End-to-End Supervised Lung Lobe Segmentation
In this project we present a fully automatic and supervised approach to the problem of the segmentation of the pulmonary lobes from a CT scan. A 3D fully convolutional neural network was used based on the V-Net wich we called Fully Regularized V-Net (FRV-Net). This work was performed in the Biomedical Imaging group at C-BER centre of INESC TEC, Portugal and it resulted in the paper "End-to-End Supervised Lung Lobe Segmentation" accepted to the IJCNN2018 conference. Here are the code and scripts to train our FRV-Net (as you select wich regularization techniques do you want) and to run the segmentations.
Running a single segmentation with a pre-trained model.
To run a single segmentation with a pre-trained model a example file called "run_single_segmentation.py" is available. It teaches you how to open a CT scan, to open the model and to predict and save the segmentation.
Train a model
If you want to train your model, a file called "train.py" is available. It allows you to set the specific regularization techniques and parameters of the desired net.
-path : Model path (path)
-train: Train data (path)
-val : Validation data (path)
-lr : Set the learning rate (float)
-load : load a pre-trained model (boolean)
-aux : Multi-task learning (float - weight in the loss function)
-ds : Number of Deep Supervisers (int - nº of layers)
-bn : Set Batch normalization (boolean)
-dr : Set Dropout (boolean)
-fs : Number of initial of conv channels (int)
The train and validation datasets has to contain two folders A and B. where the folder A contains the CT scans and the B the correspondent ground-truth. In the script file "train_session.sh", the examples used for our results are presented.
Software
Our project was developed using Python (2.7) and Keras (2.0.4) framework that are required to use it.