From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift

Sara Inácio; Hugo Proença; João C. Neves

cs.CVOct 15, 2026

From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift

Authors: Sara Inácio, Hugo Proença, João C. Neves

Organizations: University of Beira Interior, Portugal · NOVA LINCS · IT: Instituto de Telecomunicações

Abstract

Foundation models have shown significant potential across diverse fields and have been applied to medical imaging, but their robustness across patient populations remains unclear. In this work, we investigate foundation models' ability to generalize from adult to pediatric chest X-ray classification. To the best of our knowledge, this is the first benchmark to compare multiple foundation models under a common adult-to-pediatric evaluation setting while also analyzing the effect of adaptation depth. We harmonize VinDr-CXR and PediCXR into a five-label classification task and evaluate six foundation models and two CNN baselines. All models are adapted using adult data only and tested in both the in-domain and pediatric domains using linear probing, partial fine-tuning, and full fine-tuning. The benchmark reveals a clear and consistent trend: all models suffer a substantial performance degradation when transferred from adult to pediatric data, with drops exceeding 10 percentage points. While Macro-AUROC values reach approximately 92% on the adult test set, performance decreases to around 78% on the pediatric domain. Another important finding is that increasing the depth of adaptation does not significantly improve cross-population generalization. Moreover, strong in-domain performance does not translate into better performance on pediatric data.

Figures & tables

Figure 1: Illustration of the domain shift between adult and pediatric chest X-rays. The model is trained on the adult source domain and evaluated on the pediatric target domain.

Target label	VinDr-CXR	PediCXR
No Finding	No finding	No finding
Other Diseases	Other diseases	Other diseases
Lung Tumor	Lung tumor	Lung tumor
Pneumonia	Pneumonia	Pneumonia, Broncho-pneumonia, Pleuro-pneumonia
Tuberculosis	Tuberculosis	Tuberculosis
Excluded	COPD	Remaining global labels

Table 1: Label harmonization between VinDr-CXR [ 16 ] and PediCXR [ 21 ] for the five-label multi-label classification task. Labels without a corresponding category in the other dataset are excluded from the target space.

Figure 2: Overview of the proposed domain shift evaluation methodology. Models are trained or adapted using chest X-rays from the seen adult source domain and subsequently evaluated on the unseen pediatric target domain, without target-domain training.

Model	Visual Backbone	Total params.	LP trainable	PFT trainable
DenseNet-121 [ 9 ]	–	6.95M	5.1K	2.17M
EfficientNet-B0 [ 25 ]	–	4.01M	6.4K	1.14M
CLIP [ 22 ]	ViT-B/16	86.20M	2.6K	7.49M
DINOv2 [ 17 ]	ViT-B/14	86.58M	3.8K	7.09M
BiomedCLIP [ 35 ]	ViT-B/16-224	86.19M	2.6K	7.49M
MedCLIP [ 29 ]	ViT-Swin-tiny	27.52M	3.8K	14.19M

Table 2: Model configurations across adaptation strategies. Model architectures, total parameter counts, and trainable parameters under linear probing (LP) and partial fine-tuning (PFT). Under full fine-tuning, all model parameters are updated.

Model	Cross-Domain (A $\rightarrow$ P)			In-Domain (A $\rightarrow$ A)			$\Delta_{\mathrm{AUROC}}\downarrow$
Model	Macro AUROC	mAP	Macro F1	Macro AUROC	mAP	Macro F1	$\Delta_{\mathrm{AUROC}}\downarrow$
Linear Probing
DenseNet-121 [ 9 ]	56.11 $\pm$ 0.77	17.63 $\pm$ 0.11	15.38 $\pm$ 0.50	77.60 $\pm$ 0.09	41.58 $\pm$ 0.04	38.24 $\pm$ 1.02	21.49
EfficientNet-B0 [ 25 ]	46.84 $\pm$ 2.17	18.42 $\pm$ 0.21	20.47 $\pm$ 0.47	74.40 $\pm$ 0.51	36.81 $\pm$ 0.34	37.91 $\pm$ 0.26	27.56
CLIP [ 22 ]	52.98 $\pm$ 0.19	18.29 $\pm$ 0.01	21.07 $\pm$ 0.26	73.23 $\pm$ 0.01	37.43 $\pm$ 0.04	38.94 $\pm$ 0.16	20.26
DINOv2 [ 17 ]	57.15 $\pm$ 0.11	20.27 $\pm$ 0.07	22.09 $\pm$ 0.74	83.39 $\pm$ 0.02	47.31 $\pm$ 0.06	45.21 $\pm$ 0.59	26.23

Table 3: Performance under the adult-to-pediatric domain shift (

A\rightarrow P

) and in-domain adult evaluation (

A\rightarrow A

) across the three adaptation protocols. Performance metrics (%) are reported as mean

\pm

standard deviation over three random seeds.

