All kinds of models and scripts for data processing and visualizations.

Python3.12 is recommended. Libraries are managed by https://python-poetry.org/

After installing poetry, run this to start poetry venv:

poetry install

Create a jupyter kernel from venv:

poetry run python -m ipykernel install --user --name ctseg-py3.12

To edit or remove kernels:

jupyter kernelspec list
jupyter kernelspec remove KERNEL_NAME

• Main folder: 2d_seg_and_video
• Training: 2d_seg_and_video/2d_box_seg.ipynb
• Processing logic mostly in: 2d_seg_and_video/dataset.py
  • Cropping and resizing
  • Adding bbox prompt

• Green: true positive
• Red: false positive
• Purple: false negative

DeepLesion 3D subset of the ULS23 dataset

• 743 3D lesion segmentations
• 4538 2D slices
• train/val split done by lesions to avoid data leakage

Dataset that we are labeling and reviewing (2024-2025).

Script: notebooks/ct_video.ipynb

• Download all 6 .zip parts files from https://github.com/DIAGNijmegen/ULS23/
• Unzip all 6 parts such that they are in the same folder
  • The folder should contain fragmented zip files (.z01, .z02, ...) for all 6 parts
• Unzip the fragmented zip files

sudo apt update
sudo apt install p7zip-full

7z x ULS23_Part1.zip
7z x ULS23_Part2.zip
...

Labels in git repo - merge the label folder with the existing ULS23 data folder.

• Main folder: 3d_seg

• Project page: https://aim.hms.harvard.edu/ct-fm
• Model: https://huggingface.co/project-lighter/whole_body_segmentation

• patching.ipynb Scripts to split existing datasets into patches
• ct-fm.ipynb Scripts for testing and training ct-fm seg model
• lesion3D Dataset folder in the Kingston SSD

nnUNet repo All commands below should be run in the nnUNet project directory.

• nnunet_data_proc.ipynb Scripts to generate raw data
• nnUNet_raw Dataset folder in the Kingston SSD

Set env variables e.g. by creating an env.sh file:

export nnUNet_raw="path/to/nnUNet_raw"
export nnUNet_preprocessed="/path/to/nnUNet_preprocessed/"
export nnUNet_results="/path/to/nnUNet_results"

Apply them in terminal:

source env.sh

At 64GB memory, 2 processes -np 2 is the max possible on this dataset. Proocessed data will be stored at the nnUNet_preprocessed path.

nnUNetv2_plan_and_preprocess -d 1 -np 2

Train model dataset (1 in this example) on a specified cross-validation fold (0 in this example).

nnUNetv2_train 1 3d_fullres 0 --npz

Other training config are 2d, 3d_lowres, 3d_cascade_fullres. The log and results will be stored at the nnUNet_preprocessed path.

Save the best and final checkpoints from previous run before running this:

nnUNetv2_train 1 3d_fullres all --npz -pretrained_weights path/to/checkpoint.pth

• CT channel is always 0000
• Seg masks from CT-FM seg model is 0001
• Boxes and other priors are 0002

Sample dataset.json

{
   "channel_names":{
      "0":"CT",
      "1": "noNorm"
   },
   "labels":{
      "background":0,
      "lesion":1
   },
   "numTraining":782,
   "file_ending":".nii.gz"
}

• studies_ctfm_seg_mask.ipynb generates the anatomical mask channels
• studies_weak_labels.ipnyb generates the box and mask channels

The default 5 folds are used. I used the same folds every dataset by copying the splits_final.json file from the preprocessed folder of the first dataset i.e. nnUNet_preprocessed/Dataset001_3dlesion/

• 3d_seg/visual.ipynb and 3d_seg/visualize.py contains code for visualizing 3D predictions for all the studies along with the CT and label masks.
• Visualization of val set can be found on the Kingston SSD 3d_val_visualization folder

Both voxel and lesion level metrics are calculated.

• 3d_seg/metrics.ipynb contains code for calculating metrics
• 3d_seg/metrics/ contains the csv files with metrics for every lesion in val set