ypy516478793 / lung_nodule_detection_pytorch

Incidental Lung Nodule Project Documentation

Nodule Detection

Data Preparation

Useful Information

• Code location: Z:\Incidental_Lung_Project\PycharmProjects\Data_pulling
• Data location: Z:\Incidental_Lung_Project\Methodist_incidental_Kim
• Possible series name of lung window: ["lung", "LUNG", "LUNGS", "LUNG WINDOWS", "lung window", "Lung", "Lung 5.0", "LUNG 10.0 CE", "Chest", "PRE CHEST", "CT SLICES 50cm DFOV", "CT SLICES 50cm", "CTA CHEST 5.0", "HIGH RES CHEST", "HIGH RESOLUTION CHEST", "INSPIRATION LUNG" "CT Atten Cor Head In 3.75thk", "CT ATTEN COR", "SUPER D CHEST", "INSPIRATION LUNGS", "PULMON 2.0 B60f", "CHEST WITH", "Thorax 5.0 B31f", "Thorax 3.0 B31f", "CT Atten Cor Head In", "VOLUMEN MEDIASTINO S/C", "STANDARD FULL", "Abd Venous 6.0 B60f", "CHEST 6.0 B65f", "Abdomen Arteri 1.5 B20f", "Thorax 1.5 B40f", "Standard/Full", "Recon2 Standard/Full", "CTAC HEAD IN TRANS TOMO", "CT SLICES 50", "WB CT SLICES", "CT WB 5.0 B31s", "CT Atten Cor 3.75thk"]

Workflow

Create details.xlsx/checklist.xlsx --> Pull dicom images --> Get raw/normalized data saved in .npz format --> Annotate the central slices --> Generate location label file pos_label_norm.csv --> Preprocess data by lung segmentation and cropping.

1. Create details.xlsx/checklist.xlsx

• details.xlsx is used to save the main information (patient-level) for this dataset, it contains the patient Idx (pstr), MRN, annotated nodules, pathology label, has annotation flag, date of surgery, assign to. Bold items are information must be given at the beginning. Sample details.xlsx:

• checklist_{ext}.xlsx can be generated from the details.xlsx and it is used to control the data (nodule-level) we want to precessed at the round, where {ext} is an extra string to identify this checklist. Each checklist contains the patient Idx (pstr), MRN, date (dstr), Series, z, nodule Index, main nodule, PET, position, szie, compare with. date (dstr), Series, z, nodule Index, main nodule, PET, position, szie, compare with, can be obained from checking the report on the APEC system. z is not required if we don't want annotate this nodule (e.g. in inference mode). Series is necessary for external exams (CT obtained outside methodist). Sample checklist.xlsx:

2. Pull dicom images [On Windows]

In order to pull the specified dicom images, we need to first generate a list of identifiers to specify what data to pull from the database, then log into the computer that can access the database t run the batch code to pull the dicom data automatically.

Generate t of i to get identifiers.csv

python data_utils.py get_id -s={SAVE_DIR} -c={CK_DIR} -p={PRE_DIR} -ext={EXTRA_STR}

• get_id : Task is to get identifiers.csv
• --save_dir(-s): [str] Save directory
• --ck_dir(-c): [str] Checklist directory
• --pre_dir(-p): [str, optional] Directory where some data is preprocessed/pre-downloaded
• --extra_str(-ext): [str, optional] Extra string for the checklist

If --pre_dir is given, there will be a move_data.csv indicating the data to be moved. Run the following command to move those pre-downloaded data: python data_utils.py move_data -s={SAVE_DIR}

• move_data: Task is to move pre-downloaded dicom data to save directory
• --save_dir(-s): [str] Save directory

Example: If the checklist file is located at Z:\Methodist_incidental_Kim\checklist_Ben.xlsx, then -c=Z:\Methodist_incidental_Kim, -ext=Ben,

Automatically download dicom data

• Set ROOTFOLDER in pull_dicom_data.bat file by: set ROOTFOLDER={SAVE_DIR}, where {SAVE_DIR}is where identifiers.csv located. Note that {SAVE_DIR} must ends with a "\" here.
• Run .\pull_dicom_data.bat in terminal.

In the end, all dicom data will be saved in {SAVE_DIR}, Note that all series are saved, but normally we only need Lung Window (PET or not PET). The data strucuture is shown as followed:

SAVE_DIR
|   identifiers.csv
|   move_data.csv
└── Lung_patient{pstr0}-{MRN0}_{date0}
|    |   01DATAINFO.txt
|    |   uids.csv
|    └── {date0}_CT_data
|         |   slice0.dcm
|         |   ... 
└── Lung_patient{pstr0}-{MRN0}_{date1}
│    |   ... 

