Abstract: Deep learning has great potential to improve automated medical image analysis. My talk will focus on two key components of successful predictive models working with medical images – the image data and the model architectures. Given the scarcity of medical data, a common approach to classifying medical images is to finetune deep learning models pre-trained on a large amount of data from other image domains. During the tutorial, attendees will learn how to finetune a state-of-the-art deep learning model for classifying a custom image dataset in Python with Keras library. I will lay out how understanding the specifics and limitations of the data is essential prior to modeling. In particular, I will discuss how data augmentation techniques are important for building robust predictive models and how they vary based on the data we are working with. The attendees will see how to implement image augmentation and to apply it to training classification models in Keras. Furthermore, I will discuss the problem of unbalanced datasets and propose solutions for it.
PhD Candidate at the University of Chicago