Exploring PyTorch for AI assistance in Medical Imaging

Abhishek Kumar (~vibrantabhi19)


13

Votes

Description:

Just like in our everyday lives, AI and robotics are increasingly a part of our healthcare ecosystem. Currently healthcare is broken; there’s shortage of doctors; poor quality of care. There is a dire need to provide assistance to the whole medical industry to improve healthcare.

PyTorch, which is a very popular modular deep learning framework for fast, flexible experimentation is an invaluable resource for such problems. It supports seamless conversion of Numpy arrays into GPU tensors and vice versa. The dynamic computational graph allows to change the network behavior on the fly unlike static graphs and due to Its highly modular nature helps in fast debugging.

Unlike other production grade tools, Pytorch helps with lots of Research and Experimentation with novel architectures and is very useful to test ideas a bit more quickly and prototyping.

With Medical Imaging being the field most impacted by AI, our goal in this workshop is to give a good head start covering the heuristics of Medical Imaging, the concepts involved in it and how to code your way out.


This workshop would be divided into two halfs.

  • First Half: Pytorch Introduction Duration: 1 hour The first half would be a gentle introduction to PyTorch framework. We will introduce the audience with the basics of PyTorch. This workshop will cover topics like:

  • What is PyTorch? (Use cases and war stories)

  • Tensor 101
  • Ndarray/Tensor library
  • Numpy Bridge,
  • Fast CPU to GPU conversion of tensors
  • The automatic differentiation engine or autograd
  • Difference between Static and Dynamic computational graphs
  • Advantages of dynamic computational graph with examples
  • The optimization package
  • Scope of debugging
  • Linear Code flow in Pytorch (One of the core philosophy of PyTorch)
  • Saving and loading models

A 5 minute coffee/kit-kat break. :-)

Second Half: Let’s dive in. Duration: 1 hour 25 minutes.

  • Introduction to Radiology: What is radiology? What do the images look like? How is AI used here? How will AI help improve radiology practice? (5 min)
  • Introduction to Convolutional Neural Networks with Hands on experience of coding Neural Networks and CNN using PyTorch. (20 min)
  • Introduction to classification networks. (20 min)
  • A 2D classification network Hands-On session for Liver Segmentation/Classification. (20 min)
  • Challenges faced in Medical Imaging and Deep Learning in general. (5 min)
  • End the session with talking about the bridge between literature and practical implementation. (5 min)

Putting it all together

A 10 minutes Q & A session.

Prerequisites:

Zeal to learn new things would be enough but basic knowledge of Python would be good to go but the following are always encouraged:

  • Basic Knowledge of algebra.
  • Python Libraries such as Numpy.
  • Basic knowledge of working with Neural Network (not a strict requirement as we will be covering most of it).
  • We also encourage the participants to have a look into the following linked talks/videos/literature to get a head start into the topic.

The related materials from web for ideas:

https://github.com/soumith/talks/blob/master/2017-NIPS/Coding-papers-in-pytorch.pdf

https://github.com/soumith/talks/blob/master/2017-GATech-Atlanta/PyTorch-frameworks_overview_deepdive.pdf

https://www.youtube.com/watch?v=LEkyvEZoDZg

https://www.youtube.com/watch?v=VMcRWYEKmhw

https://www.youtube.com/watch?v=Rv9naeLXolY&index=3&list=PLrzfRWNHZPa0gKBEXTJ0gbDu8NsR07KEH

https://github.com/yunjey/pytorch-tutorial/blob/master/tutorials/01-basics/pytorch_basics/main.py

Content URLs:

https://github.com/vibrantabhi19/PyConIndia2018 (A Github Link to the slides and the Jupyter Notebooks) https://docs.google.com/presentation/d/1UmT3PbazC6sO_owIeiLNj5G1EdTwrdpS84JWenO-3eE/edit?usp=sharing (Introduction Slide for CNN and PyTorch)

Some more slides and notebooks as and when we come up with more ideas to make the workshop interacting and interesting.

Speaker Info:

Abhishek Kumar: Deep Learning Engineer, Predible Health, Bangalore.

I am presently working as Deep Learning Scientist at Predible Health, here,I work on the the research and development of Predible's core Imaging platform wherein we have build state of the art segmentation algorithms/models in Computer Vision. I have previously taken workshop at IIT-Bombay Techfest, I have spoken at Shri Mata Vaishno Devi University at their SFD celebrations and at MuPy (Manipal Institute of Technology's annual Python Conference), Kongu University and a few other colleges/Universities. I have been a constant contributor in the open source world and have been attending PyCon and other conferences every year. An athlete, a Real Madrid F.C follower and a part time stand-up comedian (good enough to make you laugh).

Aditya Bagari: Final year Undergrad, Indian Institute of Technology, Madras

I am a final year Undergraduate student at IIT-Madras doing my Dual-Degree in Engineering Design with specialisation in Bio Medical Sciences. I have been working on Medical Imaging and PyTorch for almost a year and I have been a constant admirer of Open Source Technologies and frameworks.


Feel free to drop any suggestions or modifications that you want in this workshop.

See you at PyCon.

Speaker Links:

Abhishek Kumar: Website (A very outdated one), LinkedIn, Medium, Github.

Aditya Bagari: LinkedIn

Section: Data science
Type: Workshops
Target Audience: Beginner
Last Updated:

Looking forward for this.

SUNNY (~sunnykrGupta)

Login to add a new comment.