An Unsupervised Learning Model for Deformable Medical Image Registration (1802.02604v3)

Abstract: We present a fast learning-based algorithm for deformable, pairwise 3D medical image registration. Current registration methods optimize an objective function independently for each pair of images, which can be time-consuming for large data. We define registration as a parametric function, and optimize its parameters given a set of images from a collection of interest. Given a new pair of scans, we can quickly compute a registration field by directly evaluating the function using the learned parameters. We model this function using a convolutional neural network (CNN), and use a spatial transform layer to reconstruct one image from another while imposing smoothness constraints on the registration field. The proposed method does not require supervised information such as ground truth registration fields or anatomical landmarks. We demonstrate registration accuracy comparable to state-of-the-art 3D image registration, while operating orders of magnitude faster in practice. Our method promises to significantly speed up medical image analysis and processing pipelines, while facilitating novel directions in learning-based registration and its applications. Our code is available at https://github.com/balakg/voxelmorph .

• The paper introduces VoxelMorph, an unsupervised CNN-based framework for fast 3D medical image registration.
• It employs a UNet-style encoder-decoder with skip connections and a spatial transformer to compute deformation fields.
• Evaluation on 7,829 MRI scans shows Dice scores comparable to SyN while reducing registration time from hours to seconds.

An Unsupervised Learning Model for Deformable Medical Image Registration

The paper presents an innovative unsupervised learning framework for deformable, pairwise 3D medical image registration utilizing a convolutional neural network (CNN). The proposed model, termed VoxelMorph, seeks to address the computational inefficiencies of traditional registration methods by learning a global function optimization aggregated over a dataset during a training phase, rather than optimizing independently for each image pair.

Overview and Methodology

The authors define the registration task as a parametric function using a CNN to map input images to a shared representation. This model can take new image pairs and perform rapid registration by evaluating a pre-trained function. Notably, the method requires no supervised data, such as ground truth correspondences, during training, which is a significant departure from many existing techniques that rely on supervised data.

The architecture of VoxelMorph is based on an encoder-decoder with skip connections, akin to UNet structures, designed to produce a displacement field $\phi$ mapping voxels between images. The network applies convolutional layers and Leaky ReLU activations to produce the registration field. A spatial transformer is used to warp images, facilitating the differentiable computation of image similarity which serves as part of the loss function.

Evaluation and Performance

The paper evaluates VoxelMorph using a large-scale, multi-site dataset consisting of 7,829 MRI brain scans. The benchmark comparison with Symmetric Normalization (SyN), a state-of-the-art algorithm from the ANTs software package, reveals that VoxelMorph achieves comparable accuracy but operates orders of magnitude faster. Dice scores, which measure anatomical overlap, show the model performs on par with ANTs. Specifically, VoxelMorph models registered images in under a second using a GPU and offered results within two minutes on a CPU, compared to hours required by ANTs.

VoxelMorph is robust across a variety of population subgroups, yielding enhanced Dice scores for targeted datasets like ABIDE, which consists of subjects with autism. This adaptability suggests VoxelMorph can tailor its parameters for specific subpopulations, enhancing registration precision.

Key Contributions and Implications

The authors demonstrate that learning-based registration can achieve state-of-the-art accuracy without the computational burden associated with traditional methods. The implication for medical imaging is profound, potentially accelerating the processing of patient data and facilitating real-time analyses in clinical settings. Furthermore, since VoxelMorph is not limited to medical images, it opens pathways for applying similar methodologies to other imaging and registration domains.

Future Directions

The paper discusses the potential for improvements in learning-based registration by incorporating diffeomorphic properties common in traditional methods, which guarantee invertibility and topological preservation. Future work may explore integrating or approximating these properties in the VoxelMorph framework to broaden its application in analyses requiring these guarantees.

Conclusion

This work provides substantial evidence for the efficacy and efficiency of unsupervised, learning-based approaches in medical image registration. By eliminating the need for supervised inputs and dramatically reducing computational time, VoxelMorph represents a significant advancement in the field, promising to transform current medical image processing workflows and drive novel applications in imaging research.