse PC-Mask prior that was composed with the token. FreeEcho matches those two outputs with ground-truth segmentation and temperature-scaled mask distillation, and it withdraws the student’s prompt on a curriculum. Coarse-map supervision targets the intermediate localization behavior that the prompt-free student must recover from the image.

For memory-based models, FreeEcho uses the prompt-free interface in Section 3.2 to initialize memory from the student’s initial prediction and preserve subsequent propagation. It aligns object-level memory representations after the initial frame using cosine similarity (Equation 8). This supplies supervision beyond the predicted masks while preserving the model’s memory-based prediction pipeline.

FreeEcho retains the backbone prediction pipeline while coupling prompt-free adaptation with architecture-specific supervision. In-domain comparisons show the benefit across all four datasets. EchoONE reaches 0.937 / 0.926 / 0.908 / 0.867 on CAMUS, EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA, compared with 0.928 / 0.901 / 0.872 / 0.781 for the adapted DistillSAM objective and 0.925 / 0.919 / 0.900 / 0.842 for KnowSAM. MedSAM2 reaches 0.929 / 0.919 / 0.903 / 0.857, compared with 0.918 / 0.912 / 0.892 / 0.839 and 0.915 / 0.911 / 0.888 / 0.392. Together with the zero-shot results and auxiliary-loss ablations, these comparisons establish the effectiveness of the proposed prompt-free conversion framework across per-frame and memory-based architectures.

Q2. How does the performance change when using a weaker teacher model?

A weaker teacher lowers the empty-prompt student's absolute Dice to that teacher's own level. Whether the student also stays at that level depends on the prompt. Per-frame students reach the teacher's ceiling. Memory students reach it on CAMUS, then fall below the clicked teacher once the first-frame click must be retained across a domain shift.

Table. Distillation from weaker teachers. Same architecture throughout.

TeacherPromptCAMUSEchoNetPediatricCardiacUDA
EchoONEper-frame click0.939 → 0.9370.898 → 0.8980.877 → 0.8670.673 → 0.614
Medical-SAM2per-frame click0.920 → 0.9220.864 → 0.8660.838 → 0.8400.533 → 0.524
MedSAM2ED click0.937 → 0.9290.900 → 0.8780.884 → 0.8630.739 → 0.535
memSAMED click0.923 → 0.9270.802 → 0.7680.733 → 0.6440.397 → 0.378

For a per-frame click, the student saturates the teacher it is given. Medical-SAM2, the weaker frame-wise teacher, stays within 0.01 on every set (0.920 → 0.922, 0.864 → 0.866, 0.838 → 0.840, 0.533 → 0.524). EchoONE keeps the click score on EchoNet (0.898 → 0.898), loses 0.002 on CAMUS and 0.010 on Pediatric, and loses 0.059 on CardiacUDA (0.673 → 0.614). The weaker per-frame teacher does not add a further drop.

For an ED click written into memory, CAMUS still reaches the teacher: MedSAM2 0.937 → 0.929 and memSAM 0.923 → 0.927. Out of domain, both students fall below the clicked teacher, and the extra gap is not ordered by teacher strength. memSAM is the weak memory teacher (0.802 / 0.733 / 0.397), and its student is lower still (0.768 / 0.644 / 0.378): the added drops are 0.034, 0.089, and 0.019. MedSAM2 is the strong memory teacher, yet its largest added drop is CardiacUDA, 0.739 → 0.535 (−0.204), while EchoNet and Pediatric lose only 0.022 and 0.021. The absolute score still follows the teacher: the memSAM student remains below the MedSAM2 student on all three transfer sets.

Q3. Can the authors provide more analysis of temporal consistency in video segmentation?

We thank the reviewer for this question. The Dice values in Table 1 are computed only on annotated end-diastolic (ED) and end-systolic (ES) frames, so they do not describe whether the segmentation stays consistent between those two frames.

We therefore evaluated temporal consistency on the CAMUS test set, where every frame of the ED-to-ES half-sequence has a left-ventricular mask. EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA do not provide dense masks over this interval, so the same curve comparison cannot be repeated there. The comparison covers 100 half-sequences from 50 patients in the two-chamber and four-chamber views. For each sequence we compare the predicted left-ventricular area curve with the ground-truth area curve, and we compute Dice on every frame, which is different from ED/ES Dice in Table 1.

