Introduction to DeepReg Demos¶
DeepReg offers multiple built-in dataset loaders to support real-world clinical scenarios, in which images may be paired, unpaired or grouped. Images may also be labeled with segmented regions of interest to assist registration.
A typical workflow to develop a registration network using DeepReg includes:
Select a dataset loader, among the unpaired, paired and grouped, and prepare data into folders as required;
Configure the network training in the configuration yaml file(s), as specified in supported configuration details;
Train and tune the registration network with the command line tool
Test or use the final trained registration network with the command line tool
Besides the tutorials, a series of DeepReg Demos are provided to showcase a wide range of applications with real clinical image and label data. These applications range from ultrasound, CT and MR images, covering many clinical specialties such as neurology, urology, gastroenterology, oncology, respiratory and cardiovascular diseases.
Each DeepReg Demo provides a step-by-step instruction to explain how different scenarios can be implemented with DeepReg. All data sets used are open-accessible. Pre-trained models with numerical and graphical inference results are also available.
DeepReg Demos are provided to demonstrate functionalities in DeepReg. Although effort has been made to ensure these demos are representative of real-world applications, the implementations and the results are not peer-reviewed or tested for clinical efficacy. Substantial further adaptation and development may be required for any potential clinical adoption.
The following DeepReg Demos provide examples of using paired images.
This demo registers paired CT lung images, with optional weak supervision.
This demo registers paired preoperative MR images and 3D tracked ultrasound images for locating brain tumours during neurosurgery, with optional weak supervision.
This demo registers paired MR-to-ultrasound prostate images, an example of weakly-supervised multimodal image registration.
The following DeepReg Demos provide examples of using unpaired images.
This demo compares three training strategies, using unsupervised, weakly-supervised and combined losses, to register inter-subject abdominal CT images.
This demo registers unpaired CT lung images, with optional weak supervision.
This demo aligns hippocampus on MR images between different patients, with optional weak supervision.
This demo registers 3D ultrasound images with a 9-fold cross-validation. This strategy is applicable for any of the available dataset loaders.
The following DeepReg Demos provide examples of using grouped images.
This demo registers grouped masks (as input images) of prostate glands from MR images, an example of feature-based registration.
This demo registers grouped CMR images, where each group has multi-sequence CMR images from a single patient.
The following DeepReg Demos provide examples of using classical registration methods.
This demo registers head-and-neck CT images using iterative affine registration.
This demo registers prostate MR images using iterative nonrigid registration.