Source-linked AI summary
Non-Rigid Image Registration Using Self-Supervised Fully Convolutional Networks without Training Data
Hongming Li, Yong Fan
TL;DR
The paper addresses the limitation that supervised registration models depend on training transformations and proposes a self-supervised FCN framework that directly optimizes image similarity. Its multi-resolution method jointly learns spatial transformations without training deformation fields and outperformed the compared methods on reported brain-MR registration measures.
Problem
Supervised registration methods depend on training datasets with known spatial transformations and are limited to transformations represented in those datasets.
Method
A multi-resolution FCN framework jointly learns deformation fields by maximizing similarity between fixed and transformed moving images with smoothness regularization, without training deformation fields.
Results
0.720 mean Dice across 54 LPBA40 regions and 0.905 mean NCC on ADNI were obtained, with the method exceeding ANTs and alternative deep models on the reported comparisons.
Takeaways & Limitations
The method provides a self-supervised approach for non-rigid registration of 3D structural brain MR images with reported accuracy and computational-speed advantages.
Abstract
from arXiv · showhide
A novel non-rigid image registration algorithm is built upon fully convolutional networks (FCNs) to optimize and learn spatial transformations between pairs of images to be registered in a self-supervised learning framework. Different from most existing deep learning based image registration methods that learn spatial transformations from training data with known corresponding spatial transformations, our method directly estimates spatial transformations between pairs of images by maximizing an image-wise similarity metric between fixed and deformed moving images, similar to conventional image registration algorithms. The image registration is implemented in a multi-resolution image registration framework to jointly optimize and learn spatial transformations and FCNs at different spatial resolutions with deep self-supervision through typical feedforward and backpropagation computation. The proposed method has been evaluated for registering 3D structural brain magnetic resonance (MR) images and obtained better performance than state-of-the-art image registration algorithms.
1. INTRODUCTION
Existing registration methods learn spatial transformations from annotated training data or image features, limiting them to transformations represented in those datasets. Self-supervised approaches instead optimize image similarity between fixed and transformed moving images, motivating a data-free alternative.
- Supervised registration models predict spatial relationships from image patches using training datasets containing spatial transformations.
- Models trained on known transformations are limited to estimating spatial transformations captured by their training datasets.
- Recent self-supervised methods optimize image similarity between fixed and transformed moving images to estimate spatial transformations.
2. METHODS
The method formulates non-rigid registration as image-similarity optimization and uses FCNs with total-variation regularization to predict dense deformation fields. A multi-resolution, deeply self-supervised architecture jointly optimizes transformations without training deformation fields.
- 2.1. Image registration driven by image similarity metric: Non-rigid registration seeks a dense deformation field that establishes voxel-wise correspondence between fixed and moving images.
- 2.1. Image registration driven by image similarity metric: The optimization maximizes image similarity between the fixed image and transformed moving image while regularizing the deformation field.
- 2.1. Image registration driven by image similarity metric: A total variation regularizer is used to encourage spatially smooth and physically plausible transformations.
- 2.2. Image registration using FCNs: FCNs receive concatenated fixed and moving images and learn feature representations and spatial transformations through voxel-to-voxel regression.
- 2.3. Multi-resolution image registration with deep self-supervision: The multi-resolution network jointly optimizes deformation fields at multiple spatial resolutions using feedforward computation and backpropagation.
- 2.3. Multi-resolution image registration with deep self-supervision: No training deformation fields are required; image similarity and deformation-field smoothness provide the self-supervision, and the trained network registers image pairs directly.
3. RESULTS
The method was evaluated on ADNI and LPBA40 brain MR datasets against deep-learning baselines and ANTs. Multi-resolution registration achieved the strongest reported similarity and label-overlap results, while registered mean images retained detailed textures.
- Datasets and evaluation: The evaluation used hippocampus-labeled ADNI MR images and LPBA40 MR images with 56 brain-region labels.
- Datasets and evaluation: The method was compared with coarse-grained deep-learning registration models and ANTs on the same datasets.
- ADNI results: 0.905 mean NCC was obtained by the multi-resolution model, exceeding 0.616 without pooling layers and 0.848 with pooling layers.
- Visual results: Registered mean images preserved detailed textures, and the proposed method showed sharper visual contrast than ANTs.
- LPBA40 results: 0.720 mean Dice across 54 ROIs was obtained by the proposed method, compared with 0.697 for ANTs and 0.639 before registration.
- LPBA40 results: The proposed method achieved significantly higher Dice values than ANTs for 35 of 54 ROIs, with p<0.02 by the Wilcoxon signed rank test.
4. CONCLUSION
The paper presents a deep-learning non-rigid registration algorithm for learning spatial transformations between image pairs. Experiments on 3D structural MR images indicate promising accuracy and computational speed.
- The proposed algorithm learns spatial transformations between pairs of images for non-rigid registration.
- Experiments on 3D structural MR images demonstrated promising image-registration accuracy and computational speed.