acceptodds
Under review as a conference paper at ICLR 2027

Pretrained Systems for Low-Data Myocardial Segmentation: Direct Versus LV-Supervised Fine-Tuning

Abstract

Choosing a pretrained system is important when myocardial annotations are limited. We compare Pixio-L with Mask2Former and EchoFM with UNETR for myocardial (MYO) segmentation on CAMUS apical four-chamber (A4C) echocardiography, evaluated against official provider-supplied reference annotations. Pixio-L is a released student distilled from the web-scale Pixio-5B teacher; EchoFM is pretrained on echocardiographic videos. Each system starts from its released encoder weights and is trained either directly on MYO or first on EchoNet-Dynamic A4C left ventricular (LV) cavity segmentation and then on MYO. Under full fine-tuning, across target-training patient fractions of 5%, 10%, 25%, 50%, and 100% and three training seeds, Pixio-L achieves higher mean MYO Dice under both routes. At 5%, the cross-system differences favoring Pixio-L are 13.38 percentage points for Direct fine-tuning and 6.52 points after LV supervision. The LV stage improves mean Dice in both systems at every fraction, with the largest gains at 5%: 2.29 points for Pixio-L and 9.15 for EchoFM. An additional 60-run frozen-MYO-encoder ablation shows that the effect of encoder updates depends on the system and initialization route. These exploratory results on one fixed patient split identify Pixio-L with Mask2Former as a strong low-data baseline and support intermediate LV supervision for both systems. Differences in architecture, input, and training recipe prevent attributing the ranking to pretraining domain or distillation alone.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
myocardial segmentation, echocardiography, low-data learning, transfer learning, fine-tuning, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4XTFOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies low-data myocardial segmentation in echocardiography by comparing two existing pretrained segmentation systems, Pixio-L with Mask2Former and EchoFM with UNETR, on CAMUS A4C myocardium segmentation. The authors evaluate two fine-tuning routes: direct fine-tuning on CAMUS MYO labels and an intermediate LV-supervised route that first fine-tunes on EchoNet-Dynamic LV cavity segmentation before adapting to MYO. Across five target-data fractions and three training seeds, Pixio-L achieves higher mean MYO Dice than EchoFM under both routes, while LV supervision improves both systems, especially in the lowest-data regime and particularly for EchoFM. The paper also includes a frozen-encoder ablation showing that the effect of encoder updates depends on the system and initialization route

Critical strengths
  1. Studying whether LV supervision helps MYO segmentation is well motivated, especially in echocardiography where annotation types often differ across datasets.
  2. The paper separates absolute final MYO performance from gain due to LV supervision.This avoids the common mistake of claiming that a method is superior simply because it has a larger transfer gain.
Critical weaknesses
  1. The technical contribution of the paper is limited. The paper does not introduce a new model, loss, adaptation method, pretraining objective, or theoretical insight. Its main contribution is an empirical comparison of two existing pretrained systems under direct and LV-supervised fine-tuning.
  2. The benchmark scope is narrow. Only two systems are compared, but they differ in architecture, decoder, input format, temporal context, resolution, and optimization setup. As a result, the study cannot isolate the effects of pretraining, architecture, or domain specificity. Key baselines are missing, including nnU-Net, training from scratch, ImageNet-pretrained models, architecture-matched controls, and other ultrasound or foundation models.
  3. The evaluation uses only one CAMUS A4C split and one nested subset draw, which is especially concerning in the low-data setting where patient selection can strongly affect results. The study lacks external validation and does not evaluate other views or datasets.
Questions for authors
  1. Why are only two pretrained systems evaluated? How should readers generalize from this study beyond the specific Pixio-L + Mask2Former and EchoFM + UNETR pipelines?
  2. Why is there no strong non-foundation-model baseline such as nnU-Net?
  3. Could the advantage of Pixio-L be due to Mask2Former rather than Pixio-L pretraining? Have the authors tested EchoFM with Mask2Former or Pixio-L with UNETR-style decoding?
  4. Can the authors verify whether CAMUS or EchoNet-Dynamic patients/images were included in the released pretrained checkpoints?
  5. Have the authors considered parameter-efficient fine-tuning methods such as LoRA, adapters, or partial-layer fine-tuning? Since the target setting is explicitly low-data myocardial segmentation, these methods may be more appropriate than full fine-tuning.
7 comments
AuthorsOpenReview ↗

