PaperlayerView exact revision on arXiv

Source-linked AI summary

Wrong Organ, Right Physics: Transferring Echocardiography Pretraining to Lung Ultrasound for Tuberculosis Screening

Christiaan M. Geldenhuys, Joshua M. Jansen van Vüren, Véronique Suttels, Trevor Brokowski, Ablo P. Wachinou, Mary-Anne Hartley, Rensu P. Theart, Grant Theron, Thomas R. Niesler

arXiv:2610.03290v1eess.IVcs.CVcs.LG

TL;DR

LUS is a potential primary-care TB screening tool, but labelled data are scarce and interpretation remains a challenge. The study compares fixed pretrained encoders and feature conditioning for LUS classification; conditioning improves all seventeen encoders, while the cohort cannot resolve encoder differences.

  • Problem

    LUS interpretation remains a challenge, and the study asks whether echocardiography-pretrained representations can transfer to LUS despite the small labelled cohort.

    Source: §I · p. 1, A. Lung ultrasound for tuberculosis… · p. 2

  • Method

    The study compares seventeen fixed encoders for LUS screening, training a patient-level classifier and evaluating feature conditioning.

    Source: A. Input preprocessing and encoders · p. 4, C. Training and hyperparameter… · p. 5, §VII · p. 8

  • Results

    Feature conditioning improved all seventeen encoders, and every encoder selected on development folds exceeded the baseline on held-out test AUROC by up to +2.57 percentage points.

    Source: §VII · p. 8

  • Takeaways & Limitations

    The development folds establish feature conditioning as an intervention, but this cohort cannot settle which encoder is best.

    Source: §VII · p. 8

    With replicated still frames as video input, the study cannot separate a pretraining-domain effect from applying a video encoder to static frames.

    Source: §VII · p. 9

Abstract

from arXiv · show

Lung ultrasound (LUS) is attractive for tuberculosis (TB) screening at primary-care level, but labelled cohorts are small. Echocardiography carries no such constraint, while sharing the same underlying ultrasound imaging physics, signal processing and B-mode appearance as LUS. We ask whether an encoder pretrained on that high-resource ultrasound domain carries representations that remain usable in the low-resource one. Only the encoder varies, across seventeen encoders spanning three architecture families. Among them, a latent-predictive video encoder pretrained on generic video (V-JEPA2-L) and its echocardiography counterpart (EchoJEPA-L) differ in pretraining corpus alone. The choice among these encoders does not resolve the classification, the whole family spanning 2.50 percentage points against a measurement resolution of 2.71. What moves the task instead is feature conditioning. Standardising the features between the encoder and the classifier improves all seventeen encoders by a mean of +1.23 percentage points at $p=1.5\times10^{-5}$. On the held-out test set every encoder selected on the development folds stands above the baseline system by up to +2.57 percentage points of area under the receiver operating characteristic curve (AUROC), and specificity at 90% sensitivity reaches 79.3% against 60.3%. The contrast specified in advance, EchoJEPA-L against V-JEPA2-L, measures -0.16 percentage points at $p=0.926$. We therefore find no evidence that shared ultrasonic physics alone makes echocardiography a more productive pretraining corpus than generic video, and any advantage, if present, is smaller than this cohort can resolve. The video encoders receive replicated still images, however, so whether this absence of an effect reflects the pretraining domain or a video encoder applied to static frames cannot be separated. The limiting factor is the labelled cohort rather than the encoder.

I. INTRODUCTION

LUS could support primary-care TB screening, but its interpretation lacks a strong evidence base and labelled cohorts are scarce. This study tests whether echocardiography pretraining transfers to LUS and separates encoder choice from feature conditioning.

