minghaobao / lung_nodule_detector

lung nodule detector using 3d resnet using focal loss

lung_nodule_detector

lung nodule detector using 3d resnet

Prerequisites

python 3.5.2
pytorch 0.2.0 pyqt 5

Docker Image

docker pull likebullet86/luna16_detector

Using Viewer

• settings (detector_viewer/xai_viewer.py)
  • self.init_openpath
  • self.label_dirpath
  • self.detect_resume
  • self.gpu
• excute
  • python xai_viewer.py
• for using viewer in docker
  • xhost + (excute in host side)
  • excute docker with following options: -v "$HOME/.Xauthority:/root/.Xauthority:rw" -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=$DISPLAY

FROC curve

(0.74487168, 0.125)
(0.87307787, 0.25)
(0.89820999, 0.5)
(0.92245173, 1.0)
(0.93532807, 2.0)
(0.98055845, 4.0)
(0.99051809, 8.0)
0.906431

Data Download

• https://luna16.grand-challenge.org/download/
  • download data and candidates_V2.csv
• https://wiki.cancerimagingarchive.net/display/Public/LIDC-IDRI
  • download Radiologist Annotations/Segmentations (XML)

pretrained weight

• https://drive.google.com/drive/u/1/folders/15d8BsRLNkroAl0iypHaIrlRorhBImzYk
• download detector.ckpt

Training

• set luna and lidc path: set config_training.py
• export path: source export_path.sh
• train, val idx npy make: python preprocess/make_validate_npy.py
• preprocess data make: python preprocess/prepare.py
• train: sh train.sh

Make froc curve

• make test result bbox: sh test.sh
• make froc sumbit: python make_FROC_submit_native.py
• calc froc curve: sh eval.sh

Todo

• apply UNET base lung segment

Reference Code

https://github.com/lfz/DSB2017
https://github.com/juliandewit/kaggle_ndsb2017