Mitigating class imbalance and promoting fairness in medical image classification with limited data using deep learning

Class imbalance, limited data, and demographic bias are significant challenges in medical image classification, particularly in skin lesion analysis, where heterogeneous populations and privacy constraints hinder the development of robust and fair models. This thesis proposes deep learning strategies that improve classification performance, robustness, and fairness in both centralised and federated settings. First, SIA-SMOTE is introduced as a Siamese network-based oversampling method for imbalanced image data. By exploiting feature similarity in high-dimensional spaces, it generates more representative synthetic samples than conventional SMOTE variants, leading to consistent performance gains across benchmark and medical imaging datasets. Second, a self-attention fusion strategy for Vision Transformers is proposed to enhance skin lesion classification under imbalanced conditions. By fusing complementary attention mechanisms at the final transformer stage while preserving pretrained representations, the method improves balanced accuracy and F1 performance on the ISIC 2017 dataset. Third, this thesis presents FedCIAL, a federated colour-invariant adversarial learning framework for mitigating demographic bias in decentralised skin lesion classification. Using Fitzpatrick skin type transformations and adversarial domain adaptation, FedCIAL improves both accuracy and fairness while preserving data privacy on the Fitzpatrick17k dataset. Finally, a fairness-aware federated learning framework combining domain adaptation and class imbalance mitigation is developed and evaluated on the Diverse Dermatology Images (DDI) dataset. Results show that jointly addressing imbalance and domain shift yields synergistic improvements in subgroup fairness and overall performance. Overall, this thesis advances oversampling, transformer-based representation learning, and fairness-aware federated learning for developing robust and equitable medical imaging systems.