Benchmark 1 (Anatomy1) Results

The following PDF files contain the segmentation results and the landmark detection results.

The information that participants have provided about their techniques is below (linked to the tables in the PDF documents by the abbreviations in square brackets):

[SJ]
Title: A New Rule-Based Approach For Body Multi-Organ Automatic Segmentation in CT Scans
Authors: Assaf B. Spanier and Leo Joskowicz, School of Eng. and Computer Science, The Hebrew Univ. of Jerusalem, Israel University
Abstract: We describe a new generic method for the automatic rule-based segmentation of multiple organs from 3D CT scans in the VISCERAL challenge. The rules determine the order in which the organs are isolated and detected from simple to difficult. Following the isolation of the body, the breathing system organs are segmented: the trachea and the left and right lungs. Next, the organs with high blood content are segmented: the spleen, liver and the left and right kidneys. The segmentation of each organ itself is then performed in three steps: 1) definition of the Organ Binary Inclusive Region Of Interest (BI-ROI) from the target organ intensity values; 2) identication of the organ's Largest Axial Cross Section Slice (LACSS), the slice where the organ has the largest axial area, and; 3) organ segmentation by 3D region growing to adjacent slices from the BI-ROI and the LACSS while preserving smoothness and curvature constraints between two adjacent slices. The key advantages of our method are that it uses the organs segmentations, known locations, and anatomical context to guide the automated the segmentation process and that the organ segmentation itself follows a generic three-step process. Our experimental results on the 7 CT training dataset of VISCERAL Challenge Anatomy1 Benchmark yield a Dice volume overlap similarity of 96.4 for the left lung, 96.6 for the right lung, 79.1 for the trachea, 89.2 for the spleen, 92.8 for the left kidney, 92.0 for the liver, 90.0 for the right kidney. For the 11 CT scans test datasets, the Dice scores are 84.8 for the left lung, 97.5 for the right lung, 78.5 for the trachea, 69.0 for the spleen, 63.1 for the left kidney, 74.7 for the liver, 63.1 for the right kidney. Our method was ranked first in the segmentation of the lung among 6 methods participated in the challenge.

[HJ]
Title: Automatic Liver Segmentation using Multiple Prior Knowledge Models and Free-Form Deformation
Authors: Cheng Huang and Fucang Jia, Research Lab. for Medical Imaging and Digital Surgery, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences.
Abstract: An automatic and robust coarse-to-fine liver image segmentation method is proposed. The workflow can be divided into four steps: liver localization, shape model fitting, appearance profile fitting and free-form deformation. For liver localization, an atlas image based rigid registration with correlation coefficient histogram metric is used to detect liver region of interest (ROI), an AdaBoost classifier is trained via multiple low level image features including intensity, gradient and context features, then liver probability map is generated. For generation of prior models, shape correspondence is established by deformation registration of manual segmented liver image to the atlas image, then Statismo toolkit is used to build the shape and appearance model, both intensity and gradient profile information inside, outside and at the true liver boundary are sampled and KNN classifier is used to construct appearance model. For model fitting, registration of the distance map image of AdaBoost classified liver mask and the point sets of the mean shape model is used, the mesh vertexes of deformed shape are moved to liver boundary location more accurately with major shape variation constraints. For free-form deformation, simplex mesh deformable model with gradient and intensity KNN classifier as external force is used to conquer local specific variation of liver shape. Fifty manually segmented datasets, fourteen VISCERAL challenge CT and CTce training datasets are used to train prior models.

