Melanoma is the most lethal type of skin cancer. Early diagnosis is crucial to increase the survival rate of those patients due to the possibility of metastasis. Automated skin lesion analysis can play an essential role by reaching people that do not have access to a specialist. However, since deep learning became the state-of-the-art for skin lesion analysis, data became a decisive factor in pushing the solutions further. The core objective of this M.Sc. dissertation is to tackle the problems that arise by having limited datasets. In the first part, we use generative adversarial networks to generate synthetic data to augment our classification model’s training datasets to boost performance. Our method generates high-resolution clinically-meaningful skin lesion images, that when compound our classification model’s training dataset, consistently improved the performance in different scenarios, for distinct datasets. We also investigate how our classification models perceived the synthetic samples and how they can aid the model’s generalization. Finally, we investigate a problem that usually arises by having few, relatively small datasets that are thoroughly re-used in the literature: bias. For this, we designed experiments to study how our models’ use data, verifying how it exploits correct (based on medical algorithms), and spurious (based on artifacts introduced during image acquisition) correlations. Disturbingly, even in the absence of any clinical information regarding the lesion being diagnosed, our classification models presented much better performance than chance (even competing with specialists benchmarks), highly suggesting inflated performances.
Improving Skin Lesion Analysis with Generative Adversarial Networks