I. INTRODUCTION

  • LUS offers a portable, radiation-free complement for primary-care screening, but interpretation remains operator-dependent and labelled data are scarce.The cohort has 504 labelled patients, while public echocardiography corpora are several orders of magnitude larger.Source: §I · p. 1
  • The study tests whether echocardiography-pretrained representations remain useful for LUS despite differences in anatomy, probe placement, and acquisition protocol.Echocardiography shares ultrasound imaging physics with LUS but not its anatomy.Source: §I · p. 1

A. Contributions •

  • The paper tests whether shared imaging physics without shared anatomy makes echocardiography a productive pretraining corpus for LUS.The test uses two encoders that differ in pretraining corpus alone.Source: A. Contributions • · p. 2
  • The study compares encoder architectures and pretraining corpora for TB classification while exchanging only the encoder.The comparison is conducted under a protocol that holds the remaining components constant.Source: A. Contributions • · p. 2
  • The study isolates feature conditioning by pairing each encoder against itself with and without conditioning.Source: A. Contributions • · p. 2

A. Lung ultrasound for tuberculosis screening

LUS interpretation lacks consensus, and transfer from other imaging domains remains incompletely validated across anatomy and acquisition shifts. This section reviews latent-predictive learning, available ultrasound encoders, and the rationale for fixed-encoder transfer and feature conditioning.

A. Lung ultrasound for tuberculosis screening

  • A review found no consensus on acquiring or analysing pulmonary TB ultrasound images and concluded their diagnostic utility remains unclear.Automated analysis could fill a role similar to automated chest-radiograph analysis without radiographic infrastructure or radiation control.Source: A. Lung ultrasound for tuberculosis… · p. 2

B. Latent-predictive self-supervised video representation learning

  • Pixel-reconstruction objectives must reproduce ultrasound speckle and other acquisition artefacts, whereas latent prediction moves the loss into representation space [6] [13].The discussion describes I-JEPA, V-JEPA [16], V-JEPA 2 [17], and echocardiography-pretrained EchoJEPA [6].Source: B. Latent-predictive self-supervised… · p. 2

C. Large-scale pretrained ultrasound and echocardiography encoders

  • US-JEPA was trained on over five million frames across 22 anatomies, including lung, but its weights are unavailable for comparison [13].The work asks whether an ultrasound-native latent-predictive model can be trained, rather than whether an existing echocardiography model can be repurposed.Source: C. Large-scale pretrained ultrasound… · p. 2

D. Cross-domain transfer

  • Validation of medical-imaging transfer across anatomy and imaging technique shifts remains inconsistently documented [24].Here, echocardiography and LUS share image-formation physics, but LUS is read largely from artefacts such as A-lines and B-lines.Source: D. Cross-domain transfer · p. 2
  • The study freezes encoder parameters because the labelled cohort is small and anatomy and acquisition shifts are large.Feature conditioning addresses whether a shortfall reflects representation quality or feature scale and spread.Source: D. Cross-domain transfer · pp. 2–3

E. Multiple-instance learning for patient-level scoring

Because patients have one TB diagnosis but images lack individual labels, the paper scores each patient from an unordered set of site images. The cohort uses patient-level labels and includes examples showing that positive patients may have different appearances at the same site.

E. Multiple-instance learning for patient-level scoring

  • Attention-based multiple-instance learning pools image embeddings into one patient-level TB score without requiring disease findings at every site.Its embedding-dependent weights make the pooling permutation invariant and compatible with variable-size image sets.Source: E. Multiple-instance learning for… · p. 3

III. DATA

  • Patients carry a single PTB label, defined by bacteriologically confirmed or clinically diagnosed TB, rather than labels for individual images.Images are acquired at 14 anatomical sites per patient using a portable probe and standardised sliding-scan protocol.Source: §III · p. 3
  • The cohort table describes its composition and identifies patient-disjoint folds, with the sixth fold held out as the test set.Source: Table I · p. 3
  • Two PTB-positive patients imaged at the same anterior apical site show aerated lung and consolidation appearances, respectively.This contrast illustrates why patient-level aggregation must accommodate an aerated appearance at an individual site in a patient with confirmed disease.Source: Figure 1 · p. 3
  • The echocardiography encoders were pretrained on MIMIC-IV-ECHO, a single-centre corpus of transthoracic studies.Source: §III · p. 3

IV. EXPERIMENTAL SETUP

The experiment keeps the pipeline constant and varies only the fixed encoder, training the patient-level classifier. Encoders share image preprocessing, while latent-predictive video models receive replicated still frames; the set spans multiple architectures and pretraining sources.

