NanoMASK

A deep learning algorithm for auto-segmentation and pharmacokinetic quantitation for PET/CT functional imaging of nanomedicines

Installation

1. Create a virtual environment

bash conda create -n nanomask python=3.8 -y conda activate nanomask

1. Install required packages

```bash

install pytorch following the official guideline

https://pytorch.org/get-started/locally/

conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia cd NanoMask pip install -e . ```

Preprocessing your mouse PET/CT data

1. Convert PET/CT data format

   Run

   bash python dcm2nii.py -i input_path -o output_path -a CT

- `input_path`: the input path contains three folders: 'CT' for CT data, 'PET' for PET data, and 'ROI' for organ contours.
- `CT`: this can only be CT or PET, which represents converting organ contours according to CT data or PET data.

1. Register PET/CT data

   Install Greedy for image registration from https://greedy.readthedocs.io/en/latest/install.html.

   Use affine registration to register PET/CT data

   bash python affine_registration.py -i input_path -m CT

- `input_path`: the input path contains PET, CT data and organ contours. All in '.nii.gz' format.
- `CT`: this can only be CT or PET. 'CT' represents using CT data ('CT_0000.nii.gz') as the moving image and PET data ('PET_0001.nii.gz') as the fixed image. Apply Affine transformation to register CT and organ contours ('organs-CT.nii.gz') to PET. 'PET' represent using PET data as the moving image and CT data as the fixed image. 

Organ segmentation for mouse PET/CT data

Download the pretrained model and put it into ./NanoMask/nnunet/nnUNet_data/nnUNet/3d_fullres. Run

bash nnUNet_predict -i input_path -o output_path -m 3d_fullres -t 006 -f 0

Notes: - input_path: Please use absolute path and the path cannot have spaces. CT and PET data should have the suffix _0000.nii.gz, _0001.nii.gz, respectively. E.g., Case1_0000.nii.gz, Case1_0001.nii.gz - ouput_path: the segmentation results will be saved in this path.

Train the model based on your own data

Organize the dataset as follows

Task001_MousePETCT - imagesTr - case1_0000.nii.gz # CT - case1_0001.nii.gz # PET - imagesTs - labelsTr - case1.nii.gz # ground truth - dataset.json Run

bash nnUNet_train 3d_fullres nnUNetTrainerV2 Task001_MousePETCT 0

Evaluate your segmentation result

1. Compute Dice Similarity Coefficient and Volume Difference

   bash python compute_metrics.py -s segmentation_path -r ground_truth_path -o output_path -n Heart Lungs Liver Spleen Kidneys Tumor

   Notes:

   • Heart Lungs Liver Spleen Kidneys Tumor: the mapped integers of each organ in ascending order
   • ouput_path: the evaluation results, DSC and VD, will be saved in this path in a csv file.

2. Compute clinical metrics based on three groups metrics data: machine, ground truth contours, and segmentation contours

   This includes mean voxel intensity, total region volume, total/max/min/std intensity in the contour, percent injected dose per cubic centimeter

   bash python compute_clinical.py -s segmentation_path -r ground_truth_path -i image_path -o output_path -m machine_path -n Heart Lungs Liver Spleen Kidneys Tumor

   Notes:

   • machine_path: the path of xlsx file containing machine metrics including time of scan start, time of injection (datatime) and injected dose (mCi), machine percent injected dose per cubic centimeter (%ID/cc) for each timepoint.
   • Heart Lungs Liver Spleen Kidneys Tumor: the organs in ascending order according to their mapped integers.
   • ouput_path: the clinical metrics will be saved in this path in an xlsx file.

Acknowledgement

This customized model is based on the nnUNet framework. Thanks for the nnUNet team very much!