Executive Summary

Key Figures

Source: Pedersen, Sydendal, Cheplygina & Sourget, arXiv:2608.12086

Oakden-Rayner and colleagues laid out shortcut learning in chest X-ray models clearly back in 2020. A DenseNet-121 trained on NIH-CXR14 scored an AUROC of 0.940 on pneumothorax images that contained a chest drain, and dropped to 0.770 on images without one. A drain is the tube inserted to treat a pneumothorax, so it tags along with the positive images. The model had learned the mark of treatment rather than the disease, and the pooled average shows none of that split.

MedCLIP, the model in this paper, comes a generation later. It is a vision-language model pretrained on CheXpert, MIMIC-CXR, COVID and RSNA pneumonia data, and the core of its design is loosening the requirement that each image be matched to its own report. Paired images and text are scarce in clinical practice, and strict pairing pushes semantically similar cases into the negative pile for no better reason than the absence of a match. MedCLIP replaces that pairing with semantic similarity drawn from the Unified Medical Language System, which pulls unpaired data into training and scales the corpus up considerably. The outcome is still familiar. Probe accuracy landed at 0.839 for NIH-CXR14 pneumothorax, 0.905 for PadChest cardiomegaly and 0.875 for PadChest pneumothorax, all comfortable numbers.

Calibration is where the comfort ends. All three configurations were overconfident about their own predictions, and PadChest pneumothorax was the worst of them. Calibration was poorest in the IDC scanner group, which also holds the fewest positive cases, so the metric itself is fragile when positives are thin. Earlier work found the same shape: six CLIP-family models performed better on images with a drain and clustered their predicted probabilities near 0.5, while AUROC stayed in the normal range.

Set accuracy, subgroup gaps and calibration side by side across the three configurations and it looks like this.

A single pooled number cannot tell you what a model was looking at when it got the answer right. So rather than read the finished prediction, the researchers opened up the stages where that prediction gets built, one layer at a time. They froze the ResNet-50 that serves as MedCLIP's vision encoder and attached one linear classifier to each of its 16 internal blocks plus the final average pooling layer. That makes 17 probes, each taking the pooled features and predicting the presence of the finding through a single linear layer. Since the backbone weights never move and only the probes train, a probe that predicts well means the representation at that depth already carries the information.

The training loss is the sum of the cross-entropy of all 17 probes. With the backbone frozen and no weights shared between probes, each probe receives gradients only from its own loss. One layer's result cannot disturb another layer's probe, which is what makes the depths readable separately. All three dataset configurations ran under the same settings: 100 epochs, batch size 32, learning rate 0.00001, with early stopping if validation loss failed to improve for 15 epochs.

What gets measured is confidence rather than accuracy. The distance of the predicted probability from 0.5, the point of maximum uncertainty, is computed per subgroup and plotted in order of depth. When one subgroup's curve alone jumps at a particular depth, the representation there is holding something that separates that subgroup. Each probe's curve carries a 95% confidence interval from 1,000 bootstrap resamples, and how wide that interval grows for small subgroups becomes a constraint on interpretation later on.

Three settings served as the stage. NIH-CXR14 pneumothorax draws on 112,120 images from 30,805 patients, with drain status marked only in the test set. PadChest cardiomegaly comes from 160,861 images of 67,625 patients collected at San Juan Hospital in Alicante, Spain, and was chosen for its high prevalence and balanced sex distribution. Pneumothorax was run separately on the same PadChest, where prevalence is only 0.4%.

What was removed during preprocessing also matters for reading the results. PadChest mixes five projection views, and the researchers discarded everything shot standing, keeping only PA, AP and AP-horizontal. Rows whose disease label read unsuitable for imaging, exclude, or unchanged were dropped as well. NIH-CXR14 used the split shipped with the dataset, while PadChest was divided 80/20 with a further 20% carved out of the training side for validation. All three configurations split on patient identifier so that images from the same patient never appear in both training and test.

