University Archives

Convolutional Neural Networks (CNNs) have demonstrated remarkable performance in medical imaging tasks, including pneumonia detection from chest X-rays. However, these models are known to exploit spurious correlations (superficial patterns in training data that do not reflect true causal relationships) rather than learning clinically meaningful features. This phenomenon, known as shortcut learning, poses significant risks to the fairness, robustness, and trustworthiness of deployed medical AI systems. This study investigates shortcut vulnerability across five established CNN architectures: AlexNet, VGG-16, ResNet50, InceptionV3, and DenseNet121. Unlike prior work that relies on post-hoc analysis of biases naturally occurring in datasets, this research artificially injects spurious correlations into a pediatric chest X-ray dataset. Models are trained on both clean and corrupted datasets, and their susceptibility to shortcut learning is quantified by measuring accuracy degradation and test-time performance shifts when spurious cues are present or absent. To identify the internal representations driving shortcut behavior, CNN feature visualization techniques are applied to isolate feature maps that encode learned spurious patterns. The goal is to identify which architectures are most susceptible to shortcut learning under controlled corruption, locate the specific filters and feature maps responsible for encoding spurious features, and extract practical insights for building models that are less prone to shortcut reliance.