ArchitectureMethodPearson rSequences with r > 0.9Relative area errorFrame DiceES area < ED area
EchoONEFreeEcho0.92393%7.4%0.93599%
EchoONEDistillSAM0.91791%8.3%0.92899%
EchoONEKnowSAM0.91990%8.4%0.92598%
MedSAM2FreeEcho0.93294%8.8%0.92298%
MedSAM2DistillSAM0.93293%9.3%0.91098%
MedSAM2KnowSAM0.92990%10.2%0.90498%

Pearson r is the correlation between the predicted and ground-truth area curves of one sequence. Relative area error is the mean of |predicted area − ground-truth area| / ground-truth area. Frame Dice is averaged over all frames of a half-sequence and then over the 100 sequences. The last column is the fraction of sequences in which the predicted ES area is smaller than the predicted ED area.

On EchoONE, FreeEcho reaches a Pearson r of 0.923, with r > 0.9 in 93% of sequences, a relative area error of 7.4%, and a frame Dice of 0.935. The predicted cavity becomes smaller from ED to ES in 99% of sequences. On MedSAM2, the corresponding values are 0.932, 94%, 8.8%, 0.922, and 98%. Under the same protocol, FreeEcho has a lower relative area error than DistillSAM and KnowSAM on both architectures, and a higher frame Dice, most clearly on MedSAM2 (0.922, compared with 0.910 and 0.904).

Reviewer_PSPhOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FreeEcho, a knowledge-distillation framework for converting prompt-based echocardiography video segmentation models into prompt-free counterparts. A frozen prompt-based teacher receives external localization prompts during training, while the student is progressively trained to operate without such prompts. FreeEcho combines ground-truth segmentation supervision, teacher soft-mask distillation, architecture-specific auxiliary supervision, and a prompt-removal curriculum. For per-frame models, intermediate coarse-mask predictions are distilled as spatial supervision, whereas for memory-based models, object-level memory representations are aligned as temporal supervision. The framework is instantiated with two different teacher architectures and evaluated on four public echocardiography datasets under both in-domain and CAMUS-to-other-source zero-shot transfer settings.

Critical strengths
  1. The paper addresses a practically meaningful deployment gap and evaluates it under a strong cross-source protocol. Removing the need for external localization prompts is useful for scalable echocardiography analysis. The zero-shot evaluation, in which all task-specific models are trained or adapted on CAMUS and then evaluated on multiple unseen data sources without target-data training or checkpoint selection, is particularly valuable.

  2. The empirical study is broad and the architecture-specific distillation signals appear consequential. The framework is evaluated with both per-frame and memory-based teacher architectures, against direct prompt removal, existing distillation approaches, and prompt-free baselines. The ablations consistently favor the full objective and the prompt-removal curriculum. In particular, the strong degradation of the memory-based student when temporal-memory distillation is removed provides useful evidence that transferring architecture-specific internal representations is important.

Critical weaknesses
  1. The methodological novelty appears relatively incremental at the level of the underlying learning principle. FreeEcho can largely be viewed as privileged-input distillation: the teacher receives localization information unavailable to the deployed student, while the student combines ground-truth supervision, teacher prediction/intermediate-representation distillation, and a curriculum that progressively removes the privileged input. These ingredients are individually well established. The architecture-specific spatial/temporal losses and their application to prompt removal in echocardiography are useful, but I am not yet convinced that the paper establishes a sufficiently general new distillation principle rather than an effective task-specific adaptation of existing ideas.

  2. Some of the broad comparison and generalization claims are stronger than the experimental controls support. The comparisons against existing prompt-free models involve substantially different architectures, parameter counts, and pretraining. For example, the EchoONE-based student inherits a much larger SAM-based architecture than several prompt-free baselines. Thus, its superior accuracy cannot be attributed to FreeEcho alone. The more convincing evidence is the matched comparison among the same teacher/student architecture under direct prompt removal and different distillation strategies. In addition, the reported standard deviations are computed across annotated frames rather than independent training runs or patients/videos; frames from the same video are correlated. Therefore, the current results establish strong empirical trends but not statistical significance or optimization robustness.