Response to reviewer 4XTF (part 1/5): main result, what we added, W1

Main result. MYO labels in echocardiography are scarce, LV cavity labels are abundant, and the LV cavity border is the inner border of the myocardium. Training on LV before MYO (LV→MYO) raised MYO Dice on CAMUS when MYO labels were scarce, in every system we tested: with 5% of the training patients, by +2.29 pp [+1.40, +3.32] for Pixio-L, +9.15 [+7.80, +10.47] for EchoFM, +1.39 [+0.76, +2.07] for nnU-Net and +6.12 [+4.84, +7.41] for EchoNet, with intervals above zero at 10% as well and smaller gains as MYO labels grew (Section 4.3, Table 2).

What we added in response to your review.

  1. nnU-Net v2 2D, trained from scratch for 100 epochs, both routes, five fractions (Appendix A.2; W2, Q2).
  2. EchoNet, an ImageNet-initialized DeepLabV3-ResNet50, both routes, five fractions (Appendix A.2; W2).
  3. One additional CAMUS patient split, all four systems at 5% and 10% (Appendix D; W3).
  4. One external site, HMC-QU, with all models applied without retraining (Section 4.3, Appendix F; W3).
  5. A rationale, zero-shot LV localization on CAMUS and a boundary-wise analysis of the gain (Sections 3.2 and 4.3, Appendix E; W1).
  6. LoRA on the encoder at 5% for Pixio-L and EchoFM (Appendix H.1; Q5); results in part 5/5.

Not evaluated in this manuscript (Section 5): architecture-matched controls, Pixio-L or EchoFM trained from scratch, other foundation models, adapters other than LoRA, selected-layer tuning, head reinitialization after LV training, freezing during the LV stage, and other views.

W1: "The technical contribution of the paper is limited. The paper does not introduce a new model, loss, adaptation method, pretraining objective, or theoretical insight. …"

We agree that the paper introduces no new model, loss, adaptation method or pretraining objective, and we prove no theorem. Your reading of the main contribution as an empirical comparison of two systems matches how our submitted abstract read, and we have rewritten the abstract. The contribution is controlled, patient-paired evidence about the intermediate stage: within each system, the same target patients are trained with and without the LV stage. On insight, what we add is empirical: the gain is largest when MYO labels are scarce and is not confined to the border the two labels share. The rationale (item 5) notes that the myocardium is the band between the LV cavity border and the epicardial border, so LV labels supervise one of the two target borders and where the ventricle lies. Two results added after review bear on it (Section 4.3, Appendix E). Although the LV stage used no CAMUS image, all four LV-stage models already located the CAMUS LV cavity (LV Dice 0.91 for Pixio-L, 0.90 for EchoNet and nnU-Net, 0.85 for EchoFM; Table 19). In a boundary-wise analysis, the LV stage reduced error near both MYO borders in Pixio-L and EchoFM at 5% and 10%, more near the outer border for Pixio-L, so the gain is not confined to the shared inner border (Table 18). Together with the localization result, this suggests that the LV stage supplies the ventricle's position and overall shape rather than one border.

AuthorsOpenReview ↗

Response to reviewer 4XTF (part 2/5): W2

W2: "The benchmark scope is narrow. Only two systems are compared … Key baselines are missing, including nnU-Net, training from scratch, ImageNet-pretrained models, architecture-matched controls, and other ultrasound or foundation models."

