Deep Learning for Neuroimaging

Team: Andrew Doyle, Joseph Paul Cohen, Thomas Funck, Christopher Beckham

Date: December 11th, 9h-17h. Breakfast/registration at 8h30.

Location: Amphithéâtre “le groupe Maurice”, CRIUGM

Summary: Deep learning is one of the most promising avenues towards achieving artificial general intelligence, and a strong new tool for the analysis of neuroimaging data. This course will offer an introduction into the theory behind how representations are automatically learned from data, and offer students an introduction into how to use the Keras library to formulate and solve a variety of deep learning problems using hands-on examples.

Learning Objectives:

• Understand how representations are learned in deep neural networks
• Implement a convolutional neural network in Keras on neuroimaging data
• Learn how embeddings can be learned

Schedule:

Morning (9h-12h30): Introduction & Segmentation with Deep Learning

9:00 am – 10:00 am: Introduction to Deep Learning for Neuroimaging (Andrew Doyle)

10:00 am – 11:00 am: Deep Learning in Keras – Hands-on Defacing Detector (Andrew Doyle)

11:00 am – 11:15 am: Break

11:15 am – 12:30 am: Deep Learning for Segmentation - with hands-on U-Net (Thomas Funck)

12:30 pm - 1:30 pm: Lunch

Afternoon (13h30-17h00): Getting Deeper

1:30 pm – 2:45 pm: Looking Inside the Black Box - with Interpretability Hands-on (Andrew Doyle)

2:45 pm – 4:00 pm: Clinical data successes using machine learning - with Word2vec hands-on (Joseph Paul Cohen)

4:00 pm – 5:00 pm: Generative Adversarial Networks - with hands-on GAN (Christopher Beckham)

Requirements

• Basic familiarity with programming in Python is an asset, but not a requirement.
• Examples will be presented in Google Collaboratory, and participants should create a Google account to run & write code along with the instructors: https://colab.research.google.com.
• For students who wish to continue their analyses after the course, Python 3.x should be installed (ideally through Anaconda) with the Keras package.