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

eess.IVOct 2, 2026

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

Authors: Christiaan M. Geldenhuys, Joshua M. Jansen van Vüren, Véronique Suttels, Trevor Brokowski, Ablo P. Wachinou, Mary-Anne Hartley, Rensu P. Theart, Grant Theron, +1 more

Organizations: Department of Electrical and Electronic Engineering, University of Stellenbosch, Stellenbosch, South Africa · Laboratory for intelligent Global Health and Humanitarian Response Technologies (LiGHT) Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland · National Teaching Hospital for Tuberculosis and Pulmonary Diseases (CNHU-PPC), Cotonou, Benin · South African Medical Research Council Centre for Tuberculosis Research (CTR) Division of Molecular Biology and Human Genetics, Faculty of Medicine and Health Sciences Stellenbosch University, Cape Town, South Africa

Abstract

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.

Figures & tables

Property	Value
Cohort
Patients with a label	504
Patients with cleaned images	502
Cleaned images	$13\,458$
Images per patient, median	28
Anatomical site codes	14

TABLE I : Composition of the lung ultrasound cohort. Folds are patient-disjoint, and the sixth is held out as the test set.

Panel of Figure 2 — Fig. 1 : Two cleaned images from the target cohort, both acquired at the same anterior apical site. In 1(a) a bright pleural line lies above a stack of equidistant horizontal reverberation artefacts, the appearance of aerated lung. In 1(b) the region beneath the pleural line carries a tissue-like echotexture broken by punctate hyperechoic foci, the appearance of consolidation. Both patients carry a positive PTBlabel. A patient with confirmed disease can therefore present an aerated appearance at any individual site, which is what the patient-level aggregation is designed to accommodate (see Section IV-B ).

Fig. 2 : Experimental procedure, in which only the encoder varies. Each patient contributes an unordered set of cleaned LUSimages, one per anatomical site, each padded to square, resized to

224\text{\times}224\text{\,}\mathrm{px}

and encoded independently. A latent-predictive video encoder instead receives the image replicated to

16

identical video frames. The features may pass through an optional conditioning stage, and attention pooling with an additive site embedding aggregates the set into a single TBscore. The four encoders drawn here stand for the seventeen compared, each evaluated with its parameters held fixed.

		Dev. AUROC(sd)
Encoder	Pretraining corpus	Raw features	Standardised features	$\Delta$
Baseline system [ 7 ]	ImageNet-1k	$0.8717$ $\pm$ $0.0389$	–	–
Convolutional, supervised on natural images
ResNet-18 a	ImageNet-1k	$0.8694\pm 290$	$0.8817\pm 198$	$+1.23$
ResNet-34	ImageNet-1k	$0.8567\pm 453$	$0.8680\pm 303$	$+1.13$
ResNet-50	ImageNet-1k	$0.8704\pm 409$	$0.8715\pm 447$	$+0.11$

TABLE II : Seventeen encoders on the five development folds, each evaluated with and without feature standardisation.

System or contrast

Test AUROC

\Delta

test AUROC

p

Specificity (

\mathrm{\%}

)

\Delta

spec.

Baseline system [ 7 ]

0.8881

[

0.8209

0.9554

]

–

60.3

[

23.6

87.1

]

–

Encoders against the baseline system, each paired on identical patients

ViT-L/16, supervised

0.9138

[

0.8555

0.9721

]

+2.57 [

+0.12

+5.01

]

0.040

74.1

[

25.9

91.8

]

+13.79

[

-6.0

+27.6

]

ViT-L/16, MAE

0.9118

[

0.8512

0.9724

]

+2.37

[

-0.89

+5.62

]

0.154

79.3

[

26.8

92.3

]

+18.97

[

-8.6

+33.3

]

V-JEPA2-L

0.9058

[

0.8435

0.9680

]

+1.76

[

-0.88

+4.41

]

0.191

69.0

[

28.8

91.4

]

+8.62

[

-17.0

+24.1

]

EchoJEPA-L

0.9042

[

0.8469

0.9615

]

+1.60

[

-2.09

+5.29

]

0.394

70.7

[

39.6

88.6

]

+10.34

[

-8.2

+27.1

]

TABLE III : The held-out test set of 101 patients, 43 positive and 58 negative. Every figure is a mean-probability ensemble over the five development-fold models, whose scores are averaged per patient before one AUROCis computed on those 101 patients.

Comparison	Ahead	$\Delta$ dev. AUROC			$p$
Does feature conditioning improve downstream classification?
Standard. vs raw	17/17	$+1.23$ [	$+0.83$ ,	$+1.70$ ]	$1.5\text{\times}{10}^{-5}$
Batch norm. vs raw	15/17	$+1.12$ [	$+0.67$ ,	$+1.65$ ]	$0.0023$
Layer norm. vs raw	4/17	$-0.35$ [	$-0.66$ ,	$-0.02$ ]	$0.049$
Should features be normalised across the dataset, or merely learned?
Standard. vs layer norm.	17/17	$+1.57$ [	$+1.16$ ,	$+2.00$ ]	$1.5\text{\times}{10}^{-5}$

TABLE IV : Which normaliser, paired within each encoder. Each row names two conditions and reports the first against the second, so a positive

\Delta

favours the first.