We added two of the missing baselines, nnU-Net and an ImageNet-initialized EchoNet (Appendix A.2, below). We agree that the original two systems cannot isolate pretraining, architecture or domain, and the revision attributes no difference between systems to any of them. The LV gain does not need that isolation, because it is measured with the system held fixed.

  • nnU-Net v2 2D, trained from scratch for 100 epochs. On the original CAMUS split, with both trained directly, it was not clearly different from Pixio-L at 5% (+0.73 [−0.30, +1.72]) and scored higher at 10–100% (+1.40 to +1.55) (Section 4.4, Table 9). Its LV gain was +1.39 [+0.76, +2.07] at 5% and +0.60 [+0.13, +1.12] at 10%, and not resolved at 25–100% (Table 2).
  • EchoNet, an ImageNet-initialized DeepLabV3-ResNet50 trained at 112×112 with an EchoNet-derived recipe. With both trained directly, it scored lower than Pixio-L at every fraction on this split (−8.87 [−10.85, −6.94] at 5% to −1.60 [−2.52, −0.79] at 100%; Table 9), and its mean Dice was above EchoFM's (Table 7). The LV stage raised its mean MYO Dice at 5% (+6.12 [+4.84, +7.41]), and the intervals were above zero at the other four fractions too.

The items still missing are listed in part 1/5.

AuthorsOpenReview ↗

Response to reviewer 4XTF (part 3/5): W3

W3: "The evaluation uses only one CAMUS A4C split and one nested subset draw … The study lacks external validation and does not evaluate other views or datasets."

One additional split (Section 4.1, Appendix D). We ran one additional CAMUS patient split (360/90/50 patients; the 50 test patients are disjoint by patient ID from the original test set; new nested 5% and 10% subsets) for Pixio-L and EchoFM, both routes, three seeds; its eight contrasts were fixed before any of its test scores existed. All eight kept the direction reported on the original split, with 95% intervals excluding zero (Table 15): Pixio-L minus EchoFM, Direct, 5%: +13.00 pp [+10.96, +15.19]; Pixio-L minus EchoFM, LV, 5%: +5.46 pp [+4.20, +6.78]; LV minus Direct, Pixio-L, 5%: +1.96 pp [+1.35, +2.59]; LV minus Direct, EchoFM, 5%: +9.50 pp [+7.74, +11.43]; Pixio-L minus EchoFM, Direct, 10%: +10.29 pp [+8.37, +12.27]; Pixio-L minus EchoFM, LV, 10%: +5.34 pp [+3.93, +6.79]; LV minus Direct, Pixio-L, 10%: +1.80 pp [+1.17, +2.48]; LV minus Direct, EchoFM, 10%: +6.76 pp [+5.27, +8.26].

nnU-Net and EchoNet, outside these 8 contrasts, also had LV gains above zero at 5% and 10% on this split (Table 16).

HMC-QU (Section 4.3, Appendix F). We added one external site, HMC-QU (109 patient IDs, 2,349 A4C frames, clinician-verified pseudo-labels covering the LV wall); all models were applied without retraining. Two results, kept separate. First, the LV gain held: under full fine-tuning, the LV gain of the original-split models was above zero in every system at every fraction; at 5% it was +3.66 [+3.05, +4.23] for Pixio-L and +12.58 [+11.58, +13.58] for EchoFM (Table 8). With the encoder frozen, it was also above zero in Pixio-L and EchoFM at every fraction (Table 31; at 5%, EchoFM +12.60 [+11.58, +13.65]). Second, four frozen-encoder comparisons between systems or adaptation modes, none of them an LV gain, reversed at 5% relative to CAMUS (Table 20): Pixio-L minus EchoFM, −9.88 [−11.14, −8.59] under Direct (also reversed at 50% and 100%) and −1.05 [−1.96, −0.09] under the LV-supervised route; frozen minus full fine-tuning after the LV stage, Pixio-L −1.75 [−2.45, −1.09] (CAMUS +0.76) and EchoFM +1.89 [+1.66, +2.13] (CAMUS −0.39). Our conclusions about system ordering and freezing therefore stay limited to CAMUS. For models trained on the additional split (Table 21), the LV gain was above zero in seven of eight cells; the exception was nnU-Net at 10% (−3.33 [−4.96, −1.77]). Absolute HMC-QU scores are not compared with CAMUS.

