Uncertainty-Aware Skin Lesion Image Synthesis using Diffusion Transformer for Imbalanced Classification

Class imbalance is a significant challenge in skin lesion classification, as dermatological datasets often contain very few samples for rare lesion types, but too many samples for common ones. To solve this problem, existing diffusion-based augmentation approaches commonly rely on classification accuracy to determine the generation of class-wise synthetic samples. However, in imbalanced settings, accuracy may provide misleading optimization signals, as majority classes can dominate the learning process, despite substantial uncertainty in minority classes. To handle this limitation, we propose an uncertainty-aware image synthesis framework for imbalanced skin lesion classification using Diffusion Transformers with Representation Autoencoders (DiT-RAE). The proposed method introduces an entropy-guided adaptive sampling strategy based on the intuition that not all lesion classes contribute equally informative representations during training. The framework estimates class-wise uncertainty and dynamically allocates synthetic samples according to the information complexity and learning difficulty of each lesion class. We validate our framework through extensive experiments on the HAM10000 dataset. The proposed framework achieves 86.00\% Macro F1-score, 84.04\% Balanced Accuracy and 84.29\% MCC, improving baseline by 4.03\%, 2.54\% and 3.65\%, respectively.