Multi-modal, multi-organ deep segmentation of salivary and lacrimal glands (SLG) in PSMA PET/CT images

This repository contains a deep nerual net model for fully-automated segmentation of all salivary and lacrimal glands on PSMA PET/CT images. The imaging tracer used in model development was [18F]DCFPyL.

Required model inputs: (a) DICOM PET series acquired with a PSMA tracer (tracers other than [18F]DCFPyL may work); (b) DICOM CT series.

The PET and CT inputs must be co-registered.

For details regarding model architecture, training methodology, and testing results please refer to the following publication: Toosi A, Chaussé G, Chen C, Klyuzhin IS, Benard F, Rahmim A. Multi-modal, multi-organ deep segmentation of salivary and lacrimal glands in PSMA PET/CT images. JNM Aug 2022; 63 (supplement 2): 2587. https://jnm.snmjournals.org/content/63/supplement_2/2587

---

Requirements

• Python 3.8+
• rt_utils
• Other Python packages: see requirements.txt

Installation

• You should be able to run the test inference script (main.py) from a native or virtual Python/Conda environment
• Ensure that the trained weights file _unet_512x512__UNET-BN_PET_ENCOREGCT__focal-loss_LR0.0001_20-epochs-constant_NB-24epochs.hdf5 is included in your Python search path

Configuration

In "main" set the variables CT_FOLDER and PET_FOLDER containing CT and PET series, respectively (for only one patient/study). Set "output_filename" to the desired output file location and name.

Usage (inference)

1. Copy DICOM PET and CT series in the pre-specified directories
2. Execute the file main.py

Note: Inference results will be saved as an RTSTRUCT file containing segmentation ROIs. The package "rt_utils" is used to generate the RTSTRUCT file.

Project Structure

psma-slg-pub/
├── main.py                # Python script for evaluating the model + utility functions for image pre-processing and QC
├── _unet_512x512__UNET-BN_PET_ENCOREGCT__focal-loss_LR0.0001_20-epochs-constant_NB-24epochs.hdf5        # Model trained parameters
├── README.md              # Project documentation
└── LICENSE                # License file

Licence

This project is licenced under the MIT License.

How to cite

If you are including this model or associated code into your projects, kindly include the following citation:

Toosi A, Chaussé G, Chen C, Klyuzhin IS, Benard F, Rahmim A. Multi-modal, multi-organ deep segmentation of salivary and lacrimal glands in PSMA PET/CT images. JNM Aug 2022; 63 (supplement 2): 2587. https://jnm.snmjournals.org/content/63/supplement_2/2587

Acknowledgments

This project was supported by the Canadian Institutes of Health Research Project grant PJT-162216, National Institutes of Health / Canadian Institutes of Health Research QIN grant 137993, and Mitacs Accelerate grant IT18063. Azure Cloud compute credits were provided by Microsoft for Health.