High Precision Localization of Pulmonary Nodules on Chest CT utilizing Axial Slice Number Labels

This is a repository accompanying the manuscript of the same name.

Note:

This model utilizes the pytorch Retinanet implementation available at https://github.com/yhenon/pytorch-retinanet/. The relevant code has been included in this repo, but see their documentation for further information.

Data:

To train and test model, users will need to obtain access to LUng Nodule Analysis (LUNA) data and National Lung Cancer Screening Trial (NLST) data.

Installation Instructions:

1. Clone this repository

2. Install pytorch and torchvision: https://pytorch.org/

3. Install https://github.com/yhenon/pytorch-retinanet/ dependencies:

pip install cffi
pip install pandas
pip install cython
pip install opencv-python
pip install requests

4. Install additional python dependencies

• numpy
• matplotlib
• argparse
• pydicom
• skimage
• SimpleITK
• scipy
• XGBoost
• sklearn
• imgaug
• joblib
• csv

Pipeline:

1. Generate Maximum Intensity Projection (MIP) LUNA data:

generate_data_LUNA_coronalaxial_combined.py \
 --LUNA_DATA_DIR "directory to LUNA .mhd files" \
 --LUNA_ANNS "directory to LUNA annotations" \
 --no_masking \
 --mip 25 \
 --positives_only \
 --SAVE_DIR "directory where to save formatted LUNA data" \

2. Train retinanet model:

train.py \
  --csv_train "path to annotations_train.csv in the formatted LUNA data directory" \
  --csv_classes "path to classes.csv in the formatted LUNA data directory" \
  --csv_val "path to annotations_val.csv in the formatted LUNA data directory" \
  --csv_test "path to annotations_test.csv in the formatted LUNA data directory" \
  --depth 101 \
  --model_output_dir "directory to save model" \
  --lr 1e-5 \
  --epochs 30 \
  --train_all_labeled_data

Note: csv_train, csv_val, and csv_test will be grouped together for training if --train_all_labeled_data is specified. Otherwise, train and val will be grouped to become the train data, and test set will become the val data. No test data will be used.

3. Inference raw NLST data:

analysis_nlst.py \
  --datadir "path to NLST CT data" \
  --anns "path to NLST nodule annotations 'sctabn.csv'" \
  --model "path to saved from from train.py" \
  --results_dir "directory where to save raw inference results" \
  --mip 25 \
  --no_mask \
  --pids "(Optional) specify specific PIDs to inference on"

4. Perform DBSCAN unsupervised clustering and apply False Positive Reducer:

clustering.py \
  --nlst_dir "path to NLST CT data" \
  --nlst_anns "path to NLST nodule annotations 'sctabn.csv'" \
  --inference_dir "directory where raw inference results were saved" \
  --save_dir "directory where to save clustered & fp_reduced results" \
  --conf_thresh 0.10 \
  --fp_reducer './fp_reducer_final.joblib' # path to FP reducer (XGBoost Classifier)

5. Apply axial slice localizer:

localize_with_slicenum.py \
  --inferences_dir "directory where raw inference results were saved" \
  --nodules_filepath "where clustered nodule results \ (nodules.csv) from clustering.py are located" \
  --nlst_dir "path to NLST CT data" \
  --nlst_anns "path to NLST nodule annotations 'sctabn.csv'" \