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

Technical Summary: Real-World Chest X-ray Shortcuts in MedCLIP

Problem Statement
While Vision-Language Models (VLMs) like MedCLIP have achieved state-of-the-art (SOTA) performance in medical artificial intelligence, recent literature suggests they remain vulnerable to "shortcut learning"—relying on spurious correlations rather than clinically relevant features. Previous studies indicate that models often exploit artifacts such as chest drains in pneumothorax classification or scanner-specific noise. However, it remains unclear how these shortcuts manifest across the intermediate layers of modern VLMs and whether the models are well-calibrated. Furthermore, existing research often relies on synthetically biased data; there is a need to investigate these phenomena using real-world medical datasets.

Methodology
The authors investigate the internal behavior of MedCLIP, a VLM utilizing a frozen ResNet-50 vision encoder, by employing a linear probing approach.

• Architecture: 17 linear classification probes are attached to the intermediate building blocks of the frozen ResNet-50, plus one final probe at the average pooling layer. Only the probe weights are trained; the backbone remains frozen to preserve pre-trained representations.
• Datasets: The study utilizes three configurations on real-world data:
  1. NIH-CXR14: Pneumothorax classification, stratified by the presence of chest drains (using NEATX annotations and automated labels).
  2. PadChest: Cardiomegaly classification, stratified by X-ray scanner type (IDC vs. PMS).
  3. PadChest: Pneumothorax classification, stratified by scanner type and patient sex.
• Evaluation Metrics:
  • AUROC: To measure overall discriminative performance across subgroups.
  • Calibration Curves: To assess if predicted probabilities match true positive frequencies.
  • Layer-wise Confidence Curves: To track the development of model confidence (deviation from 0.5 uncertainty) across network depth, hypothesizing that localized shortcuts (e.g., drains) emerge in later layers while diffuse shortcuts (e.g., scanner noise) emerge earlier.
• Manual Analysis: A qualitative review of specific images was conducted to verify data quality and annotation accuracy.

Key Results

• Performance: The models achieved high AUROC scores across all configurations (e.g., 0.905 for Cardiomegaly on PadChest, 0.839 for Pneumothorax on NIH-CXR14). However, performance gaps were observed between subgroups (e.g., different scanner types).
• Calibration: All models exhibited poor calibration, often displaying overconfidence. The degree of miscalibration varied by dataset configuration and subgroup, with the PadChest pneumothorax task showing the most significant issues, likely due to the extremely low prevalence of positive cases (0.4%).
• Shortcut Emergence:
  • NIH-CXR14 (Drains): Confidence curves for positive cases (with drains) diverged from negative cases starting at building block 13, spiking to higher confidence levels. This aligns with the hypothesis that localized shortcuts emerge in deeper layers.
  • PadChest (Scanner Noise): For the pneumothorax task, spikes in confidence occurred in earlier layers (around block 3), resembling the pattern of diffuse shortcuts (e.g., scanner-specific noise) described in prior synthetic studies.
  • Cardiomegaly: Confidence curves remained relatively low and stable, with spikes not clearly resembling shortcut patterns.
• Data Quality Issues: Manual inspection revealed significant errors in the datasets, including:
  • NIH-CXR14: Incorrect "no drain" annotations for images containing visible drains, non-uniform grey images, and incorrect patient ages.
  • PadChest: Conflicting sex metadata, duplicate images, and the inclusion of non-chest X-rays (e.g., skull X-rays) labeled with lung diseases.

Key Contributions

1. Implementation of Linear Probes on Real-World Data: The authors applied linear classification probes to MedCLIP using real-world medical datasets rather than curated synthetic biased data.
2. Layer-wise Analysis of MedCLIP: The study reveals that MedCLIP's vision encoder exhibits poor calibration and learns shortcuts at varying depths depending on the shortcut type (localized vs. diffuse).
3. Data Quality Audit: A manual analysis highlighted critical data quality and annotation errors in widely used public datasets (NIH-CXR14 and PadChest), which directly impact model evaluation.

Significance and Claims
The paper concludes that even SOTA models like MedCLIP are not inherently robust to shortcuts, despite their scale and multi-modality. The findings suggest that:

• Shortcut learning in real-world medical data behaves similarly to synthetic shortcuts, manifesting at specific network depths.
• High AUROC scores do not guarantee robustness or clinical reliability, as models can be overconfident and miscalibrated.
• Data quality is paramount: The presence of label errors, metadata conflicts, and annotation inaccuracies in public datasets undermines the validity of conclusions drawn from them. The authors emphasize that solid conclusions require high-quality, well-annotated data, and that current evaluation practices may be insufficient to detect these vulnerabilities.