Anatomy2 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). More detailed descriptions of some of these methods will appear in the proceedings of the MICCAI Medical Computer Vision 2014 workshop.

Title: Rule-Based Ventral Cavity Multi-Organ Automatic Segmentation in CT Scans
Authors: Assaf B. Spanier, Leo Joskowicz, School of Eng. and Computer Science, The Hebrew Univ. of Jerusalem, Israel University
Abstract: We describe a new method for the automatic segmentation of multiple organs of the ventral cavity in CT scans. The method is based on a set of rules that determine the order in which the organs are isolated and segmented, from the simplest one to the most difficult one. First, the air-containing organs are segmented: the trachea and the lungs. Then, the organs with high blood content: the spleen, the kidneys and the liver, are segmented. Each organ is individually segmented with a generic four-step procedure that consists of: 1) ROI Identication; 2) Thresholding; 3) 2D-seed identification; 4) Slice region growing with clustering classication algorithms. Our method is unique in that it uses the same generic segmentation approach for all organs and in that it relies on the segmentation diculty of organs to guide the segmentation process. Experimental results on 20 CT scans of the VISCERAL Anatomy2 Challenge training datasets yield a Dice volume overlap similarity score of 79.5 for the trachea, 97.4 and 97.6 for the left and right lungs, 89.2 for the spleen, 92.8 and 89.2, respectively, for the left and right kidney, and 83.5 for the liver. For the 10 CT scans test datasets, the Dice scores are 85.1, 97.0, 96.8, 82.2, 82.9 and 87.0, respectively. Our method achieved an overall DICE volume overlap similarity score of 88.5. For the segmentation of air containing organs (i.e. lungs and trachea) in CTce we ranked first among other methods that participated in the challenge.

Title: Automatic Liver Segmentation using Statistical Prior Models and Free-form Deformation
Authors: Xuhui Li, Cheng Huang, Fucang Jia, Zongmin Li, Chihua Fang, Yingfang Fan. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China University of Petroleum and Southern Medical University, China
Abstract: An automatic and robust coarse-to-fine liver image segmentation method is proposed. Multiple prior knowledge models are built to implement liver localization and segmentation: voxel-based AdaBoost classifier is trained to localize liver position robustly, shape and appearance models are constructed to fit liver these models to original CT volume. Free-form deformation is incorporated to improve the models’ ability of refining liver boundary. The method was submitted to VISCERAL big data challenge, and had been tested on IBSI 2014 challenge datasets and the result demonstrates that the proposed method is accurate and efficient.

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 anatomical 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.

Title: Automatic segmentation of abdominal organs in a framework based on Active Appearance Models
Author: Graham Vincent and Richard Haworth, Imorphics Ltd., Manchester, UK
Abstract: We present a single fully automatic model based framework for segmenting the aorta, kidneys, liver, lungs and the psoas major muscles in thoracic contrast enhanced and wide field of view CT images. These anatomies were chosen as a representative subset of the structures in the full VISCERAL database, in order to contribute to the VISCERAL Anatomy2 benchmark. The models were built from approximately 35 manually segmented examples provided by the VISCERAL organisers. The segmentation framework is generic and has been successfully used for segmentation in CT and/or MR for bone and soft tissue in the hands, knee, hip and ankle, spine, for prostate and for sub-cortical structures in the brain.
The framework is based on Active Appearance Models (AAM). High quality correspondences for the AAM are generated using a Minimum Description Length group-wise image registration method. A hierarchical multi-start optimisation scheme is used to robustly match the AAMs to new images which fits low density low resolution models followed by increasingly detailed and high resolution models. The model result helps define a region of uncertainty in a narrow halo around the model boundary. The voxels in this region are assigned a probability of belonging to the structure using a non-linear regression function, trained using a PAC-learning method. Finally, the probability image can be turned into a binary image by thresholding at p=0.5 for comparison with binary golden data.
The results on the VISCERAL data are very good across all anatomies. In DICE overlap scores, the method ranked top for the aorta, left and right psoas major muscles and left kidney for both wide beam and thoracic images, and the right lung for contrast enhanced images.
Note: This was the only method to produce fuzzy segmentations, with grey levels indicating the probability of membership of the organ. For this reason, two sets of results were produced. "-f" contains the metrics calculated directly on the fuzzy images; "-b" contains the metrics calculated on binary images produced from the fuzzy images with a threshold at 0.5.

