Rotational-CoordConv

Duke BME590 (Medical Machine Learning) Project 2: Rotational Variation of CoordConv

Notebooks

'Basic_CoordConv_Classification_MNIST': Image classification comparing CNN and CNN+CoordConv performance.

'CoordConv_Baseline_Classification_Application_Pneumonia_CXR': Pipeline for image classification of pneumonia. (Note: not used in project due to long training time).

'Rotational CoordConv for Lung Image Segmentation': Includes CoordConv layer class and Learned Rotational CoordConv layer class. Both can be plugged in to any Keras model for immediate use.

Presentation Slides

https://docs.google.com/presentation/d/1M339Q8ohdX3mB3Lw4PwImxou3RTLJyAhfe0_d-G0FoE/edit?usp=sharing

References:

An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution
Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski Such, Eric Frank, Alex Sergeev, Jason Yosinski
https://arxiv.org/abs/1807.03247

Spatial Transformer Networks
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu
https://arxiv.org/abs/1506.02025

Lung Segmentation:
    Data: https://www.kaggle.com/kmader/finding-lungs-in-ct-data

    Model: https://www.kaggle.com/toregil/a-lung-u-net-in-keras/