Michael Kaus

Statistical validation of image segmentation quality based on a spatial overlap index.

RATIONALE AND OBJECTIVES To examine a statistical validation method based on the spatial overlap between two sets of segmentations of the same anatomy. MATERIALS AND METHODS The Dice similarity coefficient (DSC) was used as a statistical validation metric to evaluate the performance of both the reproducibility of manual segmentations and the spatial overlap accuracy of automated probabilistic fractional segmentation of MR images, illustrated on two clinical examples. Example 1: 10 consecutive cases of prostate brachytherapy patients underwent both preoperative 1.5T and intraoperative 0.5T MR imaging. For each case, 5 repeated manual segmentations of the prostate peripheral zone were performed separately on preoperative and on intraoperative images. Example 2: A semi-automated probabilistic fractional segmentation algorithm was applied to MR imaging of 9 cases with 3 types of brain tumors. DSC values were computed and logit-transformed values were compared in the mean with the analysis of variance (ANOVA). RESULTS Example 1: The mean DSCs of 0.883 (range, 0.876-0.893) with 1.5T preoperative MRI and 0.838 (range, 0.819-0.852) with 0.5T intraoperative MRI (P < .001) were within and at the margin of the range of good reproducibility, respectively. Example 2: Wide ranges of DSC were observed in brain tumor segmentations: Meningiomas (0.519-0.893), astrocytomas (0.487-0.972), and other mixed gliomas (0.490-0.899). CONCLUSION The DSC value is a simple and useful summary measure of spatial overlap, which can be applied to studies of reproducibility and accuracy in image segmentation. We observed generally satisfactory but variable validation results in two clinical applications. This metric may be adapted for similar validation tasks.

Adaptive, template moderated, spatially varying statistical classification.

A novel image segmentation algorithm was developed to allow the automatic segmentation of both normal and abnormal anatomy from medical images. The new algorithm is a form of spatially varying statistical classification, in which an explicit anatomical template is used to moderate the segmentation obtained by statistical classification. The algorithm consists of an iterated sequence of spatially varying classification and nonlinear registration, which forms an adaptive, template moderated (ATM), spatially varying statistical classification (SVC). Classification methods and nonlinear registration methods are often complementary, both in the tasks where they succeed and in the tasks where they fail. By integrating these approaches the new algorithm avoids many of the disadvantages of each approach alone while exploiting the combination. The ATM SVC algorithm was applied to several segmentation problems, involving different image contrast mechanisms and different locations in the body. Segmentation and validation experiments were carried out for problems involving the quantification of normal anatomy (MRI of brains of neonates) and pathology of various types (MRI of patients with multiple sclerosis, MRI of patients with brain tumors, MRI of patients with damaged knee cartilage). In each case, the ATM SVC algorithm provided a better segmentation than statistical classification or elastic matching alone.

Automated segmentation of the left ventricle in cardiac MRI

We present a fully automated deformable model technique for myocardium segmentation in 3D MRI. Loss of signal due to blood flow, partial volume effects and significant variation of surface grey value appearance make this a difficult problem. We integrate various sources of prior knowledge learned from annotated image data into a deformable model. Inter-individual shape variation is represented by a statistical point distribution model, and the spatial relationship of the epi- and endocardium is modeled by adapting two coupled triangular surface meshes. To robustly accommodate variation of grey value appearance around the myocardiac surface, a prior parametric spatially varying feature model is established by classification of grey value surface profiles. Quantitative validation of 121 3D MRI datasets in end-diastolic (end-systolic) phase demonstrates accuracy and robustness, with 2.45 mm (2.84 mm) mean deviation from manual segmentation.

Development and evaluation of an efficient approach to volumetric arc therapy planning

An efficient method for volumetric intensity modulated arc therapy (VMAT) planning was developed, where a single arc (360 degrees or less) is delivered under continuous variation of multileaf collimator (MLC) segments, dose rate, and gantry speed. Plans can be generated for any current linear accelerator that supports these degrees of freedom. MLC segments are derived from fluence maps at relatively coarsely sampled angular positions. The beam segments, dose rate, and gantry speed are then optimized using direct machine parameter optimization based on dose volume objectives and leaf motion constraints to minimize arc delivery time. The method can vary both dose rate and gantry speed or alternatively determine the optimal plan at constant dose rate and gantry speed. The method was used to retrospectively generate variable dose rate VMAT plans to ten patients (head and neck, prostate, brain, lung, and tonsil). In comparison to step-and-shoot intensity modulated radiation therapy, dosimetric plan quality was comparable or improved, estimated delivery times ranged from 70 to 160 s, and monitor units were consistently reduced in nine out of the ten cases by an average of approximately 6%. Optimization and final dose calculation took between 5 and 35 min depending on plan complexity.

Automated 3-D PDM construction from segmented images using deformable models

In recent years, several methods have been proposed for constructing statistical shape models to aid image analysis tasks by providing a priori knowledge. Examples include principal component analysis of manually or semiautomatically placed corresponding landmarks on the learning shapes [point distribution models (PDMs)], which is time consuming and subjective. However, automatically establishing surface correspondences continues to be a difficult problem. This paper presents a novel method for the automated construction of three-dimensional PDM from segmented images. Corresponding surface landmarks are established by adapting a triangulated learning shape to segmented volumetric images of the remaining shapes. The adaptation is based on a novel deformable model technique. We illustrate our approach using computed tomography data of the vertebra and the femur. We demonstrate that our method accurately represents and predicts shapes.

