- 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
-
The paper addresses a practical problem. Removing the need for manual prompts is important for automatic echocardiography analysis.
-
The proposed framework can be applied to both per-frame and memory-based segmentation models, showing good flexibility.
-
The experiments are relatively comprehensive, covering different models, multiple datasets, and cross-dataset evaluation.
-
The ablation studies demonstrate the contributions of different distillation losses and the prompt-removal strategy.
Critical weaknesses
-
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.
-
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.
-
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.
-
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
-
What are the main differences between FreeEcho and existing knowledge distillation methods?
-
How does the performance change when using a weaker teacher model?
-
Can the authors provide more analysis of temporal consistency in video segmentation?
3 comments
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.
| Teacher | Prompt | CAMUS | EchoNet | Pediatric | CardiacUDA |
|---|---|---|---|---|---|
| EchoONE | per-frame click | 0.939 → 0.937 | 0.898 → 0.898 | 0.877 → 0.867 | 0.673 → 0.614 |
| Medical-SAM2 | per-frame click | 0.920 → 0.922 | 0.864 → 0.866 | 0.838 → 0.840 | 0.533 → 0.524 |
| MedSAM2 | ED click | 0.937 → 0.929 | 0.900 → 0.878 | 0.884 → 0.863 | 0.739 → 0.535 |
| memSAM | ED click | 0.923 → 0.927 | 0.802 → 0.768 | 0.733 → 0.644 | 0.397 → 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.
| Architecture | Method | Pearson r | Sequences with r > 0.9 | Relative area error | Frame Dice | ES area < ED area |
|---|---|---|---|---|---|---|
| EchoONE | FreeEcho | 0.923 | 93% | 7.4% | 0.935 | 99% |
| EchoONE | DistillSAM | 0.917 | 91% | 8.3% | 0.928 | 99% |
| EchoONE | KnowSAM | 0.919 | 90% | 8.4% | 0.925 | 98% |
| MedSAM2 | FreeEcho | 0.932 | 94% | 8.8% | 0.922 | 98% |
| MedSAM2 | DistillSAM | 0.932 | 93% | 9.3% | 0.910 | 98% |
| MedSAM2 | KnowSAM | 0.929 | 90% | 10.2% | 0.904 | 98% |
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).
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.