A multi classes image classifier, based on convolutional neural network (CNN) using Keras and Tensorflow.
A multi-label classifier (having one fully-connected layer at the end), with multi-classification (2 classes, in this instance)
Classifying 2 cancer types (LGG, GBM tiles generated from around 1200 WSI slides) downloaded from candygram.emory.edu
Thumbnail extraction and tile extraction from WSI slides are implemented using Girder Calls and HistomicsTK functions (https://girder.readthedocs.io/en/latest/)
Used Keras.ImageDataGenerator for Training/Validation data augmentation and the augmented images are flown from respective directory
train folder has training images and validation images which are autogenerated by keras has val images with cross-validation ratio of 80-20 (train-val)
test folder has all test images
Environment: A docker container having Keras, TensorFlow, Python-3 with GPU based execution
After !pip install nvidia-ml-py3 --user
Model in use is, Google Inception V3 with image size 150*150
Number of FCNs (Fully Connected Networks) - starting from 1024 to 100000, vary between models Initially only two cancer classes lgg and GBM and is extending to another 25 cancer classes like BRCA etc