Clinical evaluation of deep learning for tumour segmentation on PET/CT scans of head and neck cancer

This repository contains links to the exact versions of our computer code underlying the article "Clinical evaluation of deep learning for tumour segmentation on PET/CT scans of head and neck cancer". This includes the code used to develop five deep learning models and supporting code used for pre-processing and model evaluation.

Background

Our research group implemented established deep learning frameworks and investigated their ability to delineate tumour volumes on PET/CT images.

Information about this study inluding the protocol and statistical analysis plan (SAP) is available at:

https://www.rigshospitalet.dk/english/departments/centre-of-diagnostic-investigation/department-of-clinical-physiology-nuclear-medicine-and-pet/research/research-projects/Pages/default.aspx

and

https://www.isrctn.com/ISRCTN16907234?q=16907234&filters=&sort=&offset=1&totalResults=1&page=1&pageSize=10

Please cite the following paper, if you use the data underlying this publication or the tools shared in this repository: (Paper will be added once published).

The purpose of the developed models was to delineate the gross tumour volume based on [18F]FDG-PET/CT scans (GTV-PET). A clinical example is provided in the figure below.

Introduction

This repository holds all code we used to implement five different openly available Deep Learning methods. The links marked "our copy" hold the exact versions of the code, that were used in the publication. Here, it is possible to review the exact (few) changes we made to the original code of the included methods.

Methods implemented:

• DeepMedic. Original repo. Our copy.
• nnU-Net. Original repo. Our copy.
• Tureckova. Original repo. Our copy.
• Swin-UNETR. Original repo. Our copy.
• InnerEye. Original repo. Our copy.

The MEDIcaTe package in this repository contains the tools used to pre-process and analyze image data (this draws heavily on well-established tools - you might prefer to use those directly.)

The folder "method_implementation" uses the MEDIcaTe package and contains all code necessary use the deep learning methods, i.e. to ensure the data adheres to the specifications of each method, and code to calculate evaluation metrics and visualise the output of the methods.

System information

Operating System: Ubuntu 20.04.4 LTS

• Kernel: Linux 5.13.0-51-generic
• Architecture: x86-64
• CUDA/cuDNN version: pytorch 1.11.0 py3.7_cuda11.3_cudnn8.2.0_0 pytorch cudatoolkit 11.3.1 h2bc3f7f_2 '''

GPU models and configuration:

• NVIDIA A6000
• NVIDIA-SMI 470.57.02
• Driver Version: 470.57.02
• CUDA Version: 11.4