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

arXiv:2610.21809 · cs.CV · Submitted 2026-10-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "From Adults to Children".

Jane: The gist This work investigates foundation models’ ability to generalize from adult to pediatric chest X-ray classification,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we've got the setup down. Now let's talk about what they actually found in this paper, "From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift".

Jane: They are showing that when you move from adult X-rays to pediatric ones, almost every single model suffers a drop in performance, and these drops can be quite substantial, sometimes exceeding ten percentage points two <ref:2610.21809#pg1>.

Lu: The main point they're making is that there isn't some magic way to make these models robust across populations just by tweaking them slightly.

Meng: They’ve found that the level of adaptation doesn't actually help much in improving this cross-population generalization two <ref:2610.21809#pg1>. It seems like you can either get a decent transfer, or you don't.

Lalam: They specifically observed that the choice of adaptation strategy really depends on what kind of representation the model learned during its pretraining two <ref:2610.21809#pg1>. For instance, some models benefit from stronger adaptation while others actually get worse when you fine-tune them further.

Tom: It’s interesting that they found this inconsistency between different model architectures and their performance under the adult-to-pediatric protocol described in this paper two <ref:2610.21809#pg1>.

The paper's summary: Jane: So to summarize what they are doing, the researchers are systematically comparing different foundation models against standard CNNs on a standardized adult-to-pediatric setting for chest X-ray classification.

Lu: They harmonize the datasets into that five-label classification task, and they’re testing them on both adult images and pediatric images without using any pediatric data during the adaptation phase two <ref:2610.21809#pg1,five-label classification task, and>.

Meng: The evaluation protocol is very specific: they test each model using linear probing, which is just a baseline, partial fine-tuning, and full fine-tuning.

Lalam: And their primary metric for seeing the shift is this difference in AUROC scores between adult and pediatric performance, where a lower value means less degradation two <ref:2610.21809#pg1>.

Tom: The summary highlights that they are looking at how different pretraining domains and learning objectives affect the results when evaluated under this common adult-to-pediatric protocol two <ref:2610.21809#pg1>.

Jane: So what does this mean for us right now? It shows that just having a big foundation model isn't enough; we need to see how it handles real-world domain shifts, like moving from one patient group to another.

The paper's improvements: Tom: The authors of this work suggest a few key things about this research, mainly focusing on what they learned about model adaptation.

Lu: They point out that increasing the depth of adaptation doesn't significantly improve cross-population generalization, which is a pretty direct finding two <ref:2610.21809#pg1>.

Meng: They also emphasize that the choice of adaptation strategy isn't universal; there isn't one single strategy that works for every model two <ref:2610.21809#pg1>.

Lalam: They show examples where some models like DenseNet-one hundred twenty-one and BioViL-T benefit from stronger adaptation, but others, like MedCLIP and BiomedCLIP, actually perform best with just linear probing two.

Jane: The paper suggests that strong performance on adult data doesn't automatically guarantee robustness when you move to a pediatric setting two <ref:2610.21809#pg1>.

Tom: So the implication is that we can't just rely on in-domain performance as a proxy for how well a model will perform across different patient populations, which is something they really push for two <ref:2610.21809#pg1>.

Conclusion: Jane: So to wrap up this look at "From Adults to Children: Benchmarking Foundation Models for Chest X-Ray Classification under Domain Shift," the main message is that we see a clear trend across all tested models.

Lu: All models suffer a substantial performance degradation when transferred from adult to pediatric data, with drops exceeding ten percentage points two <ref:2610.21809#pg1,All models suffer a substantial>.

Meng: The paper confirms that cross-population robustness depends heavily on both the pretrained representation and how it's adapted for the downstream task two <ref:2610.21809#pg1>.

Lalam: They’ve shown that some models, like RAD-DINO or MedCLIP and BiomedCLIP, retain comparatively strong pediatric performance with limited adaptation because they are strong in their own domain two <ref:2610.21809#pg1>.

Tom: So to put it plainly, this study is the first systematic benchmark of multiple foundation models under a common adult-to-pediatric CXR evaluation protocol two <ref:2610.21809#pg1,multiple foundation models under a common adult-to-pediatric>. It really forces us to evaluate them directly across patient populations instead of just checking their adult performance two <ref:2610.21809#pg1>.

Jane: Exactly. This work provides a standardized way to see where these foundation models stand when they encounter real demographic shifts, and it points toward needing more careful consideration of how we adapt them for specific medical applications two <ref:2610.21809#pg1>.