Fig. 3 : The held-out test set, on the 101 patients of Table I , for the baseline system, for EchoJEPA-L and V-JEPA2-L, and for the encoder that leads there. Upper panel: each curve is the pointwise mean over

1\,000\,000

bootstrap resamples at the

59

achievable false-positive rates, and the legend gives the empirical AUROCwith a DeLong

95\text{\,}\mathrm{\%}

interval [ 39 ] . The grey band is the pointwise

95\text{\,}\mathrm{\%}

bootstrap interval of the baseline system alone. The hatched region marks operating points meeting both Type 2 minimal targets of the 2025 screening profile of the WHO,

90\text{\,}\mathrm{\%}

sensitivity and

60\text{\,}\mathrm{\%}

specificity. Lower panel: the operating points instead, as specificity at a sensitivity of at least

90\text{\,}\mathrm{\%}

on the same axis, where the box is the interquartile range of the bootstrap, the whisker its

95\text{\,}\mathrm{\%}

interval with the threshold re-estimated inside every resample, the marker the observed value rather than the bootstrap median, and the right-hand figures the true negatives of 58. The dashed rule is the Type 2 specificity.

Explore similar work

How well do routinely collected demographic and clinical variables aid point-of-care lung ultrasound TB classification
Oct 5, 2026Joshua M. Jansen van Vüren, Christiaan M. Geldenhuys, Devendra S. Parihar +7Medical Diagnosis Medical Image Classification
Department of Electrical and Electronic Engineering, University of Stellenbosch, Stellenbosch, South Africa · Laboratory for intelligent Global Health and Humanitarian Response Technologies (LiGHT) Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland · National Teaching Hospital for Tuberculosis and Pulmonary Diseases (CNHU-PPC), Cotonou, Benin +1
Which Pretext Task Transfers? Self-Supervised Pretraining Objectives for Lung Ultrasound
Sep 15, 2026Moein Heidari, Junbo Rao, Jai Choraria +3Contrastive Learning Masked Autoencoders
University of British Columbia, Vancouver, BC, Canada · Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, USA
Hierarchy-Aware and Anatomy-Guided Learning for Lung Ultrasound Video Classification
Jul 20, 2026Alya Almsouti, Lotfi Mecharbat, Noha Aboukhater +5Hierarchical Classification Medical Image Classification
Mohamed bin Zayed University of Artificial Intelligence, UAE · SEHA, UAE · Sheikh Khalifa Medical City, UAE +2

Oct 5, 2026J/K move · Enter open · S save

How well do routinely collected demographic and clinical variables aid point-of-care lung ultrasound TB classification

Joshua M. Jansen van Vüren, Christiaan M. Geldenhuys, Devendra S. Parihar, Véronique Suttels+6

Department of Electrical and Electronic Engineering, University of Stellenbosch, Stellenbosch, South Africa · Laboratory for intelligent Global Health and Humanitarian Response Technologies (LiGHT) Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland · National Teaching Hospital for Tuberculosis and Pulmonary Diseases (CNHU-PPC), Cotonou, Benin+1

Medical Diagnosis Medical Image Classification Ultrasound Imaging Multimodal Fusion

Open paper AskPDF arXiv

We consider the fusion of lung ultrasound images with routinely-collected clinical and demographic data for the purpose of automated tuberculosis (TB) screening using deep-learning. Such deep-learning based screening tools for TB could meaningfully support the health care system in Africa, where the burden of disease is severe and resources are constrained. Beginning with an established ResNet baseline for classification of lung ultrasound images, which achieves an area under the receiver operating characteristic (AUROC) curve of 0.91 [0.86,0.96] (95% CI), we consider the incorporation of the clinical and demographic data using three fusion approaches. We find that a simple average-based fusion of the output scores of separately-trained image and clinical data classifiers consistently matches or outperforms a more complex approach where the data is fused earlier and a combined classifier is trained. Fusing the image and the clinical classifiers in this way leads to a classifier with an overall AUROC of 0.95 [0.91,0.99] (specificity of 0.76 at sensitivity 0.93) which is an improvement of 4% absolute over the image-only baseline. We also find that greedy feature selection can be used to reduce the number of clinical and demographic inputs without sacrificing classification performance. Finally, when we differentiate between clinical and demographic data that are self-reported, that require some basic measurement or calculation, and that require a point-of-care (POC) test, we find the inclusion of the POC tests included in this study to be of minimal benefit to classification performance. We conclude that the incorporation of routinely-collected clinical and demographic data is a promising way to improve the performance of lung ultrasound based automatic classification.

Abstract

Explore similar work

How well do routinely collected demographic and clinical variables aid point-of-care lung ultrasound TB classification

Which Pretext Task Transfers? Self-Supervised Pretraining Objectives for Lung Ultrasound

Hierarchy-Aware and Anatomy-Guided Learning for Lung Ultrasound Video Classification