Other views. Only the four-chamber view is evaluated; other views are listed as not evaluated (Section 5).

Reviewer_1s4AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper compares the myocardial (MYO) segmentation performance of Pixio-L, Maask2Former, EchoFM, and UNETR on the apical four-chamber (A4C) echocardiographic views from the CAMUS dataset, using the officially provided reference annotations as the evaluation benchmark. The experimental results demonstrate that differences in architecture, input, and training schemes make it impossible to attribute the observed rankings solely to the pre-training domain or knowledge distillation.

Critical strengths

The author evaluated various pre-trained networks using metrics such as Dice (including boundary-based assessment), sensitivity, and precision. The evaluation approach is comprehensive, and the experiments are thorough and credible.

Critical weaknesses

The originality of this paper is difficult to define; it does not propose a new network but rather provides a comparison of existing ones, making it perhaps more suitable for a review-type journal.

Questions for authors
  1. What is the core innovation of this paper?
  2. Based on the results of various comparative experiments, what is the most significant conclusion?
  3. Given the respective pros and cons of existing pre-trained networks, is there a better solution? I believe this is the most important point.
1 comment
AuthorsOpenReview ↗

Response to reviewer 1s4A

Main result. MYO labels in echocardiography are scarce, LV cavity labels are abundant, and the LV cavity border is the inner border of the myocardium. Training on LV before MYO (LV→MYO) raised MYO Dice on CAMUS when MYO labels were scarce, in every system we tested: with 5% of the training patients, by +2.29 pp [+1.40, +3.32] for Pixio-L, +9.15 [+7.80, +10.47] for EchoFM, +1.39 [+0.76, +2.07] for nnU-Net and +6.12 [+4.84, +7.41] for EchoNet, with intervals above zero at 10% as well (Section 4.3, Table 2).

Summary: "… differences in architecture, input, and training schemes make it impossible to attribute the observed rankings solely to the pre-training domain or knowledge distillation."

That sentence of our abstract is a caveat, not our result; the submitted abstract gave the result above only one sentence. The revised abstract is built around it and moves the caveat to Section 5.

"The originality of this paper is difficult to define; it does not propose a new network but rather provides a comparison of existing ones, making it perhaps more suitable for a review-type journal."

Our object of study is a training route, not a set of networks: does an intermediate stage supervised by abundant LV labels improve MYO segmentation when MYO labels are scarce, and by how much? It is not a review: every number comes from our own training runs under one paired design (the same target-training patients for both routes, the same test patients, five fractions, three seeds). There is no new network.

Q1. "What is the core innovation of this paper?"

The core contribution is a measurement of a training stage, not a new network: how much training first on abundant LV labels adds to MYO segmentation when MYO labels are scarce, measured on the same patients within each of four systems at five label fractions. It does two things a comparison of networks does not: it reports the stage's gain separately from the final ranking (a larger gain did not mean a better system), and it tests one explanation: the gain was not confined to the inner border that the LV label shares with the myocardium (Section 4.3, Table 18).

Q2. "Based on the results of various comparative experiments, what is the most significant conclusion?"

On CAMUS, the LV stage raised MYO Dice in both original systems at every fraction, most when MYO labels were scarcest; the two systems added after review also gained at 5% and 10%. On one additional patient split, the gains at 5% and 10% were also above zero in all four systems (Table 16). On HMC-QU (one external site, clinician-verified pseudo-labels), the LV gain of the original-split models was above zero in every system and fraction (Tables 8 and 31); what reversed there, at 5%, were four frozen-encoder comparisons between systems or between adaptation modes, none of them an LV gain (Table 20), and because of them we limit our conclusions about system ordering and freezing to CAMUS.