The hypothesis came from earlier synthetic experiments. Boland and colleagues found in 2025 that diffuse shortcuts spread across the whole image, such as uniform noise, surface in early layers, while localized shortcuts such as a red square in a fixed position appear in later ones. A chest drain is a physical device inside the body and sits closer to the localized end, whereas image quality differences between scanner models cover the entire frame and sit closer to the diffuse end. If that holds, drains should show up late and scanner differences early.

In NIH-CXR14 pneumothorax, the positive and the negative images were each split again by whether a drain was present, which gives four curves. All four stay low and flat together through the first twelve blocks. They begin to separate at block 13, with confidence on pneumothorax-positive images climbing to between 0.35 and 0.41 while the negatives remain below 0.25. Between the two positive curves, the one with a drain stays higher all the way to the end. Patients with the same disease draw more confidence from the model when a tube is in place.

The vertical axis is confidence and the horizontal axis is the block where the probe is attached. Through the first half the four curves sit almost on top of one another, and the separation is crowded into the right-hand end.

PadChest pneumothorax produced the opposite end. All four curves jump once at block 3 and then settle onto their own trajectories. The IDC negative curve without a drain climbs as high as 0.42 along the way, though it ends near the others. The direction matches the prediction that diffuse shortcuts appear in early layers, and it also lines up with earlier work reporting that imaging equipment becomes a shortcut under imbalanced conditions. Here the researchers step back, however. Images taken on IDC machines carry a burned-in R marking the right side, which they cannot rule out as a localized shortcut. With only four positive IDC samples, the curve itself is unstable.

PadChest cardiomegaly produced no pattern worth calling a shortcut. Small spikes do appear between blocks 4 and 6 and again at block 13. But confidence sits low across the whole range, between 0.23 and 0.28, and the four scanner curves nearly overlap within that same band, so the researchers did not read those bumps as traces of a shortcut. Splitting by sex showed no significant difference either. That the same method finds a signal in one dataset and nothing in another is itself the point: a shortcut is a product of the conditions the data was collected under, not a fixed property of the model.

NIH-CXR14's labels are the presence or absence of 14 thoracic conditions, pulled from radiology reports by natural language processing. What treatment the patient received and what machine took the image are not in there. The drain labels for this experiment therefore had to come from somewhere else. The NEATX project manually annotated 3,709 images showing a drain, and the no-drain side was filled in automatically by a separate classifier trained on those annotations. An experiment checking for shortcuts rests on another model's judgment.

Rather than take those labels on faith, the researchers checked them by hand. Of five images marked as having no drain, three showed a small drain tucked into a corner. The three cases printed in the paper's figure all fall into this group, and all three are classified as pneumothorax negative, with the tube running across the upper-left or upper-right corner. When the tubes the classifier missed sit on the negative side, the baseline for any drain-versus-no-drain comparison is compromised. Every comparison blurs by however much that error amounts to, and the authors wrote as much when they asked readers to interpret the results cautiously.

The same audit surfaced other things. NIH-CXR14 contained records with impossible ages and an image filled entirely with grey. PadChest yielded contradictory sex entries, duplicated images and duplicated metadata, and one skull X-ray carrying a lung disease label. Both datasets are public assets that thousands of papers in medical AI have cited.

In the end the label decides what can be asked at all. The two datasets divide between what their labels record and what survives only in the pixels.

The nature of the problem shifts at this point. This study could establish that the model responds to chest drains only because another team spent years building drain labels. Anything without that supplementary annotation never becomes a question in the first place. Whether the cardiomegaly experiment found no shortcut, or whether the model was holding some other unrecorded context that the scanner split simply failed to catch, cannot be distinguished with the labels available today.

Data captured by sensors carries the conditions of the site along with it, and that happens everywhere. In manufacturing inspection, camera model, lighting and line number get entangled with defect rates. If one line produces more defects, the model can learn the color temperature of that line's lighting rather than the shape of the flaw. In autonomous driving, sensor configuration, weather and regional road standards travel alongside the driving labels. In either case, the moment only the outcome label gets stored, whatever plays the role of the chest drain stays inside the image.

The questions worth asking when designing or taking delivery of a dataset therefore sit outside the accuracy metrics. What does our labeling schema record next to the outcome label? Does it keep the device identifier, the capture conditions, the history of interventions applied to the subject? And if it does, can you tell which of those values a person verified and which another model filled in? The drain labels in this paper failed the third question, and the interpretive range of the whole experiment narrowed as a result.