Questions for authors

Could the authors provide repeated-seed or patient/video-level uncertainty estimates for the main comparisons? The current mean ± SD values characterize variation across annotated frames, which are correlated within videos, rather than training variability or confidence in the differences between methods.

How were DistillSAM and KnowSAM adapted to the EchoONE and MedSAM2 teacher/student architectures? Were optimization schedules, initialization, training duration, available ground-truth supervision, and teacher outputs matched to FreeEcho? More implementation detail would help establish that the advantage over these distillation baselines comes from the proposed distillation design rather than different adaptation protocols.

The paper emphasizes superiority over existing prompt-free models, but those models differ substantially in model capacity and pretraining. Could the authors either provide a more capacity/pretraining-matched prompt-free control or narrow this claim to emphasize the stronger same-architecture comparisons against direct prompt removal and alternative distillation strategies?

3 comments

Q1. Could the authors provide repeated-seed or patient/video-level uncertainty estimates for the main comparisons? The current mean ± SD values characterize variation across annotated frames, which are correlated within videos, rather than training variability or confidence in the differences between methods.

We thank the reviewer for requesting an estimate of training variability. The standard deviations in the submitted tables are computed across annotated frames. Frames from the same video are correlated, so those values describe dispersion within a test set rather than variability from optimization. We repeated the matched KnowSAM, DistillSAM, and FreeEcho training at seeds 7, 1234, and 2026. Each entry below is the mean of the three seed-wise Dice means, and the standard deviation is the sample standard deviation across those three seeds. Empty-prompt Dice. One seed is mean ± std over frames. Three seeds (7, 1234, 2026) are the mean of the three seed-wise means ± std across seeds. Columns are CAMUS, EchoNet, Pediatric, and CardiacUDA.

Table S1. EchoONE, one seed versus three seeds.

MethodSeedsCAMUSEchoNetPediatricCardiacUDA
KnowSAM30.921±0.0040.867±0.0070.843±0.0110.539±0.012
DistillSAM30.928±0.0010.885±0.0020.833±0.0070.590±0.055
FreeEcho (ours)30.938±0.0010.897±0.0040.865±0.0010.608±0.006

Table S2. MedSAM2, one seed versus three seeds.

MethodSeedsCAMUSEchoNetPediatricCardiacUDA
KnowSAM30.912±0.0040.865±0.0060.856±0.0100.507±0.014
DistillSAM30.918±0.0020.869±0.0030.848±0.0090.461±0.004
FreeEcho (ours)30.927±0.0040.878±0.0020.863±0.0010.540±0.009

Averaging three seeds changes the reported means only slightly. Relative to the original single run, every EchoONE entry moves by at most 0.010, and every MedSAM2 entry moves by at most 0.008. FreeEcho moves by at most 0.006. DistillSAM is unchanged on CAMUS, EchoNet, and Pediatric for EchoONE, and its CardiacUDA mean changes from 0.596 to 0.590.

FreeEcho remains above both distillation baselines on all four sets for both students. On EchoONE, its margin over DistillSAM is 0.010, 0.012, 0.032, and 0.018. On MedSAM2, the corresponding margins are 0.009, 0.009, 0.015, and 0.079. These margins are larger than the seed standard deviations of the two methods, except on EchoONE CardiacUDA, where DistillSAM varies by 0.055 across seeds and the 0.018 gap should be read with that spread.

The repeated-seed standard deviations are much smaller than the original frame-wise standard deviations because they estimate a different quantity. On CAMUS, EchoNet, and Pediatric, the seed standard deviation is at most 0.011. FreeEcho is the most stable student on these sets, with a seed standard deviation of 0.001 to 0.004. The larger sensitivity is confined to CardiacUDA, especially EchoONE DistillSAM (0.590±0.055).

Q2. How were DistillSAM and KnowSAM adapted to the EchoONE and MedSAM2 teacher/student architectures? Were optimization schedules, initialization, training duration, available ground-truth supervision, and teacher outputs matched to FreeEcho? More implementation detail would help establish that the advantage over these distillation baselines comes from the proposed distillation design rather than different adaptation protocols.

