| Resumo: | Transfer learning is a popular solution to the common problem in deep learning that is the lack of data or the computational resources to train large models from scratch, which skin lesion classification is a prime candidate for because high quality medical imaging data in this domain is scarce. This dissertation studies transfer learning in the domain of skin lesion classification by exploring pre-trained models of the VGG16 architecture (originally trained on ImageNet) and repurposing them for skin lesion classification on the ISIC 2018 dataset. Specifically, models of VGG16 are tested by exhaustively testing the layers at which weights are extracted from and up to which they are frozen from further training, concluding that extracting all layers from VGG16 and fine-tuning the last two convolutional blocks to the ISIC 2018 dataset is the most performant configuration. However different choices of optimizer and learning rates could unveil better models. For comparison, two custom CNN architectures are explored and trained from scratch in a typical endto- end learning scheme, from which it can be seen that end-to-end learning of CNN is much harder due to the many different hyperparameters that need to be cross-validated on a wide range of values which is computationally intensive to do thoroughly. In conclusion, transfer learning is a much more practical strategy for skin lesion classification and most other computer vision problems.
|
|---|