Building a novel deep learning architecture

I want to work on ISIC skin cancer classification dataset
1. Here we want to use novel GAN model for data balancing add dataset is highly imbalanced.
2. Develop a novel U-Net based architecture for image segmentation.
3. Develop a novel deep residual and skin convoluted or state of art architecture to achieve atleast 95% accuracy on ISIC17,ISIC18,ISIC19 and ISIC20 dataset coding to be done using tensorflow and environment is preferably Google Colab.