Problem

Vision-language models (VLMs) have achieved state-of-the-art results in medical artificial intelligence, particularly in tasks like chest X-ray classification. However, a significant concern is their potential reliance on "shortcuts," where models learn to associate image artifacts (like chest drains) or imaging characteristics (like scanner-specific noise) with disease labels rather than genuine clinical features. Previous research has indicated that CLIP-based models, including MedCLIP, can exhibit these shortcut behaviors and often show poor calibration. This study aims to understand how these real-world shortcuts emerge and manifest across the different intermediate layers of MedCLIP's vision encoder, a frozen ResNet-50, to evaluate its robustness and calibration in a linear probing setting.

Method

To investigate shortcut learning, the researchers attached 17 linear classification probes to the intermediate building blocks of MedCLIP's frozen ResNet-50 vision encoder, with a final probe at the last average pooling layer. These probes were trained on three distinct dataset configurations: pneumothorax classification using the NIH-CXR14 dataset (augmented with drain annotations), and both cardiomegaly and pneumothorax classification using the PadChest dataset. The training process only updated the weights of the linear probes, keeping the ResNet-50 backbone frozen to leverage its pre-trained representations. Evaluation involved analyzing subgroup-based calibration curves, which plot mean predicted probabilities against the fraction of true positive cases, and layer-wise confidence curves, defined as the deviation from maximum uncertainty (0.5), to observe how classification confidence develops through the layers. A manual analysis of selected images was also conducted to identify potential data quality issues in the datasets.

Results

The study found that the final linear probes achieved high AUROC scores across all test configurations, with pneumothorax on NIH-CXR14 reaching 0.839, cardiomegaly on PadChest reaching 0.905, and pneumothorax on PadChest at 0.875. Despite these high performance metrics, all models exhibited poor calibration and overconfidence to varying degrees. The layer-wise confidence analysis suggested that different types of shortcuts emerge at different depths: patterns consistent with localized shortcuts, such as chest drains, appeared in later layers, while patterns resembling diffuse shortcuts (like scanner-specific noise or dataset inconsistencies) emerged earlier. However, for PadChest pneumothorax, the distinction between diffuse and localized shortcuts was less clear due to confounding factors like "R" markers in X-rays. Furthermore, a manual inspection of the NIH-CXR14 and PadChest datasets revealed significant data quality issues, including incorrect patient ages, conflicting patient sexes, duplicate images, and unreliable automated drain annotations, suggesting that conclusions drawn from these datasets must be interpreted cautiously.

Implications

The findings underscore that even state-of-the-art vision-language models like MedCLIP, despite their scale and strong performance metrics like high AUROC, remain highly vulnerable to shortcut learning. This vulnerability indicates that models may not be learning clinically relevant features, questioning their robustness and reliability in real-world medical applications. The study also highlights that synthetic shortcuts behave similarly to real-world shortcuts in medical datasets, reinforcing previous research. The pervasive poor calibration across all models further complicates their practical use, as clinicians rely on accurate probability estimates. Crucially, the extensive data quality issues identified in widely used public datasets like NIH-CXR14 and PadChest emphasize the paramount importance of high-quality, meticulously annotated data for training and evaluating reliable medical AI models. Without robust data, even advanced architectures and methods may yield misleading conclusions, necessitating a critical re-evaluation of dataset curation practices.