Heinrich Schulz

A new method for robust organ positioning in CT images

Robust initialization is essential for any successful segmentation process of medical images. For CT images, initialization is challenging because quality, appearance, content, and field-of-view of the images are highly variable. Furthermore, high execution speed is desirable, whereas the user tolerance to errors is low in clinical applications. We present a new method for efficient and robust positioning of organs in CT images. It is based on a novel probabilistic atlas that, given a tissue type, describes the probability density of the random variable spatial location. Random sampling is then employed to select the most informative points for matching. We present results on pelvic and abdominal images acquired for radiotherapy planning.

Validation of automatic landmark identification for atlas-based segmentation for radiation treatment planning of the head-and-neck region

Manual contouring of target volumes and organs at risk in radiation therapy is extremely time-consuming, in particular for treating the head-and-neck area, where a single patient treatment plan can take several hours to contour. As radiation treatment delivery moves towards adaptive treatment, the need for more efficient segmentation techniques will increase. We are developing a method for automatic model-based segmentation of the head and neck. This process can be broken down into three main steps: i) automatic landmark identification in the image dataset of interest, ii) automatic landmark-based initialization of deformable surface models to the patient image dataset, and iii) adaptation of the deformable models to the patient-specific anatomical boundaries of interest. In this paper, we focus on the validation of the first step of this method, quantifying the results of our automatic landmark identification method. We use an image atlas formed by applying thin-plate spline (TPS) interpolation to ten atlas datasets, using 27 manually identified landmarks in each atlas/training dataset. The principal variation modes returned by principal component analysis (PCA) of the landmark positions were used by an automatic registration algorithm, which sought the corresponding landmarks in the clinical dataset of interest using a controlled random search algorithm. Applying a run time of 60 seconds to the random search, a root mean square (rms) distance to the ground-truth landmark position of 9.5 ± 0.6 mm was calculated for the identified landmarks. Automatic segmentation of the brain, mandible and brain stem, using the detected landmarks, is demonstrated.

SU-E-J-141: Comparison of Dose Calculation On Automatically Generated MRBased ED Maps and Corresponding Patient CT for Clinical Prostate EBRT Plans

PURPOSE To analyze the effect of computing radiation dose on automatically generated MR-based simulated CT images compared to true patient CTs. METHODS Six prostate cancer patients received a regular planning CT for RT planning as well as a conventional 3D fast-field dual-echo scan on a Philips 3.0T Achieva, adding approximately 2 min of scan time to the clinical protocol. Simulated CTs (simCT) where synthesized by assigning known average CT values to the tissue classes air, water, fat, cortical and cancellous bone. For this, Dixon reconstruction of the nearly out-of-phase (echo 1) and in-phase images (echo 2) allowed for water and fat classification. Model based bone segmentation was performed on a combination of the DIXON images. A subsequent automatic threshold divides into cortical and cancellous bone. For validation, the simCT was registered to the true CT and clinical treatment plans were re-computed on the simCT in pinnacle3 . To differentiate effects related to the 5 tissue classes and changes in the patient anatomy not compensated by rigid registration, we also calculate the dose on a stratified CT, where HU values are sorted in to the same 5 tissue classes as the simCT. RESULTS Dose and volume parameters on PTV and risk organs as used for the clinical approval were compared. All deviations are below 1.1%, except the anal sphincter mean dose, which is at most 2.2%, but well below clinical acceptance threshold. Average deviations are below 0.4% for PTV and risk organs and 1.3% for the anal sphincter. The deviations of the stratifiedCT are in the same range as for the simCT. All plans would have passed clinical acceptance thresholds on the simulated CT images. CONCLUSION This study demonstrated the clinical usability of MR based dose calculation with the presented Dixon acquisition and subsequent fully automatic image processing. N. Schadewaldt, H. Schulz, M. Helle and S. Renisch are employed by Phlips Technologie Innovative Techonologies, a subsidiary of Royal Philips NV.

Local motion analysis in 4D lung CT using fast groupwise registration

Automatic tools for the propagation of contours and surfaces in time series of CT images (4DCT) are necessary e.g. for dose calculation in radio-therapy-planning. The state of the art method is the registration of 3-D frames of the sequence on a static reference volume or the registration of consecutive volume pairs. Based on the computed vector fields, the contours and surfaces are then propagated through the entire 4-D series. However, the registrations frequently suffer from common image or gating artefacts in the time series. To overcome the problem, we propose a groupwise 4-D registration approach combined with a motion model to perform the registration of a sequence of images.

Motion compensation and plane tracking for kinematic MR-Imaging

The acquisition of time series of 3D MR images is becoming feasible nowadays, which enables the assessment of bone and soft tissue in normal and abnormal joint motion. Fast two-dimensional (2D) scanning of moving joints may also provide high temporal resolution but is limited to a single, predefined slice. Acquiring 3D time series has the advantage that after the acquisition image processing and visualization techniques can be used to reformat the images to any orientation and to reduce the through-plane motion and undesired gross motion superimposed on the relevant joint movement. In this publication, we first review such post-processing techniques for retrospective tracking of viewing planes according to a single moving rigid body (e.g. bone). Then, we present new numerical schemes for optimally tracking viewing planes according to the movement of multiple structures to compensate for their through- as well as in-plane motion. These structures can be specified in an interactive viewing program, and the motion compensated movies can be updated and displayed in real-time. The post-processing algorithms require a 4D motion-field estimation which also can be utilized to interpolate intermediate images to present the final movies in smooth cine-loops and to significantly improve the visual perceptibility of complex joint movement.

