Title: Quicksilver: Fast predictive image registration – A deep learning approach Paper Link
Authors: Xiao Yang, Roland Kwitt, Martin Styner, Marc Niethammer
Association: University of North Carolina at Chapel Hill, Chapel Hill, USA
Submission: Jul 2017
“Failure case for IBSR18 dataset where LDDMM optimization generated very extreme deformations. From left to right: (a): moving image; (b): target image; (c): LDDMM optimization result; (d): predictionþcorrection result (LPC); (e): heatmap showing the differences between the optimization deformation and predicted deformation in millimeters. Most registration errors occur in the area of the cerebellum, which has been inconsistently preserved in the moving and the target images during brain extraction. Hence, not all the retained brain regions in the moving image have correspondences in the target image. Best-viewed in color.”
Image registration is a key component for medical image analysis to provide spatial correspondences.
…………………………………………………………………………………………………………………
»> Background of Deep Learning in Medical Image Analysis
»> A Review: Robot-Assisted Endovascular Catheterization Technologies:
…………………………………………………………………………………………………………………
»> Image Registration Basic Knowledge
»> Image Registration Literature Review
»> Slice-To-Volume Medical Image Registration Background
…………………………………………………………………………………………………………………
Contributions
(0) introduces Quicksilver, a fast deformable image registration method.
(1) propose a deep regression model to predict deformation parameters using image appearances in a time-efficient manner.
(2) focus on predictions for the Large Deformation Diffeomorphic Metric Mapping (LDDMM) model.
(3) introduce a new correction network which greatly increases the prediction accuracy of an already existing prediction network.
(4) open-source https://github.com/rkwitt/quicksilver
Background
Image registration = an optimization problem = optimizing the parameters of a transformation model
Process:
-
Detect features. An image feature is any protion of an image that can potentially be identified. Features can be points, lines, or corners, etc.
-
Match corresponinging features.
-
Infer geometric transformation. Fom the sets of matched-feature pairs, a geometric transformation function is inferred that maps features in one image onto the locations of the matching features in the other.
-
Use the geometric transformation to align one image with the other.
type of transformation
- Low-dimensional parametric registration models: affine, rigid
- High-dimensionalparametric registration models
- Non-parametric parametric registration models
However, Non-parametric parametric registration models have a very large numbers of parameters. Therefore, numerical optimzation to solve the registration problems becomes computationally costly, even with acceleration by graphics processing units (GPUs).
Motivation
(1) “achieve the best possible agreement between a transformed source and a target image, subject to transformation constraints.”
Network Design
” Fig. 2. 3D (probabilistic) network architecture. The network takes two 3D patches from the moving and target image as the input, and outputs 3 3D initial momentum patches (one for each of the x; y and z dimensions respectively; for readability, only one decoder branch is shown in the figure). In case of the deterministic network, see Sec. 2.2.1, the dropout layers, illustrated by , are removed. Conv: 3D convolution layer. Conv⊺: 3D transposed convolution layer. Parameters for the Conv and Conv⊺ layers: In: input channel. Out: output channel. Kernel: 3D filter kernel size in each dimension. Stride: stride for the 3D convolution. Pad: zero-padding added to the boundaries of the input patch. Note that in this illustration B denotes the batch size.”
“Fig. 3. The full prediction þ correction architecture for LDDMM momenta. First, a rough prediction of the initial momentum, mLP, is obtained by the prediction network (LP) based on the patches from the unaligned moving image, Mand target image, T, respectively. The resulting deformation maps Φ?1 and Φ are computed by shooting. Φ is then applied to the target image to warp it to the space of the moving image. A second correction network is then applied to patches from the moving image M and the warped target image T∘Φ to predict a correction of the initial momentum, mC in the space of the moving image, M. The final momentum is then simply the sum of the predicted momenta, m ¼ mLP þmC, which parameterizes a geodesic between the moving image and the target image.” 2.4.
Limitations
*
Performance
*
Other Useful Info.
Object Detection helpful slides link (D2L5 Insight@DCU Machine Learning Workshop 2017)