Synopsis

This repo stores partial artifacts of training nnU-Net and U-Net-based models for the robust segmentation of COVID-19 lesions in lung CT scans. The repo contains training scripts, checkpoints, and additional scripts for maximum intensity projection, lung lobe segmentation extraction, and result checking using a dataset from the Santa Barbara Cottage Hospital.

Introduction

• Saved models are saved in ./runs (for U-Net) and ./nnunet_runs (nnU-Net)
  • Subdirectories represent Epoch number for training
  • i.e. ./nnunet_runs/500 represents nnU-Net trained on 500 epochs
• Training and inference files are run_net.py (U-Net) and run_nnUNet.py (nnU-Net)
  • These models are based on MONAI and are tested with the Grand Challenge Dataset
  • Training: /home/s_shailja/Fall2020/COVID-19-20_v2/Train
  • Testing: /home/s_shailja/Fall2020/COVID-19-20_v2/Validation
• Evaluation & Inference scripts for Cottage Hospital Data are located in ./CottageWork
  • separate README located in ./CottageWork/README.md

Running Instructions

• Command Line Usage
  • Modify source file (runnet.py / runnnUNet.py), data folder (Train / Validation), model folder (APPEND EPOCH NUMBER; i.e. --model_folder "runs/300"
  • Modify train or infer
  • python run_net.py train --data_folder "/home/s_shailja/Fall2020/COVID-19-20_v2/Train" --model_folder "runs"
    • output segmentation files to ./output
  • GPU selection at start of script

Cottage Scripts, Lung Lobe Processing, and Post Processing

Located in ./CottageWork Documentation below can also be found as a README.md in ./CottageWork

Work Instructions

• Use LungLobeProcessing.py to add lung lobe segmentation results to raw data
  • Input raw data dimensions: (x, y, num_slices)
  • Output NIFTI file dimensions: (x, y, num_slices, 7)
• Run Infer_nnUNet.py to segment data (provide path & segmentation model path)
  • python Infer_nnUNet.py infer --data_folder "/home/claire/data/nifti/COVID_nifti" --model_folder "/home/alex/nnunet_runs/500"
• Compare segmentation results with CSV-saved data, generate and print statistics
  • Complete with ResultChecker.py, which uses SegmentationParser.py and DataParser.py
• Postprocessing with PostProcessing.py - improve accuracy and metrics by applying Lung Mask
  • Make PostProcessing object with desired segmentation folder, raw unsegmented data folder, and postprocessing output folder, or run PostProcessing.py for default parameters

Note: most scripts include GPU selection code; please modify accordingly.

File Descriptions

DataParser.py - Reads annotated CSV & turns into Python data - Format: dictionary --> key = subject ID, val = [all lesion slices]

Infer_unet.py - Segments provided data & saves result to ./unetsegmentedoutput - Default parameters uses 300 epoch U-Net model

Infer_nnUNet.py - Segments provided data with 500 epoch nnU-Net model

SegmentationParser.py - Reads segmented results (from ./segmentation_output) & turns into Python data - Format: dictionary --> key = subject ID, val = [all lesion slices]

Resultchecker.py - Generate statistics given segmented results in comparison with annotated CSV - Use method print_stats()

LungLobeProcessing.py - Adds lobe segmentation to 3D NIFTI raw data in form of one-hot encoding - Output dimension: (x, y, num_slices, 7) - Provide raw data folder and output folder; defaults in script

PostProcessing.py - Iterate through input files and generates output of processed files in the same format - If facing errors --> copy segmented results into output folder and let the script override the initial files