Q3. "Given the respective pros and cons of existing pre-trained networks, is there a better solution? I believe this is the most important point."

At 5% of the MYO training patients on CAMUS, our best tested solution is LV→MYO with Pixio-L, keeping its encoder frozen or LoRA-adapted during MYO training: 82.49% and 82.85% Dice, against 81.73% with full fine-tuning, while updating about 21 M instead of 324 M parameters. It combines existing parts and is not a new method; the difference between freezing and LoRA was not resolved. The LV stage needs no extra MYO labels (the LV labels come from a public dataset, and one LV run per seed serves all MYO fractions), and on CAMUS it raised MYO Dice at 5% and 10% in every system we tested (Tables 2 and 16). Without the LV stage, freezing or LoRA cost Pixio-L 8–9 points, so the cheaper adaptation was usable only after the stage (Section 4.6; Appendix H.1, Table 29). On which network to start from: choose by final Dice, not by the size of the LV gain. Trained directly, a from-scratch nnU-Net (100 epochs) showed no clear difference from Pixio-L at 5% and scored higher at 10–100% (Section 4.4, Table 9).

If the revision and these answers address your concerns, we would be grateful if you would consider updating your score.

Reviewer_f6QcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an empirical study on training pipelines for echocardiography segmentation tasks where annotation data is limited. It explores gradual pipeline stages and transfer learning steps when a related segmentation problem exists (here, LV segmentation is used as a task with richer data; indeed, the border of the LV aligns with the MYO, so the problems are related). Then, it explores different orchestrations of encoders (e.g., Pixio, EchoFM) and decoders (e.g., Mask2Former, UNETR) with the goal of finding the optimal path for training on the MYO segmentation task.

Critical strengths

-- The study can help practitioners in this field, as it is a relevant and popular task in echocardiography.

-- The introduction of Pixio + Mask2Former seems to be a strong encoder-decoder pair for echocardiography segmentation, which is of practical value.

Critical weaknesses

The study does not address core ML problems and their solutions; rather, it addresses practitioners in this specific field.

Questions for authors

--

1 comment
AuthorsOpenReview ↗

Response to reviewer f6Qc

Main result. MYO labels in echocardiography are scarce, LV cavity labels are abundant, and the LV cavity border is the inner border of the myocardium. Training on LV before MYO (LV→MYO) raised MYO Dice on CAMUS when MYO labels were scarce, in every system we tested: with 5% of the training patients, by +2.29 pp [+1.40, +3.32] for Pixio-L, +9.15 [+7.80, +10.47] for EchoFM, +1.39 [+0.76, +2.07] for nnU-Net and +6.12 [+4.84, +7.41] for EchoNet, with intervals above zero at 10% as well (Section 4.3, Table 2).

"The study does not address core ML problems and their solutions; rather, it addresses practitioners in this specific field."

The core ML problem is intermediate (related-task) supervision under label scarcity: when, how much and why does a related, abundant label help a target task whose labels are scarce? LV cavity and myocardium are a testable instance: LV labels are plentiful, and the cavity border is one of the two myocardial borders. Three results concern the intermediate stage rather than echocardiography: (i) its benefit was largest where target labels were fewest; (ii) it was not confined to the structure the two labels share (point 3); (iii) it changed how much encoder adaptation mattered: at 5% on CAMUS, freezing the Pixio-L encoder or adapting it with LoRA cost 9.30 and 8.03 points without the stage and none after it (Sections 4.3 and 4.6). We propose no new method: the contribution is controlled evidence on when and how much the stage helps, and one test of why, on one label pair.

