Medical image segmentation is complicated by a lack of training data and by similarity of intensities in regions of interest in image data. Additionally, current best methods for segmentation, deep convolutional neural networks, can drastically change prediction values with only small changes in input. To overcome this drawback, we combine partial differential equations with deep learning to create small convolutional neural networks while still being robust to perturbation.
Learning Objective: Learn how mathematical structure of differential equations enables efficient and robust deep learning medical image segmentation methods.
Jonas Actor (Presenter)
Beatrice Riviere, Rice University
David Fuentes, MD Anderson Cancer Center