We propose a learning based approach to perform automatic segmentation of the left ventricle in 3D cardiac ultrasound images. The segmentation contour is estimated through the use of a variant of Hough forests whose object localization capabilities are coupled with a patch-wise, appearance driven, contour estimation strategy. The performance of the proposed method is evaluated on a dataset of 30 images acquired
from 15 patients using different equipment and settings.

Left Ventricle Segmentation in Cardiac Ultrasound Using Hough-Forests With Implicit Shape and Appearance Priors

As part of the CETUS challenge, we present a multi-atlas segmentation framework to delineate the left-ventricle endocardium in echocardiographic images. To increase the robustness of the registration step, we introduce a speckle reduction step and a new shape representation based on sparse coding and manifold approximation in dictionary
space. The shape representation, unlike intensity values, provides consistent shape information across different images. The validation results on the test set show that registration based on our shape representation significantly improves the performance of multi-atlas segmentation compared to intensity based registration. To our knowledge it is the first time that multi-atlas segmentation achieves state-of-the-art results for echocardiographic images.

Learning Shape Representations for Multi-Atlas Endocardium Segmentation in 3D Echo Images

A method for real-time automatic tracking of the left ventricle (LV) in 3D ultrasound is presented. A mesh model of the LV is deformed using mean value coordinates enabling large variations. Kalman filtering and edge detection is used to track the mesh in each frame. The method is evaluated using the framework of the Challenge on Endocardial Three-dimensional Ultrasound Segmentation (CETUS). The results show that the method is able to robustly track the LV in all sequences with a mean mesh difference of about 2.5 mm

Real-time Tracking of the Left Ventricle in 3D Ultrasound Using Kalman Filter and Mean Value Coordinates

Segmentation of 3D echocardiograms (3DEs) is still a challenging task due to the low signal-to-noise ratio, the limited field of view, and typical ultrasound artifacts. We propose to segment the left ventricular endocardial surface by using Active Appearance Models (AAMs). Separate end-diastolic (ED) and end-systolic (ES) AAMs were built from presegmented 3DEs of the CETUS training data and 25 previously acquired 3DEs, imaged using various 3DE equipment. The AAMs fully automatically segmented the 15 training sets in a leave-one-out cross validation, comparing two training populations and various initialization strategies. All segmentations took about 15 seconds per patient. The comparison on the CETUS training data shows that the AAM benefits from additional training data and more accurate initialization. The results on the CETUS training and testing data confirm good ED and ES segmentation accuracy on multi-center, multi-vendor, multi-pathology data, and corresponding EF estimation. Selection from different initialization strategies, based on the minimal residual error, and propagation of detected ED contours to initialize ES detection, contributed to more accurate segmentations in this heterogeneous population.

Segmentation of Multi-Center 3D Left Ventricular Echocardiograms by Active Appearance Models

In this paper, we propose a semi-automatic method for left ventricle segmentation. The proposed method utilizes a multi-scale quadrature filter method to enhance the 3D volume, followed by a model-based level set method to segment the endocardial surface of the left ventricle. The phase map from the quadrature filters is also used to weight the influence of contour points when updating the statistical model.

Model-based left ventricle segmentation in 3D ultrasound using phase image

Malignant gliomas are highly heterogeneous brain tumors with complex an-
isotropic growth patterns and occult invasion. Computational modeling of cell migration and proliferation has been subject of intensive research aiming at a deeper understanding of the tumor biology and the ability to predict growth and thus improve therapy. However, current modeling techniques follow a generative approach and make strong assumptions about underlying mechanisms. The tumor is so far treated as homogeneous entity with behavioral parameters extrapolated from previous longitudinal image information.
We present a novel way of approaching this problem by employing data driven, discrim-
inative modeling techniques that learn relevant features from observed growth patterns and are able to make meaningful predictions solely on basis of local and regional tissue characteristics at one given point in time.
We demonstrate superior performance of the proposed discriminative method (DICE score
83) compared to the state of the art generative approach (DICE score 72) on six patients and a total of nine different time intervals. Our approach can help estimating occult invasion as well as it can advance our understanding of the tumor biology and lead to valuable predictions of tumor growth patterns that could guide and improve radio therapy.

Brain Tumor Progression Modeling - A Data Driven Approach

In radiation treatment (RT) planning medical doctors need to consider a variety of information sources for anatomical and functional target volume delineation. The validation and inspection of the defined target volumes and the resulting RT plan is a complex task, especially in the presence of moving target areas as it is the case for tumors of the chest and the upper abdomen. A 4D-PET/CT visualization system may become a helpful tool for validating RT plans. We define major requirements such a visualization system should fulfill to provide medical doctors with the necessary visual information to validate tumor delineations, and review the dose distribution of a RT plan. We present an implementation of such a system, and present qualitative results of its applications for a lung cancer patient.