\Delta_{\mathrm{AUROC}}

is computed from the corresponding unrounded mean AUROC values and is defined as

\Delta_{\mathrm{AUROC}}=\mathrm{AUROC}_{A\rightarrow A}-\mathrm{AUROC}_{A\rightarrow P}

, where lower values indicate a smaller performance degradation under the domain shift. Best and second-best mean values within each adaptation strategy are shown in bold and underlined, respectively.

Figure 3: Pediatric AUROC vs. Domain-shift gap. Colors denote models; circles, squares, and triangles denote linear probing, partial fine-tuning, and full fine-tuning, respectively. Lines connect the three protocols for each model. Lower gaps and higher AUROC are better.

Figure 4: Adult in-domain versus pediatric AUROC across adaptation strategies. The dashed line indicates equal in-domain and cross-domain performance, and the vertical dotted lines show the corresponding domain-shift gap.

Medical Imaging Foundation Models

Medical imaging foundation models learn reusable representations for tasks across clinical images, modalities, and institutions.

See all 234 papers

Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift
Jul 11, 2026Giang Nguyen, Raghav Mehta, Emma A. M. Stanley +4Medical Image Classification Medical Imaging Foundation Models
College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam · Imperial College London, London, UK · Radiology Department, Vietnam National Cancer Hospital, Hanoi, Vietnam +2
Cross-dataset transportability of pediatric chest X-ray deep learning across three countries: discrimination, calibration, operating-point failure, and limited-label recovery
Sep 4, 2026Nazim-E-AlamMedical Image Classification Cross-Dataset Generalization
Department of Computer Science, American International University-Bangladesh (AIUB), 408/1 (Old KA 66/1), Kuratoli, Khilkhet, Dhaka 1229, Bangladesh
Subgroup performance analysis of adaptation strategies for chest X-ray foundation models
Aug 19, 2026Dhruv Gupta, Emma A. M. Stanley, Fabio De Sousa Ribeiro +2Chest X-Ray Classification Medical Imaging Foundation Models
Department of Computing, Imperial College London, UK · Causality in Healthcare AI Hub, UK · Department of Radiology, Royal Brompton Hospital, UK +1

Jul 11, 2026J/K move · Enter open · S save

Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift

Giang Nguyen, Raghav Mehta, Emma A. M. Stanley, Tian Xia+3

College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam · Imperial College London, London, UK · Radiology Department, Vietnam National Cancer Hospital, Hanoi, Vietnam+2

Medical Image Classification Medical Imaging Foundation Models OOD Generalization Distribution Shift Robustness Mammography Medical Image Benchmarks+1

Open paper AskPDF arXiv

Foundation models are increasingly used as image feature extractors for mammography, but their robustness under external domain shift remains unclear. We benchmark 15 foundation-model backbones across breast density, BI-RADS severity, and cancer status using a unified frozen-backbone linear-probe protocol, training on 3 source datasets and evaluating on 12 task-compatible out-of-distribution (OOD) datasets after label harmonization. Mammography-specific vision-language models (Mammo-FM and MaMA) provide the strongest mean OOD performance, but robustness is not explained by mammography exposure alone. DINOv3 remains a competitive vision-only baseline, and mammography-adapted pretraining does not consistently improve generalization. Dataset-level analysis further shows that even leading models show heterogeneous performance across datasets. Feature-space inspection reveals that useful representations can preserve clinical signal while retaining dataset and acquisition structure. These findings highlight dataset-level OOD evaluation as a central criterion for assessing mammography representations. Our code is publicly available: https://github.com/biomedia-mira/mammo-ood.

Abstract

Medical Imaging Foundation Models

Explore similar work

Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift

Cross-dataset transportability of pediatric chest X-ray deep learning across three countries: discrimination, calibration, operating-point failure, and limited-label recovery

Subgroup performance analysis of adaptation strategies for chest X-ray foundation models