DistillSAM keeps the original EchoONE or MedSAM2 decoder. The teacher is frozen and receives a centroid click. On EchoONE, the student still follows the same prompt-annealing schedule and ends with an empty prompt. Three losses are added to the original segmentation loss: mean squared error on the mask tokens, mean squared error on the dense prompt embedding against the ground-truth mask encoded at 0/1, and binary cross-entropy plus Dice on the attention map. On MedSAM2, memory propagation is unchanged. Because the model does not return its internal tokens, hooks record the mask tokens, the dense prompt, and the attention. The student uses a point prompt for the first two epochs and then an empty point on the first frame, while later frames continue to use memory. The same three losses are used. Segmentation supervision is applied only to end-diastole and end-systole frames that contain foreground.

KnowSAM does not distill prompt tokens. On EchoONE, its two views are the fine mask and the PC-Mask coarse mask from the same network. Each view is trained with segmentation loss, KL divergence toward the teacher mask, and an entropy penalty. The two views are also trained to agree with each other, and their fused prediction is supervised by the ground-truth mask. The prompt schedule and the decoder are the same as in FreeEcho. On MedSAM2, a U-Net and a V-Net supply the two views, and small adapters are inserted into Hiera. A point head converts the first-frame embedding into a point prompt, and the fused foreground of the two views is used as the first-frame dense prompt. Later frames still use memory. The views receive segmentation loss, KL divergence toward the detached MedSAM2 mask, an entropy penalty, and cross-view consistency. The MedSAM2 student itself also keeps segmentation loss on the end-diastole and end-systole frames.

The following table gives the training configuration of each distillation method. Same means the setting is identical to FreeEcho.

SettingFreeEchoDistillSAMKnowSAM
TeacherFrozen prompted checkpointSameSame
InitializationCopied from the teacherSameSame
EchoONE optimizationAdam, learning rate 1×10⁻⁴, batch size 8, input 256×256SameSame
MedSAM2 optimizationAdamW, learning rate 5×10⁻⁵, batch size 1, clip length 10, input 256×256SameSame
Teacher promptEchoONE receives a centroid and the PC-Mask. MedSAM2 receives an end-diastole centroid.SameSame
EpochsEchoONE 20, MedSAM2 20SameSame

Q3. The paper emphasizes superiority over existing prompt-free models, but those models differ substantially in model capacity and pretraining. Could the authors either provide a more capacity/pretraining-matched prompt-free control or narrow this claim to emphasize the stronger same-architecture comparisons against direct prompt removal and alternative distillation strategies?

We agree, and we have narrowed the claim. We revised five sentences so that any statement of superiority over existing prompt-free models is explicitly about students distilled from the state-of-the-art prompted teachers used in this paper:

  1. Abstract. “They also outperform existing prompt-free models” is now “Students distilled from state-of-the-art prompted teachers can outperform existing prompt-free models.”
  2. Introduction. “The resulting models also outperform existing prompt-free models” is now “The resulting models from state-of-the-art prompted teachers can also outperform existing prompt-free models.”
  3. Contributions. The same qualifier is added to the claim of superior accuracy and cross-source generalization relative to existing prompt-free methods.
  4. Zero-shot discussion of Table 1. The two prompt-free models that exceed the listed prompt-free methods are described as “distilled from state-of-the-art prompted teachers via FreeEcho.”
  5. In-domain discussion of Table 2. The same qualifier is added.
Reviewer_SKSZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

FreeEcho presents a knowledge distillation framework for converting prompt-based echocardiography video segmentation models into prompt-free models. The method combines ground-truth supervision with teacher-generated distillation signals and introduces a prompt removal curriculum to facilitate training without external localization inputs. Experiments on multiple models and public datasets suggest that the distilled student models can approach or match their teachers' performance while outperforming direct prompt removal and existing prompt-free methods. The paper also reports improved generalization across different data sources and conducts ablation studies to investigate the contributions of individual design choices.