  1. How much, and in which systems. Within each system, both routes use the same target-training and test patients at five fractions and three seeds. On the original split the gain was above zero for Pixio-L and EchoFM at every fraction; two systems added after review, nnU-Net trained from scratch (100 epochs) and an ImageNet-initialized EchoNet, gained at 5% and 10% (EchoNet at 25–100% too), as did all four on one additional patient split (Appendix D). On HMC-QU (one external site, clinician-verified pseudo-labels), the LV gain of the original-split models was above zero in every system; what reversed there, at 5%, were four frozen-encoder comparisons between systems or adaptation modes, none of them an LV gain (Appendix F).
  2. When. The gain was largest at 5% and shrank as MYO labels grew; for nnU-Net at 25–100% it was not resolved (−0.08 to +0.21).
  3. Why (Sections 3.2 and 4.3, Appendix E). The LV stage supervises where the ventricle lies and one of the two myocardial borders. In a boundary-wise analysis, the LV stage reduced error near both MYO borders in Pixio-L and EchoFM at 5% and 10%, more near the outer border for Pixio-L, so the gain is not confined to the shared inner border (Table 18). Before MYO training, all four LV-stage models already located the CAMUS LV cavity (LV Dice 0.91 for Pixio-L, 0.90 for EchoNet and nnU-Net, 0.85 for EchoFM; Table 19). Together with the localization result, this suggests that the LV stage supplies the ventricle's position and overall shape rather than one border.

"… it explores different orchestrations of encoders … and decoders … with the goal of finding the optimal path for training on the MYO segmentation task."

We do not search for an optimal pairing: encoders and decoders are not crossed, each system is one fixed pipeline, and the path we study is the training route within it. Our submitted abstract invited the other reading.

"The introduction of Pixio + Mask2Former seems to be a strong encoder-decoder pair for echocardiography segmentation, which is of practical value."

On CAMUS, Pixio-L with Mask2Former scored higher than EchoFM with UNETR under both routes at every fraction (at 5% after the LV stage, 81.73% vs 75.21%), although EchoFM gained more from that stage, so it is a strong low-data baseline relative to EchoFM (Section 4.4, Table 14). It is not the highest-scoring system at every fraction: trained directly, nnU-Net was not clearly different at 5% and scored higher at 10–100% (Table 9). The finding we stand behind is the LV stage itself, which raised MYO Dice at 5% and 10% in all four systems.

If passages remain hard to follow, please point us to them; if the revision addresses your concerns, we would be grateful if you would consider updating your score.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General response (part 1/3): the main result

Main result. Myocardial (MYO) labels in echocardiography are scarce, LV cavity labels are abundant, and the LV cavity border is the inner border of the myocardium. Training on LV before MYO (LV→MYO, the LV-supervised route) raised MYO Dice on CAMUS when MYO labels were scarce, in every system we tested. With 5% of the training patients on the original split, the gain was +2.29 percentage points (pp) [+1.40, +3.32] for Pixio-L with Mask2Former, +9.15 [+7.80, +10.47] for EchoFM with UNETR, +1.39 [+0.76, +2.07] for nnU-Net and +6.12 [+4.84, +7.41] for EchoNet (95% patient-bootstrap intervals); with 10% it was +2.11, +6.60, +0.60 and +2.76, every interval above zero (Section 4.3, Table 2). The gain was largest at 5% and shrank as MYO labels grew: at 25–100% it stayed above zero for Pixio-L, EchoFM and EchoNet and was not resolved for nnU-Net (−0.08 to +0.21). On one additional patient split, the gains at 5% and 10% were also above zero in all four systems (Appendix D, Table 16):