3. Get raw/normalized data saved in .npz format [On Windows]

This step is to de-identify the data by saving it in .npz format. python data_utils.py get_npz -r={ROOT_DIR} -c={CK_DIR} -n={NORMALIZE} -ext={EXTRA_STR}

• get_npz : Task is to save data in .npz format
• --root_dir(-r): [str] Root directory that stores the downloaded dicom data
• --save_dir(-s): [str] Save directory that stores the npz data
• --ck_dir(-c): [str] Checklist directory
• --normalize(-n): [bool] Whether to normalize data
• --pre_dir(-p): [str, optional] Directory where some data is preprocessed and saved in .npz format
• --extra_str(-ext): [str, optional] Extra string for the checklist

Same as in step 2, If --pre_dir is given. Run the following command to move those pre-processed data:
python data_utils.py move_data -s={SAVE_DIR}

In the end, all npz data will be saved in {SAVE_DIR}\normalized or `{SAVE_DIR}\raw``, *Note that only the corresponding series are saved. The data strucuture is shown as followed:

SAVE_DIR
└── normalized/raw
|    |   CTinfo.npz
|    |   move_data.csv
|    |   log
|    └── Lung_patient{pstr0}-{MRN0}
|         |   {MRN0}-{date0}.npz
|         |   {MRN0}-{date1}.npz
│         |   ... 
|    └── Lung_patient{pstr10}-{MRN10}
|    |   ... 

4. Annotate the central slices [On Linux]

Extract the central slices from the de-identified data.

python prepare.py extract -r={ROOT_DIR} -s={SAVE_DIR} -p={CK_PATH} -n={NORMALIZE}

• extract : Task is to extract central slices
• --root_dir(-r): [str] Root directory that stores the npz data
• --save_dir(-s): [str] Save directory that stores the central slices, they will be saved in SAVE_DIR/central_slices_{norm/raw}
• --ck_path(-p): [str] Checklist file full path
• --normalize(-n): [bool] Whether to normalize data

Mannually annotation based on the the central slices and checklist.xlsx

Go to open-source website make sense to annotate all the central slices by hand. Save the annotations as a single csv file (e.g. labels_my-project-name_2021-04-20-11-13-39.csv). Recommended to annotate the normalized data.

Create GT label (Change the format of annotations)

python prepare.py convert -r={ROOT_DIR} -s={SAVE_DIR} -a={ANNOT_FILE} -n={NORMALIZE}

• convert : Task is to convert annotation from make sense output to normalized ground truth pos_label_norm.csv
• --root_dir(-r): [str] Root directory that stores the npz data
• --save_dir(-s): [str] Save directory that contains the make sense annotation file and will be used to save GT label
• --annot_file(-a): [str] Name of the annotation file from make sense
• --normalize(-n): [bool] Whether the annotated data is normalized or not

5. Preprocess data by lung segmentation and cropping

python prepare.py prep_methodist -s={SAVE_DIR} -r={ROOT_DIR} -m={MASK} -c={CROP}

• prep_methodist: Task is to do preprocessing for methodist data
• --root_dir(-r): [str] Root directory that stores the npz data
• --save_dir(-s): [str] Save directory that is used to save preprocessed data
• --mask(-m): [bool] Apply unsupervised lung mask in preprocessing
• --crop(-c): [bool] Crop masked images in preprocessing

Run detection code

Train

1. Set DATA_DIR, PET_CT, POS_LABEL_FILE = "pos_labels_norm.csv" arguments in methodistFull.py file
2. python detect.py -d=methodistFull --test=False --gpu="2,3" --save-dir=CODETEST --resume="../detector_ben/results/PET_newMask_rs42_augRegular_5fold_2/029.ckpt" --start-epoch=0 --best_loss=0.3076

Test

python detect.py -d=methodistFull --test=True --gpu="2,3" --save-dir=CODETEST --resume="../detector_ben/results/PET_newMask_rs42_augRegular_5fold_2/029.ckpt"

Inference

1. Set POS_LABEL_FILE = None arguments in methodistFull.py file.
2. python detect.py -d=methodistFull --test=True --gpu="2,3" --resume="../detector_ben/results/res18-20201202-112441/026.ckpt"

Run kfold for Methodist data

1. Set DATA_DIR, PET_CT, POS_LABEL_FILE = "pos_labels_norm.csv" arguments in methodistFull.py file
2. Set DATA_DIR, SAVE_DIR arguments in PET_aug_newMask.sh file (in scripts folder).
3. Run PET_aug_newMask.sh.

Arguments

methodistFull.py (in dataLoader folder)

• MASK_LUNG: [bool] Unsupervised lung mask was applied in preprocessing
• CROP_LUNG: [bool] Cropped masked images in preprocessing
• DATA_DIR: [str] Data directory that stores the npz data
• POS_LABEL_FILE: [bool] Cropped masked images in preprocessing
• BLACK_LIST: [list] List of image to be excluded from the dataset
• AUGTYPE: [dict] Dict of augmentation options to be used

luna.py (in dataLoader folder)

• DATA_DIR: [str] Data directory that has root folder (LUNA16)
• TRAIN_DATA_DIR: [str] List of subdirectories for training
• VAL_DATA_DIR: [str] List of subdirectories for validation
• TEST_DATA_DIR: [str] List of subdirectories for test

detect.py

• --datasource(-d): [str] Datasource (options: [luna, methodist])
• --model(-m): [str] Model to be used (default: 3D Resnet-18)
• --epochs(-e): [int] Number of total epochs to run
• --start-epoch: [int] Manually set start epoch number
• --batch-size(-b): [int] Batch size
• --learning-rate(-lr): [float] Initial learning rate
• --resume(-re): [str] Path to latest checkpoint (default: none)
• --save_dir(-s): [str] Save directory that is used to save results
• --test(-t): [bool] True if in test mode, otherwise in train or inference mode
• --inference(-i): [bool] True if in inference mode, otherwise in train or test mode
• --gpu: [str] GPUs to be used
• --kfold: [int] Number of kfold for train_val
• --split_id: [int] Split id when use kfold