Critical strengths
  1. The paper addresses a practical problem in echocardiography video segmentation, as eliminating external prompts can simplify inference and improve the scalability of automated segmentation systems.
  2. The proposed distillation framework provides a potentially useful way to transfer the capabilities of existing prompt-based models to prompt-free settings, and the evaluation across multiple models and datasets is encouraging.
Critical weaknesses
  1. The novelty appears moderate. Knowledge distillation and curriculum learning are well-established techniques, and the paper needs to clarify the fundamental methodological innovations beyond combining these existing concepts for prompt-free echocardiography segmentation.
  2. The methodological motivation and technical contributions require a more rigorous explanation. In particular, it is unclear from the abstract how the adapted prediction pipeline and prompt removal curriculum differ substantially from existing distillation or prompt-free learning strategies.
  3. The paper's presentation has considerable room for improvement. The framework overview should be redesigned to communicate the overall pipeline and the interactions between its components more clearly, while the result figures and tables should provide more effective and informative comparisons.
  4. The writing also needs substantial polishing. The authors should improve the precision of technical descriptions, explain the rationale behind key design choices more explicitly, and ensure that the experimental results and ablation findings are discussed in sufficient detail.
Questions for authors
  1. Novelty: What are the key methodological innovations of FreeEcho beyond the application of knowledge distillation and curriculum learning to echocardiography video segmentation? How does the proposed approach differ from closely related prompt-free distillation methods?
  2. Technical design: What specific limitations of direct prompt removal are addressed by the adapted prediction pipeline and prompt removal curriculum? Could the authors provide a clearer explanation of their mechanisms and respective contributions?
  3. Generalization: How is cross-dataset generalization evaluated? Are the improvements consistent across different acquisition settings, patient populations, and echocardiography data sources?
  4. Ablation studies: How much does each component—ground-truth supervision, teacher-derived distillation signals, and the prompt removal curriculum—contribute to the final performance? Have the authors examined whether these components interact or whether simpler alternatives achieve comparable results?
  5. Presentation and reproducibility: Could the authors improve the framework diagram, experimental figures and tables, and technical descriptions to make the method easier to understand and reproduce?
6 comments

Q1. Novelty: What are the key methodological innovations of FreeEcho beyond the application of knowledge distillation and curriculum learning to echocardiography video segmentation? How does the proposed approach differ from closely related prompt-free distillation methods?

We thank the reviewer for the question. FreeEcho introduces an architecture-aware prompt-free conversion framework that adapts the prediction pipeline, transfers complementary teacher-derived supervision, and progressively removes the student’s external prompts. The teacher retains its prompt throughout training, while the student learns to operate without one. A shared soft-mask loss supervises the final contour, but the prompt also affects intermediate localization and temporal propagation. FreeEcho therefore places auxiliary supervision at the representations through which each architecture uses prompt information.

In a per-frame model, the prompt tells the decoder where the left ventricle is. EchoONE already forms a coarse map of that location and extent. L_spatial distills this coarse map, so the student can produce the spatial anchor from the image when the prompt is absent. Table 3 shows the effect. The full EchoONE objective scores 0.937 on CAMUS and 0.898 / 0.867 / 0.614 on EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA. Removing L_spatial lowers these to 0.926 and 0.869 / 0.851 / 0.552. On CAMUS, EchoNet-Dynamic, and CardiacUDA, this drop is larger than the drop from removing ground-truth supervision or soft-mask supervision. The largest gap is on CardiacUDA, 0.614 to 0.552. On Pediatric, removing the soft mask is larger (0.867 to 0.832). The coarse-map signal contributes on every set, with its largest measured contribution on CardiacUDA.

In a memory model, the prompt is written into an object memory on the first frame, and later frames read that memory. Matching pixels does not force the student to write the same object identity without the prompt. L_temporal aligns the teacher and student object memories from the second frame onward. The first frame is excluded because the two models are prompted differently there. Table 3 shows that this term is decisive for successful prompt-free segmentation in the memory-based student. The full MedSAM2 objective scores 0.929 and 0.878 / 0.863 / 0.535. Removing ground truth or the soft mask leaves 0.920 / 0.870 / 0.849 / 0.496 and 0.924 / 0.875 / 0.857 / 0.477. Removing L_temporal leaves 0.000 on CAMUS, EchoNet-Dynamic, and EchoNet-Pediatric, and 0.039 on CardiacUDA.