[JM]
Title: Multi-atlas based segmentation
Authors: Oscar Alfonso Jiménez del Toro, Henning Müller, University of Applied Sciences Western Switzerland (HES-SO) and University and University Hospitals of Geneva, Switzerland
Abstract: Multi-atlas based segmentation is an approach that requires little or no interaction from the user. It has been evaluated with high accuracy and consistent reproducibility in different anatomical structures. In this method, multiple atlases identify the location of one or more structures in the patient volume. The label volumes of the atlases are transformed taking the coordinate transformation obtained from image registration of each atlas to the target volume. A stochastic gradient descent optimisation is performed for the desired metric during the process. Since multiple structures are segmentation targets in the VISCERAL benchmark, a hierarchical selection of the registrations improves the segmentations of all the structures. A global affine registration is followed by individual affine registrations using a local binary mask to enforce the spatial correlation of each anatomical structure separately. These masks are obtained from the morphological dilation of the output labels of the different atlases registered in the previous step. The method is repeated for the non-rigid registration. The registrations of the bigger structures are used as a starting point for the closely related smaller structures, which are harder to segment. Most of the registrations of the initial bigger structures (liver, lungs, urinary bladder) will be reused in the method which makes it faster than segmenting each structure individually from the start. Also the creation of regions-of-interest with the local masks speeds up the image registrations and improves the output estimations. The labels from the different atlases are fused using a per-voxel majority voting threshold in a single label volume that provides a final estimate location of the structures in the target volume. The images are downsampled in all but the final step to increase even more the speed of the algorithm. The method was tested with contrast-enhanced computed tomography images and 10 different anatomical structures: liver, spleen, kidneys, lungs, urinary bladder, trachea, lumbar vertebra and gallbladder. It can be then applied to any modality and any anatomical structure using a relatively small training set.

[GG]
Title: Segmentation and Landmark Localization Based on Multiple Atlases
Authors: Tobias Gass and Orcun Goksel, Computer Vision Lab, ETH Zurich, Switzerland
Abstract: In this work, we present multi-atlas based techniques for both segmentation and landmark detection.
We focus on modality and anatomy independent techniques to be applied in a wide range of image modalities, in contrast to methods customized to a specific anatomy or modality.
For segmentation, we use label propagation from several atlases to a target image via a Markov random field (MRF) based registration method, followed by label fusion by majority voting weighted by local cross-correlations. The registrations are computed on the full images as provided by the challenge, thus requiring only one registration per image pair independent of the number of segmentation labels. The multi-atlas segmentation fusion is also performed on all labels simultaneously, assigning the anatomical label with the highest weighted vote to each voxel. Local correlations are computed efficiently using a technique based on convolutions with Gaussian kernels. This is very generic method, also demonstrated by being the only submission applied to all anatomical structures in all modalities. It has comparable results to more specific segmentation algorithms in the VISCERAL challenge.
For landmark localization, we use a consensus-based fusion of location estimates from several atlases identified by a customized template-matching approach. Here, each atlas votes for candidate landmark locations, which are found by matching a fixed region of interest around the atlas landmark to a sliding window in a fixed region of the test image. For this matching, cross-correlation is employed which can be computed very efficiently in the frequency domain for such sliding windows. The top-ranked candidate locations of all atlases are combined using the median operator. Similar to the segmentation approach presented above, our landmark detection algorithm is applied to all images of all modalities in the VISCERAL challenge, yielding satisfactory results.

[DMB]
Title: Classification Forests for Landmark Detection
Authors: Mohammad A. Dabbah, Sean Murphy, Erin Beveridge, Daniel Wyeth, Ian Poole, Toshiba Medical Visualization Systems, Ltd.
Abstract: We use a voxel-level trained solution based on classification forests. Datasets are rotated to be aligned in DICOM Patient Coordinates, then downsampled to an isotropic resolution of 4mm per voxel with Gaussian smoothing to avoid aliasing effects. Features are simple densities in Hounsfield units at chosen random offsets to each voxel.
Each classification tree is trained using 40 datasets randomly selected from the 369 available. Additionally, random samples are taken for a background class throughout the volume. A number of background samples is taken equal to the total number of landmark samples in each dataset. At each node of the decision tree, 2,500 randomly selected features are searched for greatest information gain, the threshold being selected by an efficient incremental algorithm. Each leaf node stores the proportion by class of weighted training samples reaching that node. A classification forest of 80 trees is trained each with different randomly selected datasets.
At detection time, each downscaled voxel is passed down each tree in the forest, the resulting normalized likelihoods being averaged across each tree in the forest. For each landmark, the voxel with the greatest normalized likelihood for that landmark is selected as the potential detection point. Brent interpolation (ie fitting a quadratic) is used to deliver a sub-voxel result.
An issue often overlooked or unreported in other published work is how to deal with voxels for which some features cannot be measured either at training time or detection time, because the randomly selected offset references a voxel outside the dataset or in padding. Padding occurs in CT datasets outside the cylindrical acquisition region of the volume, and possibly elsewhere. The problem will very likely occur for voxels close to the edge of the volume, within the 52mm maximum feature offset. Our approach is to treat these unmeasurable values as missing features} in the manner described by Quinlan. In brief, when applying a decision rule at a node which involves a missing (unmeasurable) feature, that voxel is sent both ways down the tree, with modified weights. Quinlan discusses various ways of determining these weights, and we have experimented with these, settling on simply assigning the sample 50/50 to each branch, in both training and detection. Like Quinlan we also found it beneficial during training to scale the information gain for a candidate feature by the proportion of samples for which the feature was measurable (ie not missing).

