[Paper Review] Introduction to Medical Image Registration with DeepReg, Between Old and New

Nina Montaña-Brown, Yunguan Fu+11 more|arXiv (Cornell University)|Aug 29, 2020

Radiomics and Machine Learning in Medical Imaging11 references4 citations

TL;DR

This tutorial introduces DeepReg, an open-source deep learning framework for medical image registration, demonstrating its integration of classical optimization-based methods and modern learning-based approaches. It enables researchers to implement and compare rigid, non-rigid, and deep learning-based registration using real clinical data, with end-to-end code examples and GPU-accelerated training via TensorFlow 2, achieving accurate image alignment in practical clinical scenarios.

ABSTRACT

This document outlines a tutorial to get started with medical image registration using the open-source package DeepReg. The basic concepts of medical image registration are discussed, linking classical methods to newer methods using deep learning. Two iterative, classical algorithms using optimisation and one learning-based algorithm using deep learning are coded step-by-step using DeepReg utilities, all with real, open-accessible, medical data.

Motivation & Objective

• To provide a practical, code-based tutorial for researchers new to medical image registration using the open-source DeepReg framework.
• To bridge classical optimization-based registration methods with modern deep learning approaches through unified implementation in DeepReg.
• To demonstrate the use of DeepReg for real-world clinical applications, including multi-modal and longitudinal image registration.
• To enable researchers with minimal coding experience to reproduce and extend state-of-the-art registration techniques using pre-packaged, tested components.
• To promote community-driven, clinically relevant research by providing accessible, well-documented tools and examples.

Proposed method

• The tutorial implements three registration strategies: two classical iterative algorithms (using optimization) and one deep learning-based method (using a U-Net-like network) via DeepReg’s high-level API.
• It leverages TensorFlow 2 and Keras-compatible layers to define and train registration models, including loss functions for image similarity and smoothness of deformation fields.
• A pre-trained deep learning model is fine-tuned on open-access clinical datasets (e.g., T2-weighted MRI and label maps) using a paired training setup.
• The framework supports both image and label warping, with outputs visualized using slice-by-slice comparison of fixed, moving, warped, and ground truth images.
• Training is accelerated using GPU support via Google Colab, with model inference and visualization pipelines implemented in Jupyter notebooks.
• The codebase includes utilities for data loading, configuration parsing (YAML), and model checkpointing, ensuring reproducibility and extensibility.

Experimental results

Research questions

• RQ1 How can classical image registration algorithms be implemented and compared with deep learning-based methods using a unified framework?
• RQ2 To what extent does DeepReg enable seamless integration between traditional optimization-based methods and modern deep learning architectures?
• RQ3 Can a single, modular framework like DeepReg support diverse clinical registration tasks, including multi-modal and longitudinal imaging?
• RQ4 How does the performance of a learned registration model compare to classical iterative methods in terms of alignment accuracy and computational efficiency?
• RQ5 What is the practical usability of DeepReg for researchers with minimal deep learning experience in medical imaging?

Key findings

• The tutorial successfully demonstrates end-to-end implementation of classical and deep learning-based registration using DeepReg, with full reproducibility via Jupyter notebooks.
• The deep learning model achieves accurate warping of moving images and labels to align with fixed images, as evidenced by visual comparisons of predicted and ground truth slices.
• Training on a GPU via Google Colab completes in a few minutes, with inference times suitable for clinical workflow integration.
• The framework enables consistent and modular implementation of registration pipelines, reducing development time and increasing reproducibility.
• The integration of loss functions for image similarity and deformation smoothness leads to realistic and anatomically plausible deformation fields.
• The tutorial's code and configuration files are publicly available, supporting community extension and adaptation to new clinical datasets.