Frank Deinzer

Viewpoint Selection – Planning Optimal Sequences of Views for Object Recognition

In the past decades most object recognition systems were based on passive approaches. But in the last few years a lot of research was done in the field of active object recognition. In this context there are several unique problems to be solved, like the fusion of several views and the selection of the best next viewpoint.

2D/3D image registration on the GPU

We present a method that performs a rigid 2D/3D image registration efficiently on the Graphical Processing Unit (GPU). As one main contribution of this paper, we propose an efficient method for generating realistic DRRs that are visually similar to x-ray images. Therefore, we model some of the electronic post-processes of current x-ray C-arm-systems. As another main contribution, the GPU is used to compute eight intensity-based similarity measures between the DRR and the x-ray image in parallel. A combination of these eight similarity measures is used as a new similarity measure for the optimization. We evaluated the performance and the precision of our 2D/3D image registration algorithm using two phantom models. Compared to a CPU + GPU algorithm, which calculates the similarity measures on the CPU, our GPU algorithm is between three and six times faster. In contrast to single similarity measures, our new similarity measure achieved precise and robust registration results for both phantom models.

New technologies to reduce pediatric radiation doses

X-ray dose reduction in pediatrics is particularly important because babies and children are very sensitive to radiation exposure. We present new developments to further decrease pediatric patient dose. With the help of an advanced exposure control, a constant image quality can be maintained for all patient sizes, leading to dose savings for babies and children of up to 30%. Because objects of interest are quite small and the speed of motion is high in pediatric patients, short pulse widths down to 4 ms are important to reduce motion blurring artifacts. Further, a new noise-reduction algorithm is presented that detects and processes signal and noise in different frequency bands, generating smooth images without contrast loss. Finally, we introduce a super-resolution technique: two or more medical images, which are shifted against each other in a subpixel region, are combined to resolve structures smaller than the size of a single pixel. Advanced exposure control, short exposure times, noise reduction and super-resolution provide improved image quality, which can also be invested to save radiation exposure. All in all, the tools presented here offer a large potential to minimize the deterministic and stochastic risks of radiation exposure.

Automatic Robust Medical Image Registration Using a New Democratic Vector Optimization Approach with Multiple Measures

The registration of various data is a challenging task in medical image processing and a highly frequented area of research. Most of the published approaches tend to fail sporadically on different data sets. This happens due to two major problems. First, local optimization strategies induce a high risk when optimizing nonconvex functions. Second, similarity measures might fail if they are not suitable for the data. Thus, researchers began to combine multiple measures by weighted sums. In this paper, we show severe limitations of such summation approaches. We address both issues by a gradient-based vector optimization algorithm that uses multiple similarity measures. It gathers context information from the iteration process to detect and suppress failing measures. The new approach is evaluated by experiments from the field of 2D-3D registration. Besides its generic character with respect to arbitrary data, the main benefit is a highly robust iteration behavior, where even very poor initial guesses of the transform result in good solutions.

Extended global optimization strategy for rigid 2D/3D image registration

Rigid 2D/3D image registration is a common strategy in medical image processing. In this paper we present an extended global optimization strategy for a rigid 2D/3D image registration that consists of three components: a combination of a global and a local optimizer, a combination of a multi-scale and a multi-resolution approach, and a combination of an in-plane and an out-of-plane registration. The global optimizer Adaptive Random Search is used to provide several coarse registration results on a low resolution level that are refined by the local optimizer Best Neighbor on a higher resolution level. We evaluate the performance and the precision of our registration algorithm using two phantom models. We could approve that all three components of our optimization strategy lead to an significant improvement of the registration.

Integrated Viewpoint Fusion and Viewpoint Selection for Optimal Object Recognition

In the past decades, most object recognition systems were based on passive approaches. But in the last few years a lot of research was done in the eld of active object recognition, that is selectively moving a sensor/camera around a considered object in order to acquire as much information about it as possible. In this paper we present an active object recognition approach that solves the problem of choosing optimal views (viewpoint selection) and iteratively fuses the gained information for an optimal 3D object recognition (viewpoint fusion) in an integrated manner. Therefore, we apply a method for the fusion of multiple views with respect to the knowledge about the assumed camera movement between them. For viewpoint selection we formally dene the choice of additional views as an optimization problem. We show how to use reinforcement learning for this purpose and perform a training without user interaction. In this context we focus on the modeling of continuous states, continuous, one-dimensional actions and supporting rewards for an optimized recognition of real objects. The experimental results show that our combined viewpoint selection and viewpoint fusion approach is able to signicantly improve the recognition rates compared to passive object recognition with randomly chosen views.

Classifier Independent Viewpoint Selection for 3-D Object Recognition

3-D object recognition has been tackled by passive approaches in the past. This means that based on one image a decision for a certain class and pose must be made or the image must be rejected. This neglects the fact that some other views might exist, which allow for a more reliable Classification. This situation especially arises if certain views of or between objects are ambiguous.

Esophagus Segmentation by Spatially-Constrained Shape Interpolation

The segmentation and visualization of the esophagus is helpful during planing and performing atrial ablation therapy to avoid esophageal injury. Only very few studies have addressed this segmentation problem which is challenging because the esophagus has a low contrast in medical images. In this work we present a technique to segment the esophagus based on the interpolation of Fourier descriptors of manually drawn contours. The interpolation is spatially-constrained using a dedicated correction term to avoid intersections with the convex shaped left atrial posterior wall. Our technique is fast, modality independent and achieves optimal results if at least three input contours are used. We validated our technique successfully with patient data and discuss the use of our technique in the clinical workflow.

Temporal Estimation of the 3d Guide-Wire Position Using 2d X-ray Images

We present a method for realtime online 3d reconstruction of a guide-wire or catheter using 2d X-ray images, which do not have to be recorded from different viewpoints. No special catheters or sensors are needed. Given a 3d patient data set and the projection parameters, we use recursive probability density propagation to estimate a probability distribution of the current positions of guide-wire parts. Based on this distribution, we extract the optimal guide-wire position using regularization techniques. We describe the guide-wire position by a uniform cubic B-spline. Experiments on simulated and phantom data demonstrate the high accuracy and robustness of our approach.

On Fusion of Multiple Views for Active Object Recognition

In the last few years the research in 3-D object recognition has focused more and more on active approaches. In contrast to the passive approaches of the past decades where a decision is based on one image, active techniques use more than one image from different viewpoints for the classification and localization of an object. In this context several tasks have to be solved. First, how to choose the different viewpoint and how to fusion the multiple views. In this paper we present an approach for the fusion of multiple views within a continuous pose space. We formally define the fusion as a recursive density propagation problem and we show how to use the Condensation algorithm for solving it. The experimental results show that this approach is well suited for the fusion of multiple views in active object recognition.