Sara Inacio, Hugo Proenca, Joao C. Neves

University of Beira Interior, Portugal

cs.CV

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: Accepted at WACV 2027

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: The gist This work investigates foundation models’ ability to generalize from adult to pediatric chest X-ray classification, revealing that all models suffer substantial performance degradation

Key concepts

Domain Shift: This refers to a change in the statistical distribution of data between two different environments. In this study, it specifically means comparing chest X-ray images from adults (the source domain) with those from children (the target domain). The models are tested on data they haven't seen before, which causes their performance to drop.
Adaptation Strategies: These are the different methods used to modify a pre-trained foundation model after it has been trained on general data. The study compared three main approaches: linear probing (minimal modification), partial fine-tuning (some layers modified), and full fine-tuning (all layers modified). The goal is to see which modification helps the model perform best on the new pediatric task.
Representation Stability: This concept examines how well a model's learned internal features remain consistent when trained on different types of data or when subjected to changes in training. The analysis found that some models, like RAD-DINO and MedCLIP, are relatively stable; they maintain strong performance even with limited adaptation, suggesting their initial representations are robust to the domain shift.
AUROC Difference (∆AUROC): This is a specific metric used to quantify the severity of the domain shift. It measures the difference in Area Under the Receiver Operating Characteristic curve between performance on adult data and performance on pediatric data. A larger positive value indicates a greater drop in accuracy when shifting from adults to children.

Terminology

Benchmarking — 1× in this paper · explained in 4 episodes
Domain Shift — 1× in this paper · explained in 7 episodes
Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes

Summary

The gist This work investigates foundation models’ ability to generalize from adult to pediatric chest X-ray classification, revealing that all models suffer substantial performance degradation when transferred from adult to pediatric data, with drops exceeding 10 percentage points.

Problem Formulation and Datasets

The study considers multi-label chest X-ray classification under an adult-to-pediatric domain shift where samples from the two domains follow distributions PS(X, Y) and PT (X, Y), where PS(X, Y)!= PT (X, Y) The datasets used are VinDr-CXR [16] for the adult population and PediCXR [21] for the pediatric population Before constructing the benchmark splits, age-based population screening was performed using available DICOM metadata The labels were harmonized into a common five-label target space consisting of No Finding, Other Diseases, Lung Tumor, Pneumonia, and Tuberculosis

Models and Adaptation Strategies

The benchmark spans two main dimensions: the learning paradigm used during pretraining (ranging from supervised learning to image-only self-supervision and imagetext contrastive learning) and the domain of the pretraining data (ranging from natural images to biomedical and radiological data) Conventional CNN baselines include DenseNet-121 [9] and EfficientNet-B0 [25] Foundation models included DINOv2 [17] and CLIP [22] as general-purpose models, BiomedCLIP [35], MedCLIP [29], and BioViL-T [1] for medical vision-language models, and RAD-DINO [20] for CXR-specific image-only pretraining Adaptation protocols considered were linear probing (LP), partial fine-tuning (PFT), and full fine-tuning (FFT)

Evaluation Metrics and Results

Performance was evaluated using Macro-AUROC, mean Average Precision (mAP), and Macro-F1 computed over the five target labels The primary metric for domain-shift analysis is ∆AUROC = AUROCA→A − AUROCA→P, where lower values indicate a smaller performance degradation under the domain shift Results show that all models degrade under the domain shift, with drops exceeding 10 percentage points For instance, Macro-AUROC values reach approximately 92% on the adult test set but decrease to around 78% on the pediatric domain

Analysis of Robustness and Adaptation

The analysis revealed that increasing the depth of adaptation does not significantly improve cross-population generalization The choice of adaptation strategy depends strongly on the pretrained representation, as No single adaptation strategy is optimal for all models: additional fine-tuning can improve transfer, have little effect, or reduce cross-domain performance depending on the pretrained representation For example, DenseNet-121 [9] and BioViL-T [1] benefit from stronger adaptation, while MedCLIP [29] and BiomedCLIP [35], by contrast, achieve their best pediatric performance with linear probing and degrade with further fine-tuning Finally, strong in-domain performance does not necessarily imply robustness to the adult-topediatric shift

Conclusion

This is the first work to compare several foundation models under the same adult-to-pediatric setting while also varying the level of model adaptation The findings suggest that cross-population robustness depends on both the pretrained representation and how it is adapted to the downstream task This reinforces the need to evaluate foundation models directly across patient populations, rather than using in-domain performance as a proxy for cross-domain robustness.

Improvements for AI systems

The improved system can be evaluated using a common adult-to-pediatric evaluation setting while also analyzing the effect of adaptation depth, providing a standardized benchmark for assessing foundation models' robustness across patient populations.
The system can utilize a standardized evaluation protocol that compares conventional CNNs and foundation models under three distinct adaptation strategies: linear probing, partial fine-tuning, and full fine-tuning to determine how downstream modification affects cross-domain performance.
The improved system will be able to quantify the domain shift effect by reporting ∆AUROC = AUROCA→A − AUROCA→P, allowing researchers to assess how much modification of the pretrained representations is required for the downstream task.
The system can identify model-specific adaptation needs, as it reveals that No single adaptation strategy is optimal for all models: additional fine-tuning can improve transfer, have little effect, or reduce cross-domain performance depending on the pretrained representation.
The improved system will provide insights into representation stability by showing that certain models preserve performance relatively well under limited adaptation, such as RAD-DINO [20], MedCLIP [29], and BiomedCLIP [35] retain comparatively strong pediatric performance with limited adaptation.

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.