(Wednesday, 8/1/2018) 1:45 PM - 3:45 PM
Room: Davidson Ballroom B
Convolutional neural networks are the most prevalent deep learning architecture class used in medical imaging today. The first part of this session will build on the previously acquired knowledge of neural networks as we transition to working with image data and integrate convolutional layers. We will be covering convolution-specific concepts including receptive fields, weight sharing, strides, padding and pooling. As deep learning is often regarded as "black-box" method, we will showcase attempts at understanding what the network deems important through saliency heatmaps. Finally, we will outline other deep learning architectures such as recurrent neural networks and generative adversarial networks.
The second part of this session will move from the theoretical realm towards introducing more concrete case studies of deep learning applications in medical imaging. These will cover multiple tasks including the detection, segmentation and characterization of regions of interest, as well as span different imaging modalities including CT, MR and Ultrasound. Emphasis will be placed on the choice of deep learning architecture used in each application as well as the image pre- and post-processing pipelines.
Learning Objectives:
1. To outline the general mechanics of convolutional neural networks
2. To introduce methods for interpreting deep learning outcomes
3. To introduce examples from recent literature regarding deep learning applications in medical imaging
4. To tie theoretical deep learning concepts with practical medically-relevant use cases
Handouts
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
