Look What the Probes Dragged In! Real-World Chest X-ray Shortcuts in MedCLIP

2026-08-12

AI Summary

This study reveals that medical vision-language models such as MedCLIP rely on spurious shortcuts—such as the presence of drainage tubes or scanner artifacts—in chest X-ray images for prediction, compromising generalization and reliability. By inserting linear probes into intermediate layers of a frozen ResNet-50 backbone and analyzing predictions for pneumothorax and cardiomegaly across the NIH-CXR14 and PadChest datasets, the work systematically characterizes, for the first time, the distinct evolution patterns of localized shortcuts (in deeper layers) versus diffuse shortcuts (in shallower layers). This analysis integrates subgroup calibration, layer-wise confidence curves, and manual image inspection, uncovering significant label biases and data quality issues in public benchmarks. Despite high probe AUROC scores, the models exhibit markedly poor calibration, highlighting the unreliability of their decision-making processes.

Abstract

Vision-language models, such as contrastive language-image pre-training (CLIP)-based approaches, have reached state-of-the-art (SOTA) results in medical artificial intelligence. However, recent work reveals that CLIP-based models remain vulnerable to shortcuts. We investigate how real-world shortcuts manifest across different layers of the medical CLIP-based model, MedCLIP, and its vision encoder, a frozen ResNet-50. We attach 17 linear classification probes to the intermediate layers of the ResNet-50 and train them on three different dataset configurations and targets: NIH-CXR14 (pneumothorax) and PadChest (cardiomegaly and pneumothorax). This setup allows us to observe model behaviour during evaluation using subgroup-based calibration and layer-wise confidence curves. We find that the final linear probes achieve a high AUROC but poor calibration in the models. The layer-wise confidence analyses suggest that shortcuts emerge at different depths. Patterns consistent with localised shortcuts, such as drains, appear at later layers, while patterns consistent with diffuse shortcuts, such as scanner-specific noise patterns, emerge earlier, aligning with previous work. Finally, we conduct a manual analysis of the images, which reveals data quality issues in both NIH-CXR14 and PadChest. Our findings underscore that even SOTA models remain vulnerable to shortcuts, and the need for high-quality and well-annotated datasets to draw solid conclusions. Code can be found on our GitHub: https://github.com/nikodice4/MedCLIP_shortcuts.

Problem

• shortcuts
• medical AI
• chest X-ray
• dataset bias
• model calibration

Innovation

• shortcut learning
• MedCLIP
• probing analysis
• medical vision-language models
• dataset bias