Development and analysis of deep learning solutions for automated diagnosis of lung and heart disease through multi-label chest x-ray classification

The automated classification of thoracic pathologies using deep learning represents a promising application of Artificial Intelligence (AI) in medicine, particularly to reduce diagnostic workloads in high-intensity clinical environments. The frequent coexistence of multiple diseases within chest radiographs, combined with an inherent class imbalance in datasets such as NIH ChestX-ray14, complicates clinical decision-making and poses significant challenges for multi-label classification systems. Despite advances from established models such as CheXNet and Z-Net, limitations persist in optimising classification strategies and calibrating decision thresholds for individual pathological classes. These limitations are exacerbated by the highly imbalanced and long-tailed class distribution of the NIH ChestX-ray14 dataset, one of the most widely adopted benchmarks in the field of medical imaging. Despite challenges, AI models support clinicians by reducing reporting times, enhancing diagnostic accuracy in cases of multimorbidity, enabling rapid critical detection for triage, and promoting diagnostic standardisation through consistent outputs. This thesis presents a systematic approach for multi-label classification of 14 thoracic pathologies by implementing three distinct neural architectures based on modified DenseNet-121 backbones, which are adapted for multi-label output. The methodology involves a comparative analysis of different optimisers (e.g. SGD and AdamW) and loss functions (e.g. Weighted Binary Cross-Entropy, Focal Loss, and Asymmetric Loss) to effectively address class imbalance. Particular attention is given to pathology-specific threshold optimisation techniques that enhance the model's ability to distinguish between individual diseases. Performance evaluation employs comprehensive multi-label metrics, including the Area Under the Curve (AUC) and the Receiver-Operating Characteristic (ROC) curve analysis. AUC was prioritised as a key indicator of discriminative power due to its robustness to class imbalance and independence from threshold selection. Threshold tuning for each pathological class was based on maximising the F1-score, macro-F1, and Youden's J statistic to balance sensitivity and specificity. The results revealed varying effectiveness of the model under different conditions, suggesting different levels of learnability between classes. The influence of class imbalance and inter-class visual similarity underscores the complexity of the model evaluation process. This work aims to contribute to the development of clinical decision support systems for the automated diagnosis of thoracic pathologies by proposing a methodological framework focused on performance optimisation in imbalanced multi-label scenarios. While the proposed approach shows encouraging results, particularly in supporting workload management and promoting diagnostic consistency, its effectiveness varies across different pathological classes, highlighting both its potential and its current limitations. These findings suggest that deep learning techniques can play a valuable role in assisting radiologists; however, further research and refinement are needed to ensure more consistent and generalisable outcomes in real-world clinical settings.