The layer-wise probe is itself worth borrowing. Freezing the backbone and attaching linear classifiers costs nothing in retraining, and with subgroup labels in hand you can see what separates at which depth. What the researchers validated, though, is one model with a CNN architecture. Whether the same layer-wise rule holds for transformer-based models or on other datasets is still open.

Earlier articles along the same line include turning labeling clicks into audit evidence, a diffusion autoencoder study that found cardiac MRI phenotypes without labels, and the accuracy trap in scarce allocation, on conditions where accuracy widens gaps instead of closing them. The paper itself is at arXiv:2608.12086 and the code is in the GitHub repository.

Primary Source

• 1.Pedersen, N., Sydendal, R., Cheplygina, V., & Sourget, T. (2026). "Look What the Probes Dragged In! Real-World Chest X-ray Shortcuts in MedCLIP." arXiv:2608.12086. — The primary paper for this article. Attaches 17 linear probes to MedCLIP's vision encoder to show drain and scanner shortcuts emerging at different layer depths.

Prior Work on Shortcut Learning

• 2.Oakden-Rayner, L., Dunnmon, J., Carneiro, G., & Ré, C. (2020). "Hidden Stratification Causes Clinically Meaningful Failures in Machine Learning for Medical Imaging." Proceedings of the ACM Conference on Health, Inference, and Learning (CHIL), 151–159. — The classic case of shortcut learning: AUROC of 0.940 with a drain present versus 0.770 without.
• 3.Boland, C., Tsaftaris, S. A., & Dahdouh, S. (2025). "Preventing Shortcut Learning in Medical Image Analysis Through Intermediate Layer Knowledge Distillation from Specialist Teachers." Machine Learning for Biomedical Imaging, 3, 447–476. — Source of the original hypothesis, from synthetic-bias experiments, that diffuse shortcuts surface in early layers and localized shortcuts in later ones.
• 4.Geirhos, R., Jacobsen, J., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020). "Shortcut Learning in Deep Neural Networks." Nature Machine Intelligence, 2(11), 665–673. — The foundational paper that established the shortcut learning concept.
• 5.Jiménez-Sánchez, A., Juodelyte, D., Chamberlain, B., & Cheplygina, V. (2023). "Detecting Shortcuts in Medical Images - A Case Study in Chest X-rays." 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), 1–5. — Confirms the same drain-driven shortcut in CheXpert and NIH-CXR14.

Dataset and Model Background

• 6.Wang, Z., Wu, Z., Agarwal, D., & Sun, J. (2022). "MedCLIP: Contrastive Learning from Unpaired Medical Images and Text." Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 3876–3887. — The original paper for MedCLIP, the model this study analyzes.
• 7.Cheplygina, V., Damgaard, C., Eriksen, T. N., Juodelyte, D., & Jiménez-Sánchez, A. (2026). "Augmenting Chest X-ray Datasets with Non-Expert Annotations." Medical Image Understanding and Analysis (MIUA), 133–144. — The NEATX project, source of the manually annotated 3,709 drain-positive labels for NIH-CXR14.
• 8.Sourget, T., Restrepo, D., Hudelot, C., Ferrante, E., Christodoulidis, S., & Vakalopoulou, M. (2026). "Fairness and Robustness of CLIP-Based Models for Chest X-Rays." Fairness of AI in Medical Imaging, 11–21. — Prior observation that six CLIP-based models performed better on drain-present images while AUROC stayed unremarkable.
• 9.Jiménez-Sánchez, A., Avlona, N., de Boer, S., et al. (2025). "In the Picture: Medical Imaging Datasets, Artifacts, and Their Living Review." Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 511–531. — A living review of artifacts and labeling errors in public medical imaging datasets.
• 10.Rafferty, A., & Rajan, A. (2026). "Limitations of Public Chest Radiography Datasets for Artificial Intelligence: Label Quality, Domain Shift, Bias and Evaluation Challenges." arXiv:2509.15107. — A survey of label-quality limitations in public chest X-ray datasets.