Visualization of 4D-PET/CT, Target Volumes and Dose Distribution: Applications in Radiotherapy Planning

Purpose: Recent years have seen growing interest in adaptive radiation therapy (RT), but the existing software tools are not ideal for research use, as they are either expensive and closed proprietary applications or free open-source packages with limited scope, extensibility, reliability, or user support. To address these limitations, we have developed SlicerRT, an open-source toolkit for RT research, providing fast computations and user-friendly interface for researchers. It is a medium for RT researchers to integrate their workflows in, assists clinical translation of experimental approaches, and enables comparative testing.

Methods: SlicerRT builds on the widely used medical image visualization and analysis platform, 3D Slicer (www.slicer.org). In addition to the tools that 3D Slicer offers for visualization, registration and segmentation, the toolkit provides functions specifically designed for RT research. Established development processes as well as testing and validation environment ensure software quality. Standardized software development mechanisms of 3D Slicer were applied for documentation, distribution, and user support. The design and architecture of the toolkit was outlined in [1]. SlicerRT has evolved to contain the most widely used features in the field (see www.slicerrt.org), which have been defined through consensus discussions with a large pool of RT researchers.

Results: Numerous modules have been developed for RT-specific features such as DICOM-RT data import and export, as well as dose analysis tools including dose volume histogram, dose accumulation, dose comparison, and isodose line and surface generation. The toolkit also includes contour analysis modules for handling various contour representations, contour morphology, and contour comparison. Advanced registration tools are provided by the included Plastimatch library [2]. The main new features developed in the last year are RT plan and RT image import and the subject hierarchy module, which arranges the loaded data to a structure familiar to clinicians, while providing advanced automation features through the underlying extendible mechanism. The toolkit is being used by several research groups to support and evaluate adaptive radiation therapy workflows. SlicerRT is available for download through the Extension Manager component of 3D Slicer 4.3 or later (see instructions on our website).

Future work: As SlicerRT has matured to contain most of the planned functions, our focus has shifted from feature development to usability and stability. The subject hierarchy module, and its features such as bulk deformation of studies embodies our efforts toward usability, while the integration of the subject hierarchy and the advanced contour handling mechanism into the 3D Slicer core points towards stability, also benefiting the whole 3D Slicer community. There is ongoing work to evaluate and potentially standardize the way the 2D contours are rasterized into 3D volumes, thus mitigating the great variability between the different solutions for this problem. Current funding ensures continuous development for several years, so SlicerRT is expected to develop further.

Acknowledgement: This work was in part funded by Cancer Care Ontario through Applied Cancer Research Unit and Research Chair in Cancer Imaging grants, and the Ontario Consortium for Adaptive Invention in Radiation Oncology (OCAIRO).


Improvements in SlicerRT, the radiation therapy research toolkit for 3D Slicer

In radiotherapy (RT) for tumor delineation and diagnostics, complementary information of multi-modal images is used. Using high ionizing radiation, the accuracy of registered volume data is crucial; therefore a reliable and robust evaluation method for registered images is needed in clinical practice. Multi-modal image registration aligns images from different modalities like computed tomography (CT) and magnetic resonance imaging (MRI) or cone beam computed tomography (CBCT) into one common frame of reference. The gold standard validation methods are visual inspection by radiation oncology experts and fiducial-based evaluation. However, visual inspection is a qualitative measure with a range of 2-6 mm inaccuracy, it is time consuming and prone to errors. The fiducial-based evaluation is an invasive method when fiducial markers are fixated to bone or implanted in organs. Therefore, in clinical practice a robust non-invasive automated method is needed to validate registration of multi-modal images. 
The aim of this study is to introduce and validate an automatic landmark-based accuracy measure for multi-modal image rigid registration using feature descriptors. A porcine dataset with fixed fiducial markers was used to compare our accuracy measure with the target registration error of fiducial markers.In addition, the robustness of our evaluation method was tested on multi-vendor database consisted of 10 brain and 20 lung cases comparing the automatic landmark accuracy measure based on feature descriptors with manual landmark based evaluation. 
An automatic, non-invasive method based on feature descriptors for accuracy evaluation of multi-modal rigid registration was introduced. The method can be used to provide accuracy information slice-by slice on CT, CBCT and CT, MR-T1, -T2 weighted, MR-T1 contrast enhanced (ce) multi-modal images.

Automatic accuracy measurement for multi-modal rigid registration using feature descriptors

In this work, we try to develop a fast converging method for
segmentation assisted deformable registration. The segmentation step
consists of a piece-wise constant Mumford-Shah energy model while reg-
istration is driven by the sum of squared distances of both initial images
and segmented mask with a diffusion regularization. In order to solve
this energy minimization problem, a second order Gauss-Newton opti-
mization method is used. For the numerical experiments we used CT
data sets from the EMPIRE10 challenge. In this preliminary study, we
show high accuracy of our algorithm.
