Keras FCN for polyp segmentation

This project is the implementation of the paper Rethinking the transfer learning for FCN based polyp segmentation in colonoscopy

Installation

conda create -n tfv1 python==3.8.13
conda activate tfv1
pip install nvidia-pyindex
pip install nvidia-tensorflow[horovod] keras==2.1.6
pip install opencv-python sklearn pillow imageio scikit-image matplotlib

Usage

Preparation

Download the dataset:

CVC-EndoSceneStill(http://pages.cvc.uab.es/CVC-Colon/index.php/databases/cvc-endoscenestill/)

kvasir-seg.zip

Extract the dataset compressed file to DATA_SOURCE

Define the environment path in CVC2Keras.py and Kvasir2Keras.py

• DATA_SOURCE: Directory for decompressed dataset

• DATA_PATH: Directory for converted data

• SAVE_ROOT: Model saving directory

Convert the dataset

python segmentation/dataset/CVC2Keras.py
python segmentation/dataset/Kvasir2Keras.py
python classification/PatchGenerator.py

Training the network

./train.sh

Evaluate the network

Select the best model by tensorboard in evaluation dataset: cd to the directory of SAVE_ROOT

tensorboard --logdir=segmentation/ --port=6006
tensorboard --logdir=classification/ --port=6006

Update the path of weights in evaluate.py

cla_weights_PATH: Path of the best classification network weights

fcn_weights_PATH: Path of the best FCN for segmentation network weights