Loading presentation...

Present Remotely

Send the link below via email or IM


Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.


Building Cancer Detection Pipelines Using Deep-Learning

No description

Rahul Remanan

on 11 January 2019

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of Building Cancer Detection Pipelines Using Deep-Learning

How to feed the data?
What model to use?
Who am I?
Doctor, Scientist, Entrepreneur.
Trained as a medical doctor.
Over a decade of healthcare scientific computing experience.
Why Artificial Intelligence and Healthcare?
Doctor's best friend is evidence, evidence's best friend is data, data's best friend is artificial intelligence.
Why open-source?
Building Cancer Detection Pipelines Using

What is Jomiraki?
Jomiraki logo represents the electron cloud around a helium atom.
Powered by Jomiraki
Easiest AI in the cloud.
Why develop AI with Jomiraki?
Dr. Rahul Remanan

CEO & Chief Imagination Officer,
Moad Computer

Passionate about neuroscience, open-data, open-science and open-source
Currently runs a portfolio of companies including Moad Computer.
Core mission: "Technology for the progress of human kind."
Solve and Scale in healthcare achievable only with AI.
Why build scalable pipelines in healthcare?
Delivering personalized/precision healthcare is a tough problem.
In need of innovative ideas, including the use of data and artificial-intelligence as the backbone of care.
Over 10% of individuals with a diagnosis of breast cancer dies within the next 5 years.
How to reduce that number to 0% needs technology, investment and scalability.
Saves time, cost and effort to provide care to a single individual.
More time for healthcare providers to deliver better quality care.
What is in a logo?
What is in a name?
What is a
deep-learning pipeline?

Happy users
Training data:
From a consortium of universities in the European Union.
Microscopy images of breast tissue.
Sample size = 400.
Four classes: benign, carcinoma in-situ, invasive carcinoma and normal.
Normalized data (N=100) x 4.
Image augmentation: 16x increase in sample size (N=6400).
Train-Validation split of 0.875:0.125 (5600:800).
Using Keras image augmentation API, each epoch of the training data, received an unique input image.
Inception v3.
Transfer learning: Extract high level machine vision features.
Fully connected layer: Solving domain specific problem.
Minimize over-fitting: Interposing dropout and batch normalization layers.
What is next?
How to generate predictions?
The model tested using a smaller, unexposed data-set (N=32).

Previous top performing model using this test data: ~75% sensitivity.

Performance of our end-to-end deep-learning model: 94% sensitivity.
What is unique?
Inception v3 is inspired the visual cortex.
Base layers are unaltered, but the top layer is a custom architecture using multi-branch perceptron.
The interposing dropout and batch normalization layers in the top layer to prevent overfitting during training.
What are microservices?
Designing REST-ful API endpoints for platform agnostic planetary scale deployment.
Three access points; in-line with "Browser is the computer" design philosophy:
Browser through a Graphic User Interface
Browser through the address bar.
Command line interface
Where to go from here?
More data, training and fine tuning.
Generating predictions.
Continuous development.
"Open source is eating the world."
"Open architectures, whether it is open science, open data or open source is empowering to both individuals and organizations."
"Open architectures are potent forces of good, just like democracy."
Just one machine (Jom) + Iroquois native American tribe.
A.K.A: Just one marvelous knowledge intelligence platform.
Also happens to include first phonetics of founders: Rahul & Meetu.
Data, development & deployment in one platform.
Linear code development and scaling.
Zero user dependencies.
'Browser is the computer' philosophy for the end-user.
Fully open-source stack.
Python, C++, C, R, Javascript, Solidity ...
Secure two-factor authentication.
Optimized back-end infrastructure.
Real world testing of performance.
Build authentication and security.
Trusted deployment with blockchain.
Planetary scale deployment.
Horovod scalability
32 servers with 4 Pascal GPUs each connected by RoCE-capable 25 Gbit/s network
(Source: Uber, Via GitHub).
Ductal Carcinoma Invasive
Returns a JSON object with predictions
Jupyter notebook.
Browser address-bar API.
Base layers of Inception v3, with input layer in red.
Multi-branched perceptron like top layer.
It is a symbolism for the probabilistic computing era, the hallmark of machine learning and artificial intelligence.
Thank You!
Lightweight responsive front-end
Returns JSON object with predictions
Why is it useful?
Helps integrate with an existing EMR.
Prioritization workflow, selecting the most important slides to review first, only possible using AI pipelines.
Minimize errors and missed diagnosis.
Full transcript