Automatic 3D Multiorgan Segmentation via Clustering and Graph Cut Using Spatial Relations and Hierarchically-Registered Atlases
Authors: Razmig Kechichian, Michel Desvignes (Gipsa-lab, Grenoble INP; France), Sébastien Valette, Michael Sdika (Creatis, INSA de Lyon; France)
Abstract: We propose a generic method for automatic multiple-organ segmentation based on a multilabel Graph Cut optimization approach which uses location likelihood of organs and prior information of spatial relationships between them. The latter is derived from shortest-path constraints defined on the adjacency graph of structures and the former is defined by probabilistic atlases learned from a training dataset. Organ atlases are mapped to the image by a fast (2+1)D hierarchical registration method based on SURF keypoints. Registered atlases are furthermore used to derive organ intensity likelihoods. Prior and likelihood models are then introduced in a joint centroidal Voronoi image clustering and Graph Cut multiobject segmentation framework. Qualitative and quantitative evaluation has been performed on contrast-enhanced CT images from the Visceral Benchmark dataset.

Multi-Atlas Segmentation and Landmark Localization in Images with Large Field of View
Authors: Tobias Gass, Gabor Szekely, Orcun Goksel, ETH Zürich, Switzerland
Abstract: In this work, we present multi-atlas based techniques for both segmentation and landmark detection in images with large field-of-view (FOV). Such images can provide important insight in the anatomical structure of the human body, but are challenging to deal with since the localization search space for landmarks and organs, in addition to the raw amount of data, is large. In many studies, segmentation
and localization techniques are developed specifically for an individual target anatomy or image modality. This can leave a substantial amount of the potential of large FOV images untapped, as the co-localization and shape variability of organs are neglected. We thus focus on modality and anatomy independent techniques to be applied to a wide range of input images. For segmentation, we propagate the multi-organ label maps from several atlases to a target image via a large FOV Markov random field (MRF) based non-rigid registration method. The propagated labels are then fused in the target domain using similarity-weighted
majority voting. For landmark localization, we use a consensus based fusion of location estimates from several atlases identified by a template-matching approach. We present our results in the IEEE ISBI 2014 VISCERAL challenge as well as VISCERAL Anatomy1 and Anatomy2 benchmarks.

Hierarchic Multi-atlas Based Segmentation for Anatomical Structures: Evaluation in the VISCERAL Anatomy Benchmarks
Authors: Oscar Jimenez del Toro, Henning Müller, University of Applied Sciences Western Switzerland
Abstract: Computer-based medical image analysis is often initialized with the localization of anatomical structures in clinical scans. Many methods have been proposed for segmenting single and multiple anatomical structures. However, it is uncommon to compare dierent approaches with the same test set, namely a publicly available one. The comparison of these methods objectively denes the advantages and limitations for each method. A hierarchic multi-atlas based segmentation approach was proposed for the segmentation of multiple anatomical structures in computed tomography scans. The method relies on an anatomical hierarchy that exploits the inherent spatial and anatomical variability of medical images using image registration techniques. It was submitted and tested in the VISCERAL project Anatomy benchmarks. In this paper, the results are analyzed and compared to the results of the other segmentation methods submitted in the benchmark. Various anatomical structures in both unenhanced and contrast-enhanced CT scans resulted in the highest overlap with the proposed method compared to the other evaluated approaches. Although the method was trained with a small trainingset it generated accurate output segmentations for liver, lungs and other anatomical structures.

Title: 3D Landmark detection with Histograms of Oriented Gradients
Authors: Dominic Mai, Olaf Ronneberger, University of Freiburg, Germany
Abstract: We present an approach to landmark detection in volumetric images based on the popular Histograms of Oriented Gradients Descriptor (HOG) and linear support vector machines. We rigidly align the positive training examples and compute 3D HOG descriptors with 20 orientation bins for the patch surrounding the landmark location. As negative examples we randomly sample patches from the 3D volume that do not contain the sought landmark. We train a linear support vector machine on the kernel matrix of our training set to cope with the high dimensionality of the data. At test time we use a sliding window approach that we compute efficiently as a convolution in Fourier space.