The two signals are therefore specific to how each model uses a prompt. The coarse map teaches the per-frame student to recover the spatial anchor from the image. The object-level memory alignment transfers the representation used to propagate target information across frames. Table 3 separates these roles: spatial distillation improves localization across sources, while temporal distillation enables effective prompt-free memory-based segmentation.

Q2. Technical design: What specific limitations of direct prompt removal are addressed by the adapted prediction pipeline and prompt removal curriculum? Could the authors provide a clearer explanation of their mechanisms and respective contributions?

We thank the reviewer for asking us to separate these two mechanisms. Direct prompt removal has two distinct failures. The prediction pipeline is written for an external prompt, and the weights have never had to localize without one. The adapted pipeline addresses the first failure. The prompt-removal curriculum addresses the second. For a per-frame model, the prompt is the spatial anchor of that frame. Removing it at test time, with no change in weights, lowers EchoONE from 0.939 / 0.898 / 0.877 / 0.673 to 0.894 / 0.853 / 0.808 / 0.284 on CAMUS, EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA (Table 1). For a memory model, the prompt is used on the first frame to create the object that later frames track. Removing the initial prompt disrupts the target-conditioning process on which subsequent propagation depends. MedSAM2 then scores 0.000 / 0.000 / 0.000 / 0.009. The same pattern appears in-domain (Table 2). EchoONE falls to 0.684 on EchoNet-Dynamic and 0.000 on CardiacUDA, and MedSAM2 stays near zero on all four datasets.

The adapted pipeline keeps the original segmentation network and changes only the empty-prompt interface. A per-frame student still calls the prompt branch, with an empty prompt in place of a point prompt, and predicts each frame from the image alone. A memory student cannot use a fully empty input, because that input prevents memory from being initialized. FreeEcho uses a model-defined prompt-free representation. The first frame is predicted without memory, and that prediction is written into memory. Each later frame reads the previous memory and updates it with the same prompt-free representation. Later frames therefore track an object written by the student.

This interface makes prompt-free training possible. It does not by itself teach the weights to localize. In Table 4, the row trained with no prompt at any epoch already uses the adapted interface and the same distillation losses. It reaches 0.924 for EchoONE and 0.917 for MedSAM2 on CAMUS, compared with 0.894 and 0.000 for direct removal in Table 1. The adapted interface lets the student initialize memory from its own first-frame prediction, and distillation trains this prompt-free prediction and propagation process. The curriculum addresses the remaining training gap. The teacher retains its prompt, so its soft mask is already localized. A student that is empty from the first epoch must imitate that mask before it can localize. Linear annealing reduces the probability that the student receives a prompt from 1 to 0. Early epochs learn with localization guidance. Later epochs remove the prompt, so localization has to move into the weights. The teacher keeps its prompt throughout, and the supervision does not weaken with the student’s input.

Table 4 holds the interface, the losses, and the empty-prompt test fixed, and changes only how the prompt is withdrawn. Dice is listed as CAMUS, EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA.

Table4. Prompt-removal curriculum. Linear annealing is the default. Each cell is CAMUS / EchoNet / Pediatric / CardiacUDA.

ModelScheduleCAMUSEchoNetPediatricCardiacUDA
EchoONELinear annealing0.937±0.0360.898±0.0590.867±0.1120.614±0.284
EchoONEAdaptive gating0.928±0.0410.887±0.0690.863±0.1050.575±0.255
EchoONEPrompt-free0.924±0.0430.880±0.0740.862±0.1020.564±0.256
MedSAM2Linear annealing0.929±0.0390.878±0.0730.863±0.1020.535±0.257
MedSAM2Adaptive gating0.914±0.0650.872±0.0980.851±0.1160.515±0.339
MedSAM2Prompt-free0.917±0.0390.867±0.0950.850±0.1120.499±0.327

Training with no prompt is already far above direct removal, and linear annealing is higher on every column. The largest added gain is on CardiacUDA, from 0.564 to 0.614 for EchoONE and from 0.499 to 0.535 for MedSAM2. Adaptive gating does not force the prompt to disappear gradually, and it lands between the two schedules or below both. The pipeline makes an empty prompt a valid input and lets a memory model create its object from its own first-frame prediction. The curriculum then moves localization from the prompt into the weights. Table 1 shows the cost of deleting the prompt with no retraining. Table 4 shows that, once the interface is fixed, gradual withdrawal still outperforms removing the prompt from the first epoch.