IV. EXPERIMENTAL SETUP

  • Only the encoder varies across the seventeen comparisons, and its parameters remain fixed while the classifier is trained.The experimental procedure is otherwise held identical across encoders.Source: §IV · p. 3
  • Each patient contributes an unordered set of cleaned site images, encoded independently before optional conditioning and attention pooling produce a TB score.The pooling includes an additive site embedding.Source: Figure 2 · p. 4

A. Input preprocessing and encoders

  • Images are padded to square and resized to 224 px × 224 px; latent-predictive video encoders instead receive 16 identical frames.Other encoders take the resized image directly.Source: A. Input preprocessing and encoders · p. 4
  • The seventeen encoders span convolutional networks, supervised and self-supervised natural-image transformers, and latent-predictive video encoders.The convolutional group includes ImageNet-1k-supervised ResNets, while transformer variants use different pretrained representations.Source: A. Input preprocessing and encoders · p. 4
  • V-JEPA2-L and EchoJEPA-L share architecture and differ only in pretraining corpus, the study's specified contrast.No encoder is pretrained on LUS, and echocardiography is the only ultrasound domain represented in pretraining.Source: A. Input preprocessing and encoders · p. 4

B. Patient-level classifier

All encoders feed a shared patient-level multiple-instance classifier, with training and evaluation designed to compare fixed representations under consistent conditions. The experiments test encoder choice and feature conditioning using development folds and paired held-out comparisons.

B. Patient-level classifier

  • The patient-level classifier projects image features to 512 dimensions, adds anatomical-site embeddings, and attention-pools each patient's image set into one TB score.The same classifier architecture is used across encoders, with 2.9 million parameters beyond the input projection held identical.Source: B. Patient-level classifier · p. 5

C. Training and hyperparameter selection protocol

  • Five patient-disjoint development folds select each encoder's configuration and stopping epoch by mean held-out-fold loss after searching 45 hyperparameter configurations per feature condition.The sixth fold is the unchanged 101-patient test partition; training uses 20 epochs, with a five-seed repeatability check for four encoders.Source: C. Training and hyperparameter… · p. 5

D. Research questions

  • The experiments ask whether any fixed encoder improves on the baseline ResNet-18 and whether conditioning its features improves classification.Feature conditions are raw features, per-dimension standardisation, batch normalisation, and layer normalisation.Source: D. Research questions · p. 5

E. Metrics, estimand and statistical treatment

  • Held-out contrasts pair encoders on the same patients, while standardisation is tested across the seventeen paired encoders with a Wilcoxon signed-rank test.Held-out AUROC uses the ensemble of five development-fold models; test-set power is limited with 43 positives and 58 negatives, and specificity changes in one-negative steps.Source: E. Metrics, estimand and statistical… · p. 5

V. RESULTS

On the held-out test set, every encoder selected on development folds exceeds the baseline in AUROC, while the echo-versus-generic-video contrast is small and uncertain. Feature-specific operating-point results and the baseline retraining provide additional context for these comparisons.

