Overview / Usage

Pneumonia is one of the killer diseases around the world, the fast and the best way to diagnose it is MRI images, so there has been a lot of interesting to develop deep learning models to detect if there is pneumonia in the MRI picture or not using deep convolution neural network models. As known, convolution neural network requires a lot amount of data, and it’s not flexible to input affine transformation like rotation. A new deep learning architecture called “Capsule Network” was created in November 2017, by the Godfather of deep learning Geoffrey E. Hinton. It’s a flexible architecture with data input transformation and request less data for the training then the convolution neural network, and that solves the problem of the missing data in medicine.

This work was experienced on NIH Xray Dataset.

Methodology / Approach

In this work, I used Python as a programming language, Pillow for pictures processing, Tensorflow and Keras as deep learning libraries. Also I used **multiprocessing **to get faster results in pictures processing (actually it speed up the process 15 times).

My methodology was to clean the data and balance it (It was really very unbalanced), so I used data augmentation for that. For hyper-parameters choices, I made a huge search space and used Bayesian Optimization (used Skopt Library). Also I trained the deep learning model on multi gpus, so for that I used Distributed Training using Tensorflow library.

Technologies Used

Python

Tensorflow

Keras

Multiprocessing

2x GTX 1060

Intel core i7 7th generation