From Shortcut Learning to Explainable Prediction: A Framework for Complex Pulmonary Disease Detection

Medical imaging plays a critical role in diagnosing and monitoring various diseases, with chest radiographs (CXRs) being one of the most widely used tools for pulmonary disease detection. However, the interpretation of CXRs is often challenging due to overlapping tissue features, low contrast, and the presence of co-occurring diseases. Traditional deep learning approaches, which often focus on single-disease classification, fail to account for the complexities of multi-pathology presentations and raise concerns about bias and interpretability. This thesis addresses these limitations by advancing explainable, multi-label deep learning frameworks tailored for the detection and explanation of co-occurring pulmonary diseases in CXRs.
I highlight the risks of single-disease approaches, using COVID-19 detection as
a case study, and demonstrate the benefits of multi-label classification in capturing
disease interdependencies and mitigating model bias. To improve interpretability, I propose sparse prior variational autoencoder (VAE) and hierarchical VAE models, which provide precise visual explanations through gradient-guided latent traversals. These methods outperform traditional deep CNN-based explainability techniques in feature isolation and disease localisation but face challenges with reconstruction quality and predictive accuracy. By advancing explainable, multi-label frameworks, this thesis advances the development of trustworthy, transparent diagnostic tools.