A. The baseline system

  • The retrained baseline achieves mean development AUROC 0.8717 (between-fold standard deviation 0.0389) and held-out AUROC 0.8881 [0.8209, 0.9554].The earlier published figures differ slightly because their positive class included only bacteriologically confirmed TB cases.Source: A. The baseline system · p. 6, B. Encoder choice · p. 6

B. Encoder choice

  • Every encoder selected on development folds exceeds the baseline on held-out AUROC, with gains ranging from +0.24 to +2.57 percentage points.The supervised ViT-L/16 has the largest margin (+2.57); ResNet-101 has the smallest (+0.24).Source: B. Encoder choice · p. 6
  • EchoJEPA-L versus V-JEPA2-L measures −0.16 percentage points on the held-out test set, reversing the development-fold direction.The contrast has p = 0.926 and an interval of [−3.54, +3.21] percentage points.Source: B. Encoder choice · p. 6
  • At 90 % sensitivity, four of five selected encoders reach 69 % to 79.3 % specificity, versus 60.3 % for baseline; ResNet-101 reaches 56.9 %.The reported specificity gain is up to +18.97 percentage points.Source: B. Encoder choice · p. 6
  • Among held-out contrasts, only supervised ViT-L/16 versus baseline reaches statistical significance, at p = 0.040.The ResNet-101 margin has p = 0.866, and every paired specificity difference spans zero.Source: B. Encoder choice · p. 6

C. Feature conditioning

Feature standardisation improves performance across the encoder family, while encoder-to-encoder differences remain within the study's measurement resolution. The held-out evidence supports gains over baseline but does not resolve an advantage for echocardiography pretraining over generic video.

C. Feature conditioning

  • Standardisation improves all seventeen encoders by a mean +1.23 percentage points, with bootstrap interval [+0.83, +1.70]; development-fold Wilcoxon p = 1.5 × 10^-5.The gain is 48 % of the best-to-worst encoder spread; batch normalisation and standardisation differ by −0.11 points at p = 1, while layer normalisation is below raw features by −0.35 points at p = 0.049.Source: C. Feature conditioning · p. 6

VI. DISCUSSION

  • Fold assignment shifts mean development AUROC by 7.30 percentage points; twice the fold-mean standard error is 2.71 points, the smallest encoder difference resolvable, and no adjacent pair clears it.The full family’s best-to-worst spread stays within that resolution under either feature condition.Source: §VI · p. 6

A. Encoder choice (RQ1)

  • With conditioned features, every selected encoder improves on baseline held-out AUROC, including transformer gains of +2.57 and +2.37 percentage points.Their specificity gains at 90 % sensitivity are +13.79 and +18.97 percentage points.Source: A. Encoder choice (RQ1) · p. 6
  • Replacing ResNet-18 with a selected encoder is supported by their consistent direction over baseline, but choosing among replacements requires more recruitment.The comparison does not establish that one replacement is better than another because the observed band is within what the 101 test patients can measure.Source: A. Encoder choice (RQ1) · p. 6
  • If specificity rose from 60.3 % to 79.3 % at 90 % sensitivity, the point estimate is 171 to 188 fewer people sent for a second screening test.These point estimates come from one tertiary-referral cohort and the specificity gain is unresolved, so they indicate possible scale rather than an expected practice reduction.Source: A. Encoder choice (RQ1) · p. 7
  • The EchoJEPA-L versus V-JEPA2-L contrast shows no statistical advantage for echocardiography pretraining, and its interval is too wide to establish equivalence.The held-out contrast changes sign relative to development folds; the study cannot separate pretraining-domain effects from applying video encoders to replicated still images.Source: A. Encoder choice (RQ1) · p. 7

B. Feature conditioning (RQ2)

Dataset-wide feature conditioning matters more than the choice of encoder: the two normalisers that estimate across the dataset perform alike, while layer normalisation does not improve raw features.

B. Feature conditioning (RQ2)

  • Dataset-wide scale information, rather than a trainable affine transform, is the active ingredient in feature conditioning.The two dataset-wide normalisers perform alike; batch normalisation is more practical because its statistics are accumulated during training.Source: B. Feature conditioning (RQ2) · p. 7
  • Layer normalisation shows no improvement over raw features because it remains confined to one image at a time.By contrast, the effective normalisers estimate statistics across the dataset.Source: B. Feature conditioning (RQ2) · p. 7
  • Table IV pairs normaliser conditions within each encoder, with positive deltas favouring the first condition.Development AUROC is averaged across all 17 encoders; the reported delta is the mean with a bootstrap 95% interval, and p-values come from paired tests.Source: Table IV · p. 8, B. Feature conditioning (RQ2) · p. 8

C. Measurement resolution and limitations

Encoder differences and WHO operating-point status remain unresolved at this cohort’s measurement scale, while single-centre recruitment limits how far the results generalise.