SystemOriginal 5%Original 10%Additional split 5%Additional split 10%
Pixio-L + Mask2Former+2.29 [1.40, 3.32]+2.11 [1.38, 2.96]+1.96 [1.35, 2.59]+1.80 [1.17, 2.48]
EchoFM + UNETR+9.15 [7.80, 10.47]+6.60 [5.26, 7.93]+9.50 [7.74, 11.43]+6.76 [5.27, 8.26]
nnU-Net†+1.39 [0.76, 2.07]+0.60 [0.13, 1.12]+1.85 [1.05, 2.70]+0.90 [0.24, 1.64]
EchoNet (ImageNet start)+6.12 [4.84, 7.41]+2.76 [1.91, 3.65]+8.70 [7.07, 10.41]+5.37 [4.17, 6.71]

LV-supervised minus Direct, MYO Dice (pp), full fine-tuning, 50 test patients per split. †Trained from scratch for 100 epochs. nnU-Net and EchoNet were added after review.

Design. Within each system, the Direct route (MYO training only) and the LV-supervised route use the same target-training patients at five fractions (5–100%) and three seeds and are scored on the same test patients, so every gain is a paired, within-system comparison. The paper introduces no new architecture, loss or adaptation method; the systems are not crossed, and no pairing is searched. Our submitted abstract led with a ranking and gave the LV stage one sentence; the revised abstract, introduction and conclusion now open with the result above.

AuthorsOpenReview ↗

General response (part 2/3): what was added after review

  • nnU-Net v2 2D, trained from scratch for 100 epochs, both routes, five fractions (Appendix A.2).
  • EchoNet: an ImageNet-initialized DeepLabV3-ResNet50 trained at 112×112 input with an EchoNet-derived recipe, both routes, five fractions (Appendix A.2).
  • One additional CAMUS patient split (360/90/50 patients), all four systems at 5% and 10% (Section 4.1, Appendix D). Its 50 test patients are disjoint by patient ID from the original test set, with new 5% and 10% subsets; all 8 contrasts kept the direction reported on the original split, with 95% intervals excluding zero (Table 15; each listed in our reply to 4XTF).
  • HMC-QU (Section 4.3, Appendix F): one external site (109 patient IDs, 2,349 A4C frames, clinician-verified pseudo-labels); models applied without retraining. The LV gain held there: under full fine-tuning, the LV gain of the original-split models was above zero in every system at every fraction (5%: Pixio-L +3.66 [+3.05, +4.23], EchoFM +12.58 [+11.58, +13.58], nnU-Net +4.17 [+3.00, +5.50], EchoNet +15.21 [+13.78, +16.65]; Table 8), and with the encoder frozen it was above zero in Pixio-L and EchoFM at every fraction (Table 31; at 5%, EchoFM +12.60 [+11.58, +13.65]). What reversed there, at 5%, were four frozen-encoder comparisons between systems or between adaptation modes, none of them an LV gain (Table 20): Pixio-L minus EchoFM under Direct (−9.88 [−11.14, −8.59]; also reversed at 50% and 100%) and under the LV-supervised route (−1.05 [−1.96, −0.09]), and frozen minus full fine-tuning after the LV stage for Pixio-L (−1.75 [−2.45, −1.09]; CAMUS +0.76) and EchoFM (+1.89 [+1.66, +2.13]; CAMUS −0.39). Because of them, our conclusions about system ordering and freezing stay limited to CAMUS. Pixio-L with Mask2Former had the highest mean Dice of the four systems at every fraction under both routes; at 5% Direct its difference from nnU-Net was not resolved (nnU-Net minus Pixio-L −1.76 [−3.69, +0.02]; Table 9). For models trained on the additional split (Table 21), the LV gain was above zero in seven of eight cells; the exception was nnU-Net at 10% (−3.33 [−4.96, −1.77]).
AuthorsOpenReview ↗

General response (part 3/3)

A larger gain did not mean a better final system (Section 4.4). At 5% on the original split, EchoFM gained more than Pixio-L (difference in gain +6.87 [+5.44, +8.31]; Table 24), yet Pixio-L ended higher (81.73% vs 75.21%); Pixio-L scored higher than EchoFM under both routes at every fraction (Table 14), which makes Pixio-L with Mask2Former a strong low-data baseline relative to EchoFM. With both trained directly on the original split, nnU-Net was not clearly different from Pixio-L at 5% (+0.73 [−0.30, +1.72]) and scored higher at 10–100% (+1.40 to +1.55; Table 9). No system difference is attributed to pretraining.