Annotation-free probabilistic atlas learning for robust anatomy detection in CT images

A fully automatic method generating a whole body atlas from CT images is presented. The atlas serves as a reference space for annotations. It is based on a large collection of partially overlapping medical images and a registration scheme. The atlas itself consists of probabilistic tissue type maps and can represent anatomical variations. The registration scheme is based on an entropy-like measure of these maps and is robust with respect to field-of-view variations. In contrast to other atlas generation methods, which typically rely on a sufficiently large set of annotations on training cases, the presented method requires only the images. An iterative refinement strategy is used to automatically stitch the images to build the atlas. Affine registration of unseen CT images to the probabilistic atlas can be used to transfer reference annotations, e.g. organ models for segmentation initialization or reference bounding boxes for field-of-view selection. The robustness and generality of the method is shown using a three-fold cross-validation of the registration on a set of 316 CT images of unknown content and large anatomical variability. As an example, 17 organs are annotated in the atlas reference space and their localization in the test images is evaluated. The method yields a recall (sensitivity), specificity and precision of at least 96% and thus performs excellent in comparison to competitors.

Nearest neighbor 3D segmentation with context features

Automated and fast multi-label segmentation of medical images is challenging and clinically important. This paper builds upon a supervised machine learning framework that uses training data sets with dense organ annotations and vantage point trees to classify voxels in unseen images based on similarity of binary feature vectors extracted from the data. Without explicit model knowledge, the algorithm is applicable to different modalities and organs, and achieves high accuracy. The method is successfully tested on 70 abdominal CT and 42 pelvic MR images. With respect to ground truth, an average Dice overlap score of 0.76 for the CT segmentation of liver, spleen and kidneys is achieved. The mean score for the MR delineation of bladder, bones, prostate and rectum is 0.65. Additionally, we benchmark several variations of the main components of the method and reduce the computation time by up to 47% without significant loss of accuracy. The segmentation results are – for a nearest neighbor method – surprisingly accurate, robust as well as data and time efficient.

Precise anatomy localization in CT data by an improved probabilistic tissue type atlas

Automated interpretation of CT scans is an important, clinically relevant area as the number of such scans is increasing rapidly and the interpretation is time consuming. Anatomy localization is an important prerequisite for any such interpretation task. This can be done by image-to-atlas registration, where the atlas serves as a reference space for annotations such as organ probability maps. Tissue type based atlases allow fast and robust processing of arbitrary CT scans. Here we present two methods which significantly improve organ localization based on tissue types. A first problem is the definition of tissue types, which until now is done heuristically based on experience. We present a method to determine suitable tissue types from sample images automatically. A second problem is the restriction of the transformation space: all prior approaches use global affine maps. We present a hierarchical strategy to refine this global affine map. For each organ or region of interest a localized tissue type atlas is computed and used for a subsequent local affine registration step. A three-fold cross validation on 311 CT images with different fields-of-view demonstrates a reduction of the organ localization error by 33%.

Bladder segmentation in MR images with watershed segmentation and graph cut algorithm

Prostate and cervix cancer diagnosis and treatment planning that is based on MR images benefit from superior soft tissue contrast compared to CT images. For these images an automatic delineation of the prostate or cervix and the organs at risk such as the bladder is highly desirable. This paper describes a method for bladder segmentation that is based on a watershed transform on high image gradient values and gray value valleys together with the classification of watershed regions into bladder contents and tissue by a graph cut algorithm. The obtained results are superior if compared to a simple region-after-region classification.

TU‐G‐134‐03: MR‐Only‐Based Generation of Electron Density Maps and Digitally Reconstructed Radiographs of the Pelvis

PURPOSE To generate electron density (ED) maps and digitally reconstructed radiographs (DRRs) of the pelvis based only on magnetic resonance imaging (MRI). METHODS A conventional 3D fast-field dual-echo sequence was used to acquire image data of 8 healthy subjects with a Philips 3.0T Ingenia TX system in approximately 1:50 min scan time per volunteer. Dixon reconstruction of the nearly out-of-phase (echo 1) and in-phase images (echo 2) allowed for water and fat classification. A bone-enhanced image was generated by automatically thresholding the noise level of the in-phase image with subsequent background removal. ED maps were then produced by assigning known bulk electron densities to the classified bone and tissue fractions. A bone probability atlas derived from CT data was registered to the ED map in order to filter out misclassified voxels. Finally, DRRs were reconstructed from bone-enhanced images as well as from ED maps. RESULTS The proposed MRI sequence with subsequent Dixon reconstruction and probabilistic filtering makes it possible to classify cortical bone, soft tissue and adipose tissue in the pelvis and yields ED maps and corresponding DRRs. Bowel content was misclassified as cortical bone or air and compromised the segmentation in some slices as well as in the DRRs. Automatic probabilistic atlas filtering can significantly reduce artifacts induced by bowel content without affecting pelvic bone structures markedly. In total, the artifact/bone fraction dropped from 1.7 before filtering to 0.2 after filtering. The average reduction of artifact volume is 87%, and the average bone preservation is 99%. Remaining artifacts are spatially close to the true bone in areas of positive bone probability. CONCLUSION This study demonstrated the feasibility of generating realistic ED maps of the pelvis by using MRI only. The method has the potential to become an essential component of emerging applications such as MR-only-based radiation therapy planning. All authors have the following relevant financial interest or relationship to disclose with regard to the subject matter of this presentation: Company name: Philips Research; Type of relationship: Employee.