C. Measurement resolution and limitations

  • Reseeding shifts encoder performance by 0.28–0.68 percentage points, and one encoder selects five configurations across five seeds despite a 0.70-point mean development AUROC shift.For that encoder, configuration selection is better understood as a lottery over a flat region than a stable property.Source: C. Measurement resolution and… · p. 8
  • One negative patient changes held-out specificity by 1.72 percentage points, and every interval crosses the Type 2 minimum.The cohort therefore cannot resolve whether any system sits above or below the Type 2 specificity target.Source: C. Measurement resolution and… · p. 8
  • The held-out evaluation is internal validation: all patients come from one tertiary centre and one probe under one standardised protocol.The 44.6% positive fraction may differ from primary-care disease spectrum, potentially shifting sensitivity and specificity.Source: C. Measurement resolution and… · p. 8

VII. SUMMARY AND CONCLUSION

Feature conditioning improved all seventeen encoders, while encoder choice remained unresolved at the cohort’s measurement resolution. Shared imaging physics alone did not establish echocardiography as a more productive pretraining corpus under static-frame evaluation.

VII. SUMMARY AND CONCLUSION

  • The study compared seventeen pretrained encoders while holding encoder parameters fixed and varying only the encoder in the evaluation protocol.The held-out test partition contains 101 patients, with five development-fold models ensembled by averaging patient scores.Source: §VII · p. 8
  • Every encoder selected on development folds exceeded the baseline on held-out testing, by up to +2.57 percentage points of AUROC and +18.97 percentage points of specificity at 90% sensitivity.Figure 3 depicts the held-out comparisons and operating points against the Type 2 targets.Source: §VII · p. 8, Figure 3 · p. 9
  • The authors found no detected advantage for echocardiography pretraining from shared imaging physics under static-frame evaluation.They note that whether this result reflects the pretraining domain or video encoders applied to still frames cannot be separated.Source: §VII · p. 9

ETHICS AND DATA AVAILABILITY

  • The study received ethics approval from the University of Parakou committee on 18 May 2021, and participants gave written informed consent.A trusted representative assisted participants unable to read or write.Source: ETHICS AND DATA AVAILABILITY · p. 9
  • De-identified images and patient-level labels are available for research use under a data use agreement.The materials are deposited under a persistent identifier.Source: ETHICS AND DATA AVAILABILITY · p. 9
  • The echocardiography and generic-video encoders were used unmodified, while MIMIC-IV-ECHO access is credentialed and governed by a data-use license.No echocardiography frame was processed or reproduced in this work.Source: ETHICS AND DATA AVAILABILITY · p. 9

ORCID

The section lists ORCID identifiers for nine authors.

ORCID

  • Christiaan M. Geldenhuys (0000-0003-0691-0235), Joshua M. Jansen van Vüren (0000-0002-3406-4788), and Véronique Suttels (0000-0002-5261-0439) are listed with ORCID identifiers.Source: ORCID · p. 9
  • Trevor Brokowski (0000-0002-7547-7749), Ablo P. Wachinou (0000-0002-1024-9665), and Mary-Anne Hartley (0000-0002-8826-3870) are listed with ORCID identifiers.Source: ORCID · p. 9
  • Rensu P. Theart (0000-0003-0508-1690), Grant Theron (0000-0002-9216-2415), and Thomas R. Niesler (0000-0002-7341-1017) are listed with ORCID identifiers.Source: ORCID · p. 9

Related papers

Loading 2610.03290v1…