mingcheng991129/deep-nsclc

This repository contains the source code of FCCM for NSCLC.

2D CT Image Based Non-small Cell Lung Cancer (NSCLC) Classification Using Convolutional Neural Networks

This repository contains the source code of paper 2D CT Image Based Non-small Cell Lung Cancer (NSCLC) Classification Using Convolutional Neural Networks.

About this repository

• train folder:

  • models folder:
    • Baseline model
    • Image-wise model
    • Patch-wise model
    • ResNet model
    • CatNet model (proposed)
  • the dataset(csv format): Lung1.clinical.csv
  • the code for training and testing the model: generate_val_set.py, sequence_folders.py, train.py, utils.py
  • list of samples for testing: val_list.txt

• preprocess the dataset: prepare_training_data.py

Notice that the entire dataset (2D CT images) can be downloaded here.

Prerequisites

torchvision
scipy
argparse
tensorboardX
blessings
progressbar2
path.py
matplotlib
opencv-python
scikit-image
pypng
tqdm
spatial-correlation-sampler
pandas
pytorch
numpy

Train in the Ubuntu environment, on a GeForce RTX 2080Ti GPU.

Prepare data for training

The CT files can be converted into jpg images by running the command:

python prepare_training_data.py  \
--dataset_dir  /root/to/the/raw/DCM/files  \
--dump-root /root/to/the/prepared/files

Generate the list of samples for testing

After running this command, a txt file named val_list.txt will be generated which is used for testing. The txt file that we used during our training process is also provided in this repository.

python3 generate_val_set.py  \
--dataset_dir  /root/to/the/prepared/jpg/files 

Training the Baseline model

python3 train.py \
--dataset_dir /root/to/the/prepared/files  \
--label_dir /root/to/the/label/csv \
--batch-size 4  --FCCMnet Baseline  \
--lr 1e-4    --epochs 100  --name Baseline

Training Image-Wise and Patch-Wise models

python3 train.py \
--dataset_dir /root/to/the/prepared/files  \
--label_dir /root/to/the/label/csv \
--batch-size 4  --FCCMnet PatchWise  \
--lr 1e-4    --epochs 100  --name PatchWiseNetwork

python3 train.py \
--dataset_dir /root/to/the/prepared/files  \
--label_dir /root/to/the/label/csv \
--batch-size 4  --FCCMnet ImageWise  \
--lr 1e-4    --epochs 100  --name ImageWiseNetwork

Training ResNet model

python3 train.py \
--dataset_dir /root/to/the/prepared/files  \
--label_dir /root/to/the/label/csv \
--batch-size 4  --FCCMnet ResNet_FCCM  \
--lr 1e-4    --epochs 100  --name ResNet18

Training CatNet model

python3 train.py \
--dataset_dir /root/to/the/prepared/files  \
--label_dir /root/to/the/label/csv \
--batch-size 4  --FCCMnet CatNet_FCCM  \
--lr 1e-4    --epochs 100  --name CatNet18

Visualization during training process

run the command under the checkpoints folder:

tensorboard --logdir=./

and visualize the training progress by opening https://localhost:6006 on your browser.