[W]
Title:
Automatic multi-organ segmentation using fast model based level set method and hierarchical shape priors
Authors: Chunliang Wang and Örjan Smedby, Center for Medical Image Science and Visualization (CMIV), Linköping University, and Department of Radiology and Department of Medical and Health Sciences, Linköping University, Linköping, Sweden.
Abstract: An automatic multi-organ segmentation pipeline is presented. The segmentation starts with stripping the body of skin and subcutaneous fat using threshold-based level-set methods. After registering the image to be processed against a standard subject picked from the training datasets, a series of model-based level set segmentation operations is carried out guided by hierarchical shape priors. The hierarchical shape priors are organized according to the anatom¬ical hierarchy of the human body, starting with ventral cavity, and then divided into thoracic cavity and abdominopelvic cavity. The third level contains the individual organs such as lungs, liver and kidneys. Statistical shape models of structures and organs are created by registering the binary segmentation masks of individual organs against the picked standard subject. The position of a lower-level structure relative to a upper-level structure is computed by registering the statistical mean shape against a trust zone created by thresholding the probability atlas of that anatomical structure in the upper-level structure’s space. The segmentation is performed in a top-down fashion, where major structures are segmented first, and their location information is then passed down to the lower level to initialize the segmentation, while boundary information from higher-level structures also constrains the segmentation of the lower-level structures. A threshold-based speed function is used to drive the level set segmentation while the thresholds are iteratively updated based on statistical analysis of preliminary segmentation results. The proposed method was combined with a novel coherent propagating level-set method, which is capable to detect local convergence and skip calculation in those parts, thereby significantly reducing computation time. In our prelim¬inary experiments, the proposed method yielded a Dice coefficient around 90% for most major thoracic and abdominal organs in both contrast-enhanced CT and non-enhanced datasets, while the average running time for segmenting 10 organs was about 10 minutes.

[K]
Title:
3D Multiobject Segmentation via Clustering and Graph Cut Using Shortest-Path Constraints for Spatial Relations and Atlas-Based Shape Priors
Authors: Razmig Kechichian, Michel Desvignes (Gipsa-lab, Grenoble INP; France), Sébastien Valette (Creatis, INSA de Lyon; France)
Abstract: Our automatic multiple organ segmentation method is based on a multi-label Graph Cut optimization approach which uses prior information of organ spatial relationships and shape. The former is derived from shortest-path pairwise constraints defined on a graph model of structure adjacency relations and the latter is represented by probabilistic organ atlases learned from a training dataset. The pairwise prior in particular is a piecewise-constant model incurring multiple levels of penalization capturing the spatial configuration of structures in multiobject segmentation. Organ atlases are mapped to the image via a hierarchical image registration method based on SURF keypoints and are additionally used to derive image intensity statistics automatically. Prior models and intensity statistics are then introduced in a joint centroidal Voronoi image clustering and Graph Cut multiobject segmentation framework. The clustering approach we take to simplify images prior to segmentation strikes a good balance between boundary adaptivity and cluster compactness criteria furthermore allowing to control the trade-off. Compared to a direct application of segmentation on voxels, the clustering step improves the overall runtime and memory footprint of the segmentation process up to an order of magnitude virtually without compromising the quality of the result. An initial implementation has allowed to evaluate the method on the contrast-enhanced CT subset of the VISCERAL dataset.