| Resumo: | Machine learning, more specifically, deep learning is a fast-growing field that is being used for multiple medical imaging related problems, such as the early detection of skin cancer. For a long time, automated diagnosis of skin lesions from clinical images through deep learning was considered to be out of reach. However, recent advancements in deep neural networks allowed it to achieve state-of-the-art results on classification challenges and they have the potential to change the landscape of dermatology care. Moreover, there is a growing need for classification systems to reduce fatality rates of skin cancer by providing support for both dermatologists in the decision-making process and for patients that do not have access to expert physicians. This dissertation presents a systematic approach towards an multi-class (8 classes) deep learning classifier of skin lesions for the ISIC 2019 benchmark challenge. It attempts to study major open research points related to the effectiveness of state-of-the-art deep learning methods. The results indicate that recent CNN architectures can have a significant impact on the overall performance of deep learning based classifiers. However, these pre-trained models should be carefully re-purposed towards skin lesion classification through transfer learning methods. This work provides insight into major ways of improving the generalization performance of deep learning models towards skin lesion diagnosis. In this context, the impact of different image processing augmentation methods for offline and online data augmentation is studied. On the one hand, class balancing through offline data augmentation can significantly improve the generalization performance of underrepresented classes. On the other hand, online data augmentation brings considerable improvements towards overfitting reduction. Furthermore, this work provides a practical analysis of ensembles in the context of skin lesion diagnosis. Both the single model and the ensemble-model approach outperform the current state-of-the-art for the ISIC 2019 challenge, with a balanced multi-class accuracy of 0.815 and 0.846, respectively. Finally, experiments were made with different ways to detect out of training distribution data. The results contribute towards the understanding of the effectiveness of out-of-distribution detection methods in the context of deep learning based skin lesion classifiers.
|
|---|