Which Pretext Task Transfers? Self-Supervised Pretraining Objectives for Lung Ultrasound
Moein Heidari, Junbo Rao, Jai Choraria, Wenjin Chen, David J. Foran, Ilker Hacihaliloglu
cs.CV
Quick verdict
Same-backbone SSL comparison for lung ultrasound: POCUS winners fail to transfer while MoCo does — preliminary result, underpowered and baseline-free.
Summary
This preliminary study compares three self-supervised pretraining objective families for lung ultrasound (LUS) video under controlled conditions: contrastive learning (MoCo v3 adapted to video), masked reconstruction (VideoMAE, 90% tube masking), and latent prediction (V-JEPA). All share a randomly initialized ViT-S/16 backbone, the same COVID-BLUeS pretraining corpus, optimizer, schedule, and frozen-probe protocol. Encoders are evaluated with linear, kNN, and attentive probes at 5-100% label budgets for three-class classification on POCUS under patient-level five-fold cross-validation, and on the independently acquired Mendeley-Uganda dataset excluded from all training. At the full budget with linear probing, VideoMAE (66.5 plus/minus 13.1) and V-JEPA (65.4 plus/minus 11.7) lead on POCUS while MoCo (42.1 plus/minus 1.2) is near chance; on Mendeley-Uganda the ranking reverses: MoCo 62.7 plus/minus 1.0, VideoMAE 53.8 plus/minus 2.8, V-JEPA 35.1 plus/minus 4.9 (chance is 33.3). The authors conclude that in-distribution probe accuracy does not predict cross-dataset transfer. Limitations: one backbone scale, one classification task, no supervised or random-init baselines, no significance testing amid fold SDs up to 23.6 points, a single external dataset, and representation-level analyses deferred to a future paper; the work is explicitly preliminary.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under identical backbone, pretraining corpus, schedule, and frozen evaluation, the choice of SSL pretext objective changes both in-distribution accuracy and cross-dataset transfer for lung ultrasound video.
- At the 100% label budget with linear probing, VideoMAE (66.5+/-13.1) and V-JEPA (65.4+/-11.7) outperform MoCo (42.1+/-1.2) balanced accuracy on POCUS.
- On Mendeley-Uganda the ranking reverses: MoCo is best (62.7+/-1.0), VideoMAE drops to 53.8+/-2.8, and V-JEPA falls near chance (35.1+/-4.9).
- POCUS probe accuracy alone does not identify which pretext objective transfers best to a new acquisition setting.
Strengths
Controlled apples-to-apples design
All three objectives share the ViT-S/16 backbone, COVID-BLUeS corpus, optimizer, schedule, and random (non-ImageNet) initialization, and are compared only via frozen probes (Sec. 2.1-2.3), so measured gaps are attributable to the pretext objective rather than corpus or architecture confounds.
Leakage-aware and truly external evaluation
Patient-level five-fold CV on POCUS guards against inflation from near-duplicate adjacent frames (explicitly justified in Sec. 2.3), and Mendeley-Uganda is excluded from both SSL pretraining and probe fitting, making it a genuine transfer test rather than a re-split.
Dense, transparent results grid
Honest scoping of a preliminary result
Weaknesses
No reference baselines
No inferential statistics behind ranking claims
Unexamined MoCo failure mode
Single external dataset and single task
Mechanistic story deferred
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Which self-supervised objective works best for lung ultrasound?
It depends on the evaluation set. On the in-distribution POCUS dataset at full labels with a linear probe, masked reconstruction (VideoMAE, 66.5%) and latent prediction (V-JEPA, 65.4%) beat contrastive MoCo (42.1%) in balanced accuracy. On the external Mendeley-Uganda dataset the order reverses: MoCo is best (62.7%), VideoMAE reaches 53.8%, and V-JEPA drops to about chance level (35.1%, chance being 33.3% for three classes).
What datasets does the paper use and how are they split?
COVID-BLUeS lung ultrasound videos are used only for self-supervised pretraining. POCUS (COVID-19 vs bacterial pneumonia vs healthy) is used for downstream evaluation with patient-level five-fold cross-validation, so frames from the same patient never straddle train and test folds. Mendeley-Uganda is an independently acquired dataset used purely for external testing and is excluded from both pretraining and probe training.
Why does V-JEPA fail on the external dataset?
The paper does not yet know. It speculates that JEPA-style latent prediction has higher sample-complexity requirements and may be less robust to a different acquisition distribution, dropping from 65.4% on POCUS to 35.1% (near chance) on Mendeley-Uganda. The planned attention-map and similarity-map analyses are meant to answer this mechanistically, but they are not included in this preliminary version.
Is this a complete study?
No. The authors explicitly frame it as preliminary: only one backbone scale (ViT-S/16), only three-class classification (which does not test the temporal cues that motivate video pretraining), and the representation-level analyses are deferred to a planned full paper. Code is released at github.com/moeinheidari7829/LUSVideoSSL.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are available only via their captions; the visual layout of Fig. 2c (ranking flip) could not be independently verified.
- Implementation details (epochs, learning rate, batch size, COVID-BLUeS clip counts, JEPA masking ratio) do not appear in the extracted text, so the reproducibility rating relies on the stated public code release.
- Some non-ASCII author names show OCR artifacts (e.g., 'Ivezi´c'); meaning was recoverable from context.
- The paper self-identifies as preliminary; conclusions may change in the announced full version with representation-level analyses.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.