Multi-scale Gradual Integration CNN for False Positive Reduction in Pulmonary Nodule Detection

This repository contains code to train and test MGI-CNN. (https://doi.org/10.1016/j.neunet.2019.03.003)

Hardware and Software

Software requirements

For data processing: SimpleITK, Scipy

pip install SimpleITK scipy

This code requires unzipped LUNA16 dataset. (https://luna16.grand-challenge.org/Download/)

For training: Ubuntu 16.04, Python 3.6, Tensorflow 1.10

(Optional) GPUtil

pip install GPUtil

Hardware and training duration

Each fold takes about 12 hours to run 100 epochs using Nvidia GTX 1080 ti. Note that all experiments in our paper are based on 40^th epoch.

Usage

For training:

python main.py --data_path=PATH --summ_path_root=PATH --fold=0 --maxfold=5 --multistream_mode=0 --model_mode=0 --train

For testing:

python main.py --data_path=PATH --summ_path_root=PATH --fold=0 --maxfold=5 --multistream_mode=0 --model_mode=0 --test --tst_model_path=PATH --tst_epoch=40

• Specify your data path (--data_path) and path to save your results and summary (--summ_path_root). Unzipped LUNA16 dataset should be inside "(--data_path)/raw/" folder.

  Example
  --data_path=/home/jsyoon/MGICNN/dataset/
  /home/jsyoon/MGICNN/dataset/raw/1.3.6.1.4.1.14519.5.2.1.6279.6001.100225287222365663678666836860.mhd
  /home/jsyoon/MGICNN/dataset/raw/1.3.6.1.4.1.14519.5.2.1.6279.6001.100225287222365663678666836860.raw
  ...
  /home/jsyoon/MGICNN/dataset/raw/candidates_V2.csv

• Specify fold to train (--fold) and maximum number of folds (--maxfold).
• Specify which multistream mode to use (--multistream_mode). (0-element(proposed), 1- concat, 2-1x1 comv)
• Specify which model to use (--model_mode). (0-proposed, 1-RI , 2-LR, 3-ZI, 4- ZO)
• Specify train or test (--train or --test and --tst_model_path/--tst_epoch).

Results

We participated in the competition and got the following CPMs:

• MILAB_ConcatCAD: rank 3 (2017.11.25)

https://luna16.grand-challenge.org/Results/

Authors

Bum-Chae Kim, Jee Seok Yoon**, Jun-Sik Choi, and Prof. Heung-Il Suk*

Corresponding author: hisuk@korea.ac.kr
*\ For code inquiries: wltjr1007@korea.ac.kr

Machine Intelligence Lab.,\ Dept. Brain & Cognitive Engineering.\ Korea University, Seoul, South Korea.