Shape Constrained Deformable Models for 3D Medical Image Segmentation

To improve the robustness of segmentation methods, more and more methods use prior knowledge. We present an approach which embeds an active shape model into an elastically deformable surface model, and combines the advantages of both approaches. The shape model constrains the flexibility of the surface mesh representing the deformable model and maintains an optimal distribution of mesh vertices. A specific external energy which attracts the deformable model to locally detected surfaces, reduces the danger that the mesh is trapped by false object boundaries. Examples are shown, and furthermore a validation study for the segmentation of vertebrae in CT images is presented. With the exception of a few problematic areas, the algorithm leads reliably to a very good overall segmentation.

Factors Influencing the Application Accuracy of Neuronavigation Systems

Objective: The overall accuracy of neuronavigation systems may be influenced by (1) the technical accuracy, (2) the registration process, (3) voxel size and/or distortion of image data and (4) intraoperative events. The aim of this study was to test the influence of the registration and imaging modality on the accuracy. Methods: A plexiglas phantom with 32 rods was taken for navigation targeting. Sixteen fiducials were attached to the surface of the phantom forming two different attachment patterns (clustered vs. diffusely scattered). This model was scanned by MRI and CT (1-mm slices). Registration was performed using different numbers and attachment patterns of the fiducials. Using CT or MRI, the localization error was measured in image space as the Euclidean distance between targets defined in image space and those detected in the physical space. Accuracy was measured with two commercial systems, the Zeiss MKM and the StealthStation. Results: The mean localization error varied between 1.59 ± 0.29 mm (MKM, 8 scattered fiducials, CT scanning) and 3.86 ± 2.19 mm (MKM, 4 clustered fiducials, MRI). The worst localization error was 9.5 mm (MKM). In case of an optimal registration, the 95th percentile for the localization error was 2.2 (MKM) and 2.75 mm (StealthStation). The imaging modality has only minor influence on the localization error, with CT increasing accuracy minimally. Both the fiducial number and the attachment pattern critically influence the localization error: 8 fiducials and a generalized attachment pattern increase the accuracy significantly. No correlation between the calculated registration accuracy and the measured localization accuracy was found. Conclusion: The application accuracy of different neuronavigation systems critically depends on the registration. The calculated registration accuracy provided by the system does not correspond to the localization error found in reality. The accuracy of frameless neuronavigation systems is comparable to that of classical frame-based stereotactic devices.

Segmentation of Meningiomas and Low Grade Gliomas in MRI

Computer assisted surgical planning and image guided technology have become increasingly used in neurosurgery. We have developed a system based on ATmC (Adaptive Template moderated Classification) for the automated segmentation of 3D MRI brain data sets of patients with brain tumors (meningiomas and low grade gliomas) into the skin, the brain, the ventricles and the tumor. In a validation study of 13 patients with brain tumors, the segmentation results of the automated method are compared to manual segmentations carried out by 4 independent trained human observers. It is shown that the automated method segments brain and tumor with accuracy comparable to the manual method and with improved reproducibility.

Semiautomated four-dimensional computed tomography segmentation using deformable models

The purpose of this work is to demonstrate a proof of feasibility of the application of a commercial prototype deformable model algorithm to the problem of delineation of anatomic structures on four-dimensional (4D) computed tomography (CT) image data sets. We acquired a 4D CT image data set of a patients thorax that consisted of three-dimensional (3D) image data sets from eight phases in the respiratory cycle. The contours of the right and left lungs, cord, heart, and esophagus were manually delineated on the end inspiration data set. An interactive deformable model algorithm, originally intended for deforming an atlas-based model surface to a 3D CT image data set, was applied in an automated fashion. Triangulations based on the contours generated on each phase were deformed to the CT data set on the succeeding phase to generate the contours on that phase. Deformation was propagated through the eight phases, and the contours obtained on the end inspiration data set were compared with the original manually delineated contours. Structures defined by high-density gradients, such as lungs, cord, and heart, were accurately reproduced, except in regions where other gradient boundaries may have confused the algorithm, such as near bronchi. The algorithm failed to accurately contour the esophagus, a soft-tissue structure completely surrounded by tissue of similar density, without manual interaction. This technique has the potential to facilitate contour delineation in 4D CT image data sets; and future evolution of the software is expected to improve the process.

Technical accuracy of a neuronavigation system measured with a high-precision mechanical micromanipulator

OBJECTIVE This study was designed to determine and evaluate the different system-inherent sources of erroneous target localization of a light-emitting diode (LED)-based neuronavigation system (StealthStation, Stealth Technologies, Boulder, CO). METHODS The localization accuracy was estimated by applying a high-precision mechanical micromanipulator to move and exactly locate (+/- 0.1 micron) the pointer at multiple positions in the physical three-dimensional space. The localization error was evaluated by calculating the spatial distance between the (known) LED positions and the LED coordinates measured by the neuronavigator. The results are based on a study of approximately 280,000 independent coordinate measurements. RESULTS The maximum localization error detected was 0.55 +/- 0.29 mm, with the z direction (distance to the camera array) being the most erroneous coordinate. Minimum localization error was found at a distance of 1400 mm from the central camera (optimal measurement position). Additional error due to 1) mechanical vibrations of the camera tripod (+/- 0.15 mm) and the reference frame (+/- 0.08 mm) and 2) extrapolation of the pointer tip position from the LED coordinates of at least +/- 0.12 mm were detected, leading to a total technical error of 0.55 +/- 0.64 mm. CONCLUSIONS Based on this technical accuracy analysis, a set of handling recommendations is proposed, leading to an improved localization accuracy. The localization error could be reduced by 0.3 +/- 0.15 mm by correct camera positioning (1400 mm distance) plus 0.15 mm by vibration-eliminating fixation of the camera. Correct handling of the probe during the operation may improve the accuracy by up to 0.1 mm.