Enhancement and Segmentation GAN (ESGAN) We present a novel architecture based on conditional generative adversarial networks (cGANs) to improve the lesion contrast for the pixel-wise segmentation. ESGAN effectively incorporates the classifier loss into the adversarial one during training to predict the central labels of the sliding input patches.
You can find detailed results (Team name: Hamghalam) on BraTS 2013 dataset on:
- Challenge Dataset -
https://www.smir.ch/BRATS/Start2013
📒 Yellow : Edema
🔵 Blue : Enhancing tumor
📗 Green : Non-Enhancing tumor
A CUDA compatable GPU with memory not less than 12GB is recommended for training. For testing only, a smaller GPU should be suitable.
Linux or OSX
NVIDIA GPU + CUDA CuDNN
Keras
SimpleITK
TensorFlow
Put your Dataset as numpy array as:
data shape (#samples, width, lenght, 1)
X_full -------> High contrast images based on FLAIR
X_sketch -------> Original image FALIR
Target_class -------> Segmentation labels (Ground trusth)
python main.py 16 16 --backend tensorflow --nb_epoch 100 --do_plot --generator deconv --n_batch_per_epoch 400
positional arguments:
--patch_size Patch size for D (here 16x16)
--backend BACKEND theano or tensorflow
--generator GENERATOR upsampling or deconv
--n_batch_per_epoch N_BATCH_PER_EPOCH Number of training epochs
--nb_epoch NB_EPOCH Number of batches per epoch
--do_plot Debugging plot
BraTS 2013 dataset. Data can be downloaded from http://braintumorsegmentation.org/