DeepLungScreening Overview

This is the Deep Learning Open-source Lung cancer risk estimation repository, with PyTorch implementations. We prevoide source code, running tutorial, pre-trained model, docker, and example image in this repository.

Please see the requirement in the Dockerfile. Different version packages should be easy to adapt (not guarantee).

Tutorial

All the steps can be find in run_all.sh. A INPUTS and OUTPUTS example can be found here.

Fig. 1. Step 1-3 are summarized in this Figure

step 1: image pre-process

python3 ./1_preprocess/step1_main.py --sess_csv ${SPLIT_CSV} --prep_root ${PREP_ROOT} --ori_root ${ORI_ROOT}

${SPLIT_CSV} is the csv file stores the scan list (and meta data) you want to test. ${PREP_ROOT} is the data folder saves the pre-processed result. ${ORI_ROOT} is the data folder stores the scans you want to test.

step 2: nodule detection

python3 ./2_nodule_detection/step2_main.py --sess_csv ${SPLIT_CSV} --bbox_root ${BBOX_ROOT} --prep_root ${PREP_ROOT}

${BBOX_ROOT} is the data folder stores the nodule detection results.

step 3: feature extraction

python3 ./3_feature_extraction/step3_main.py --sess_csv ${SPLIT_CSV} --bbox_root ${BBOX_ROOT} --prep_root ${PREP_ROOT} --feat_root ${FEAT_ROOT}

${FEAT_ROOT} is the data folder stores the image feature results.

Fig. 2. Step 4 are summarized in this Figure

step 4: co-learning of meta data and image

python3 ./4_co_learning/step4_main.py --sess_csv ${SPLIT_CSV} --feat_root ${FEAT_ROOT} --save_csv_path ${PRED_CSV}

${PRED_CSV} is the predicted result from deep learning method.

Docker

The Dockerfile is shared in this repository. The docker image rg15/deeplung:0.3 can be downloaded from DockHub.