Q3. Generalization: How is cross-dataset generalization evaluated? Are the improvements consistent across different acquisition settings, patient populations, and echocardiography data sources?

We thank the reviewer for asking how the cross-dataset results are obtained and whether the gain is uniform. Cross-dataset performance is evaluated through zero-shot transfer. Every model uses CAMUS as its task-specific training or adaptation source. The weights are then frozen and evaluated on EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA, with no target-domain training and no checkpoint selection on the target. The CAMUS column reports source performance. Dice is reported as the mean and population standard deviation across annotated frames.

The four datasets vary in acquisition settings, patient populations, or both. CAMUS (Leclerc et al., 2019) is the task-specific training source. It contains 1,000 videos from 500 patients, with apical two-chamber and four-chamber views and left-ventricular annotations at end-diastole and end-systole. The reported source test set contains 200 annotated frames. Training on this source and fixing the resulting checkpoint provides a common starting point for evaluating transfer to the other three sources. EchoNet-Dynamic (Ouyang et al., 2020) keeps the adult population and the four-chamber view, and changes the hospital and the acquisition pipeline. It contains 10,030 routine adult apical-four-chamber examinations from Stanford Medicine, with reference tracings at end-diastole and end-systole and labels for ejection fraction. The zero-shot set uses the 2,552 end-diastole and end-systole frames from the official test videos. CAMUS training also includes the two-chamber view, so this transfer is four-chamber only. EchoNet-Pediatric (Reddy et al., 2023) changes the patient population from adults to children. The release contains 7,643 pediatric echocardiograms spanning ages 0 to 18. It provides a population-transfer evaluation from the adult training source to pediatric echocardiography. This evaluation uses the 706 apical-four-chamber end-diastole and end-systole test frames. CardiacUDA (Yang et al., 2023) changes the institution and the scanner. It contains 992 videos from two hospitals acquired with different ultrasound scanners, and it was built as a domain-shift benchmark. The zero-shot evaluation uses the full four-chamber split of 566 frames. The separate in-domain experiment uses a 15% site holdout of 86 frames. Those two frame sets are not paired. This target also moves from CAMUS training on two views to a four-chamber test. The direction of the improvement is the same on all three targets. Relative to direct prompt removal, EchoONE rises from 0.853, 0.808, and 0.284 to 0.898, 0.867, and 0.614 on EchoNet-Dynamic, EchoNet-Pediatric, and CardiacUDA. MedSAM2 rises from 0.000, 0.000, and 0.009 to 0.878, 0.863, and 0.535. FreeEcho is also higher than DistillSAM and KnowSAM on every target for both architectures (Table 1).

The per-dataset results show how much prompt-free capability is recovered across these sources. On a new adult four-chamber source, EchoONE reaches 0.898, the same value as the prompt-based teacher, against 0.853 with the prompt removed. MedSAM2 recovers from 0 to 0.878, against a prompt-based teacher at 0.900. On pediatric hearts, both students retain strong accuracy without external localization inputs. EchoONE reaches 0.867 against 0.877, and MedSAM2 reaches 0.863 against 0.884. On CardiacUDA, both students achieve their largest absolute gains over direct prompt removal. EchoONE rises from 0.284 to 0.614, against a prompt-based teacher at 0.673. MedSAM2 rises from 0.009 to 0.535, against 0.739.

