This will count as one of your downloads.
You will have access to both the presentation and article (if available).
Convolutional neural networks (CNNs) offer a promising means to achieve fast deformable image registration with accuracy comparable to conventional, physics-based methods. A persistent question with CNN methods, however, is whether they will be able to generalize to data outside of the training set. We investigated this question of mismatch between train and test data with respect to first- and second-order image statistics (e.g., spatial resolution, image noise, and power spectrum). A UNet-based architecture was built and trained on simulated CT images for various conditions of image noise (dose), spatial resolution, and deformation magnitude. Target registration error was measured as a function of the difference in statistical properties between the test and training data. Generally, registration error is minimized when the training data exactly match the statistics of the test data; however, networks trained with data exhibiting a diversity in statistical characteristics generalized well across the range of statistical conditions considered. Furthermore, networks trained on simulated image content with first- and second-order statistics selected to match that of real anatomical data were shown to provide reasonable registration performance on real anatomical content, offering potential new means for data augmentation. Characterizing the behavior of a CNN in the presence of statistical mismatch is an important step in understanding how these networks behave when deployed on new, unobserved data. Such characterization can inform decisions on whether retraining is necessary and can guide the data collection and/or augmentation process for training.
Methods: An atlas was constructed with segmented pelvis shapes containing standard reference trajectories for screw placement. A statistical shape model computed from the atlas is used for deformable registration to the patient’s preoperative CT (without segmentation). By transferring the reference trajectories and surrounding acceptance windows (i.e., volumetric corridors of acceptable device placement) from the atlas, the system automatically computes reliable Kwire and screw trajectories for guidance (overlay in fluoroscopy) and QA.
Results: A leave-one-out analysis was performed to evaluate the accuracy or registration and overlay. The registration achieved average surface registration accuracy of 1.82 ± 0.39 mm. Automatically determined trajectories conformed within acceptable cortical bone margins, maintaining 3.75 ± 0.68 mm distance from cortex in narrow bone corridors and demonstrating accurate registration and surgical trajectory definition without breaching cortex.
Conclusions: The framework proposed in this work allows for multi-atlas based automatic planning of surgical trajectory without tracker or manual segmentation. The planning information can be further used to facilitate intraoperative guidance and post-operatively quality assurance in a manner consistent with surgical workflow.
Methods: CT images from 41 subjects (21 males, 20 females) were derived from the Cancer Imaging Archive (TCIA) and segmented using manual/semi-automatic methods. A statistical shape model was constructed and incorporated in an active shape model (ASM) registration framework for atlas-to-patient registration. Further, we introduce a registration method that exploits clusters in the underlying distribution to iteratively perform registrations after selecting a patient relevant cluster (sub-atlas) that represents similar shape characteristics to the image being registered. Experiments were performed to evaluate surface-to-surface and atlas-to patient registration algorithms using this clustered iterative model. Initial investigation of improved registration based on using similar shapes, was first explored through the use of gender as a categorical way of selecting a possible sub-atlas for registration.
Results: The RMSE surface-to-surface registration error (mean ± std) was reduced from (2.1 ± 0.2) mm when registering according to the entire atlas (N=40 members) to (1.8 ± 0.1) mm when registering within clusters based on similarity of principal components (N=20 members), showing improved accuracy (p<0.001) with fewer atlas members – an efficiency gained by virtue of the proposed approach. The atlas showed clear clusters in the first two principal components corresponding to gender, and the proposed method demonstrated improved accuracy when using ASM registration as well as when applied to a coherent-point drift (CPD) non-rigid deformable registration.
Conclusions: The proposed framework improved atlas-to-patient registration accuracy and increased the efficiency of statistical shape models (i.e., equivalent registration using fewer atlas members) by guiding member selection according to similarity in principal components.
Method: Registration of 2D US (slice) images is performed via the initialization obtained from a fast dictionary search that determines probe pose within a predefined set of pose configurations. 2D slices are extracted from a static 3D US (volume) image to construct a feature dictionary representing different probe poses. Haar features are computed in a fourlevel pyramid that transforms 2D image intensities to a 1D feature vector, which are in turn matched to the 2D target image. 3D-2D registration was performed with the Haar-based initialization with normalized cross-correlation as the metric and Powell’s method as the optimizer. Reduction to 1D feature vectors presents the potential for major gains in speed compared to registration of the 3D and 2D images directly. The method was validated in experiments conducted in a lumbar spine phantom and a cadaver specimen with known translations imparted by a computerized motion stage.
Results: The Haar feature matching method demonstrated initialization accuracy (mean ± std) = (1.9 ± 1.4) mm and (2.1 ± 1.2) mm in phantom and cadaver studies, respectively. The overall registration accuracy was (2.0 ± 1.3) mm and (1.7 ± 0.9) mm, and the initialization was a necessary and important step in the registration process.
Conclusions: The proposed image-based registration method demonstrated promising results for compensating motion of the US probe. This image-based solution could be an important step toward an entirely image-based, real-time registration method of 2D US to 3D US and pre-procedure MRI, eliminating hardware-based tracking systems in a manner more suitable to clinical workflow.
Methods: We apply a statistical framework that incorporates objective image quality factors such as spatial resolution and image noise combined with a statistical representation of anatomical clutter to predict the root-mean-squared error (RMSE) of transformation parameters in a rigid registration. Model predictions are compared to simulation studies in CT-to-CT slice registration using the cross-correlation (CC) similarity metric.
Results: RMSE predictions are shown to accurately model the impact of dose and soft-tissue clutter on measured RMSE performance. Further, these predictions reveal dose levels at which the registration becomes soft-tissue clutter limited, where further increase provides no improvement in registration performance.
Conclusions: Incorporating tissue deformation into a statistical registration model is an important step in understanding the limits of image registration performance and selecting pertinent registration methods for a particular registration task. The generalized noise model and RMSE analysis provide insight on how to optimize registration tasks with respect to image acquisition protocol (e.g., dose, reconstruction parameters) and registration method (e.g., level of blur).
Methods: To establish such a framework, we derived Cramer-Rao lower bounds (CRLB) for registration accuracy, revealing the underlying dependencies on image variance and gradient strength. The CRLB was analyzed as a function of image quality factors (in particular, dose) for various similarity metrics and compared to registration accuracy using CT images of an anthropomorphic head phantom at various simulated dose levels. Performance was evaluated in terms of root mean square error (RMSE) of the registration parameters.
Results: Analysis of the CRLB shows two primary dependencies: 1) noise variance (related to dose); and 2) sum of squared image gradients (related to spatial resolution and image content). Comparison of the measured RMSE to the CRLB showed that the best registration method, RMSE achieved the CRLB to within an efficiency factor of 0.21, and optimal estimators followed the predicted inverse proportionality between registration performance and radiation dose.
Conclusions: Analysis of the CRLB for image registration is an important step toward understanding and evaluating an intraoperative imaging system with respect to a registration task. While the CRLB is optimistic in absolute performance, it reveals a basis for relating the performance of registration estimators as a function of noise content and may be used to guide acquisition parameter selection (e.g., dose) for purposes of intraoperative registration.
View contact details
No SPIE Account? Create one