Why an LV stage can help (Section 3.2, Appendix E). The myocardium is the band between the endocardial border, which is also the LV cavity border, and the epicardial border, so an LV label supervises where the ventricle lies and one of the two MYO borders.

Where the gain arises (Section 4.3, Appendix E). Although the LV stage used no CAMUS image, all four LV-stage models already located the CAMUS LV cavity (LV Dice 0.91 for Pixio-L, 0.90 for EchoNet and nnU-Net, 0.85 for EchoFM; Table 19; EchoFM's released encoder was pretrained on data including CAMUS, Section 5). A boundary-wise analysis splits the paired Dice gain into error near each reference border. On the original split, the LV stage reduced error near both MYO borders in Pixio-L and EchoFM at 5% and 10%, more near the outer border for Pixio-L (and for EchoFM at 5% in Dice points), so the gain is not confined to the shared inner border (Table 18). Together with the localization result, this suggests that the LV stage supplies the ventricle's position and overall shape rather than one border.

Low-rank adaptation (Section 4.6, Appendix H.1, Table 29). With LoRA of the encoder at 5% (rank 4; 393,216 parameters per system; added after review), Pixio-L was 8.03 points below full fine-tuning without the LV stage (−8.03 [−9.30, −6.83]) and 1.12 above it with the stage ([+0.51, +1.74]), so its LV gain under LoRA was +11.44 [+9.97, +13.07] (full fine-tuning: +2.29). For EchoFM the LV gain was +8.86 [+7.57, +10.11] (full: +9.15), and LoRA differed little from full fine-tuning.

Changes in the revision. No number from the submitted version changed; the marks in the appendix tables of the pdfdiff come from renaming the route "Transfer" (in the text also "LV-to-MYO") to "LV-supervised".

  1. Abstract, introduction and conclusion lead with LV→MYO under MYO-label scarcity.
  2. A short rationale in Section 3.2 and where the gain arises in Section 4.3; the boundary-wise analysis and zero-shot LV localization in Appendix E.
  3. nnU-Net, EchoNet, the additional split, HMC-QU and LoRA: short main-text mentions, full results in Appendices B, D, F and H.
  4. Limitations (Section 5): not evaluated in this manuscript are architecture-matched controls, Pixio-L or EchoFM trained from scratch, other foundation models, adapters other than LoRA, selected-layer tuning, head reinitialization after LV training, freezing during the LV stage, and other views; LoRA covers two systems, 5% and one rank; one additional split only.
  5. The submitted subsection "Data availability and practical interpretation" was removed; its contrasts remain in Table 13 (Appendix C).
  6. Figures: the submitted Figures 2 and 3 (full fine-tuning and frozen encoder, CAMUS) are merged into Figure 2, which adds nnU-Net and EchoNet (full fine-tuning) and HMC-QU and shows seed means (SDs in the tables). New Figure 3 shows error examples at 5% for the four systems on CAMUS and HMC-QU. Appendices I and J hold all fractions and a second gallery patient of these examples, the full HMC-QU subgroup tables, and HMC-QU tables of absolute Dice and of the LV gain under both adaptation modes (Tables 30 and 31).

Point-by-point answers follow under each review.

The authors now frame the contribution as controlled evidence on intermediate-task supervision under label scarcity, not a network comparison. Does this, plus the four-system, two-split and external-site results, change your views? Please say whether you’ll keep or update your score.

f6Qc: your review has no questions and a brief weaknesses section. Could you say what would move this paper above the bar for a general ML venue?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.