The improvement is therefore present for a new adult source, a pediatric population, and a two-hospital acquisition setting. Across all three targets, both FreeEcho students outperform direct prompt removal and the adapted distillation baselines. This consistent advantage demonstrates the value of transferring prompt-dependent segmentation capabilities into models that operate without external localization inputs across different data sources.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear AC and Reviewers,

We sincerely thank all reviewers for their time and constructive feedback. We appreciate their recognition of the following strengths of FreeEcho:

  • Practical importance. Removing external localization prompts addresses an important deployment barrier to automated and scalable echocardiography video analysis. [Reviewers SKSZ, PSPh, and K5xS]
  • Applicability across architectures. FreeEcho supports both per-frame and memory-based segmentation models, with architecture-specific supervision that transfers spatial localization and object-memory representations. [Reviewers PSPh and K5xS]
  • Comprehensive evaluation. The experiments cover multiple architectures and datasets, comparisons with direct prompt removal, alternative distillation methods, and existing prompt-free models, together with ablations of the supervision signals and prompt-removal curriculum. [Reviewers PSPh and K5xS]
  • Strong cross-source evaluation protocol. Training or adapting models on CAMUS and evaluating their frozen checkpoints on other sources without target-domain training or checkpoint selection provides a meaningful assessment of prompt-free transfer. [Reviewer PSPh] The reviewers’ suggestions have helped us strengthen the experimental evidence and clarify the framework. Our responses and manuscript revisions address the following points:
  • Clarified the methodological contribution and component roles. We explain how FreeEcho connects prompt-free prediction interfaces, architecture-specific supervision, and progressive prompt withdrawal. Spatial supervision transfers intermediate localization cues in per-frame models, while object-memory supervision supports target propagation in memory-based models. We also distinguish the role of the adapted prediction interface from that of the training curriculum. [Reviewers SKSZ and K5xS]
  • Expanded the loss ablations. The revised Table 3 adds single-term supervision experiments alongside the existing leave-one-out comparisons. These results distinguish the effects of ground-truth supervision, mask distillation, and architecture-specific supervision. Across both architectures and all four datasets, the full objective achieves higher Dice than the tested single-term alternatives. The curriculum comparison further shows the benefit of gradual prompt withdrawal under a fixed interface and objective. [Reviewers SKSZ and K5xS]
  • Added repeated-seed comparisons. We repeated FreeEcho, DistillSAM, and KnowSAM training with three random seeds and report variation across runs separately from variation across annotated frames. FreeEcho retains higher mean Dice than both distillation baselines on all four datasets for both architectures. [Reviewer PSPh]
  • Clarified comparison protocols and implementation details. Our responses describe how DistillSAM and KnowSAM were adapted, including their supervision signals, initialization, optimization settings, and prompt handling. The revised manuscript makes the evaluation protocols explicit and identifies the compared prompt-free students as models distilled from strong prompt-based teachers. [Reviewers SKSZ and PSPh]
  • Extended the analysis to additional teachers. We report prompt-free conversion results for Medical-SAM2 and memSAM alongside the original teachers. These comparisons examine how student performance relates to the capabilities of different teachers across the four datasets. [Reviewer K5xS]
  • Added temporal evaluation. We evaluate 100 densely annotated ED-to-ES sequences from 50 CAMUS patients using ventricular area-curve agreement, relative area error, and segmentation accuracy across frames. FreeEcho achieves lower relative area error and higher frame Dice than DistillSAM and KnowSAM with both architectures, supporting its effectiveness beyond the annotated endpoint frames. [Reviewer K5xS]
  • Improved presentation and reproducibility. The revised Figure 1 caption clarifies the teacher–student information flow and architecture-specific supervision. The captions of Tables 1–4 specify the evaluation protocols and controlled settings, and Appendix A.3 provides the optimization configuration in Table 7. [Reviewer SKSZ] The revised manuscript incorporates the expanded loss ablations and the methodological, protocol, and presentation clarifications described above. Additional multi-seed, teacher, and temporal analyses are provided in our point-by-point responses. Together, these results strengthen the evidence that FreeEcho transfers the capabilities of prompt-based echocardiography video segmentation models into effective prompt-free counterparts across architectures and data sources.

Thank you again for your careful consideration.

Best regards, Authors of paper #43920

Area_Chair_NR8FOpenReview ↗

All reviewers, please be responsive

Dear reviewers,

The authors have provided detailed rebuttal. Please respond to authors' rebuttal, If you have not done so. Also, please provide concise and constructive response. Thanks!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.