Automatic MRI segmentation of para-pharyngeal fat pads using interactive visual feature space analysis for classification
© The Author(s) 2017
Received: 28 June 2016
Accepted: 11 January 2017
Published: 14 February 2017
Obstructive sleep apnea (OSA) is a public health problem. Detailed analysis of the para-pharyngeal fat pads can help us to understand the pathogenesis of OSA and may mediate the intervention of this sleeping disorder. A reliable and automatic para-pharyngeal fat pads segmentation technique plays a vital role in investigating larger data bases to identify the anatomic risk factors for the OSA.
Our research aims to develop a context-based automatic segmentation algorithm to delineate the fat pads from magnetic resonance images in a population-based study. Our segmentation pipeline involves texture analysis, connected component analysis, object-based image analysis, and supervised classification using an interactive visual analysis tool to segregate fat pads from other structures automatically.
We developed a fully automatic segmentation technique that does not need any user interaction to extract fat pads. Our algorithm is fast enough that we can apply it to population-based epidemiological studies that provide a large amount of data. We evaluated our approach qualitatively on thirty datasets and quantitatively against the ground truths of ten datasets resulting in an average of approximately 78% detected volume fraction and a 79% Dice coefficient, which is within the range of the inter-observer variation of manual segmentation results.
The suggested method produces sufficiently accurate results and has potential to be applied for the study of large data to understand the pathogenesis of the OSA syndrome.
KeywordsPara-pharyngeal fat pads segmentation Upper airway segmentation Interactive visual analysis tool Obstructive sleep apnea (OSA) Magnetic resonance imaging (MRI)
Obstructive sleep apnea (OSA) is commonly associated with obesity  and is considered as a public health problem affecting, at least, 2–4% of the middle-aged population . OSA is defined as a recurrent cessation of respiration during sleep associated with the obstruction of an upper airway . Obesity is one of the known risk factors for OSA and is reported in at least 50% of adults with this syndrome . Obesity is a special one among other common factors (e.g., gender, age, craniofacial features, pharyngeal abductor and dilator muscles) in that it is reversible and, therefore, may reduce OSA severity following weight loss .
The way by which obesity affects upper airway collapse during sleep is not completely understood and needs further studies. Either increased size of para-pharyngeal fat pads or reduced distance between left and right para-pharyngeal fat pads reduce pharyngeal airway volume . A smaller pharyngeal airway space has been identified in patients with OSA syndrome  and may result from enlargement of surrounding soft tissues, small distance between left and right fat pads, altered craniofacial morphology or a combination of all . It is clear, however, that upper airway anatomy is important in the pathogenesis of obstructive sleep apnea . In order to fully understand upper airway anatomy, we need to examine the volume of the airway and surrounding upper airway structures . Fat pads segmentation from MRI is of considerable importance in providing noninvasive information about them that helps radiologists to visualize and study their anatomy . For these reasons, fat pads segmentation from head MRI is an essential part of epidemiological study of OSA, and success of the study mostly depends on the segmentation accuracy and automation.
There is a number of publications dealing with OSA pathogenesis, symptoms, and treatment, but only few publications are available discussing automatic segmentation of pharyngeal airway and surrounding soft tissues. The lack of automatic segmentation technique is a great limitation in analyzing large datasets and epidemiological cohort studies. Schwab et al.  manually performed volumetric analysis of upper airways to analyze anatomic alterations. Liu et al.  proposed a semi-automatic framework for upper airway segmentation using fuzzy connectedness. Here, an operator defined a volume of interest and seed points in T1- and T2-weighted MR images manually. Ivanovska et al.  presented a semi-automatic segmentation pipeline for pharyngeal air column. Their approach consists of three steps: smoothing, thresholding, and 2D and 3D connected component analysis. Whereas the first two steps are rather common, the third step provides a set of general rules for extraction of the pharyngeal airway. Their method takes less than one minute to delineate the pharyngeal airway. The approach needs a small amount of interaction like defining the ending axial slice of oropharynx. Shahid et al.  described a fully automatic segmentation technique to segregate pharyngeal airway from head MRI. They performed pre-processing, object-based image analysis (OBIA), visual analysis of the feature space, and a refinement and extraction of the pharyngeal air column. They utilized the concept of dividing the image into small groups called “objects” having same basic features at voxel level. In this way, they explored additional features of different regions at the object level to classify them using interactive visual analysis tool. Their approach does not need any user interaction and consumes less than half of a minute to produce the complete segmentation results.
In summary, over the last few years we noticed major advancements in the field of upper airway segmentation from MRI, prompted by an increased use of MRI for soft tissues. Although there has been an effort towards developing an automatic segmentation technique for upper airway analysis, segmentation algorithms for fat pads have not been developed yet.
The first step towards the endeavor of an epidemiological cohort study is to develop a reliable automatic segmentation technique to segregate para-pharyngeal fat pads from the head MRI for the pathogenesis of OSA. In this paper, we describe a reliable automatic segmentation pipeline for the fat pads from MR images. In our segmentation methodology, we extract coarse objects as candidates, analyze their feature values using a multidimensional feature view, develop a supervised classifier using an interactive visual analysis tool, and finally refine the shape.
Some of the challenges are due to the employment of MRI instead of computed tomography (CT) for airways imaging. Although, the X-ray CT is used more commonly for airways imaging and considered to be the gold standard for this purpose, its use in an epidemiological cohort study for research purposes is ethically not justified. Here, MRI as a non-radiation based scanning method gains an increasing popularity. MR imaging provides a good contrast between soft tissue structures and is used to study brain and throat tissues . MRI makes it possible to conduct large-scale epidemiological cohort studies to analyze the role of fat pads in sleep apnea syndrome because it involves little or no risk to healthy volunteers. State-of-the-art segmentation methods for X-ray CT images are not applicable to the MRI, due to large differences in structure and tissue properties, and intensities.
Further challenges are imposed by the requirement of full automation of the segmentation technique. In practice, manual segmentation of fat pads is a tedious and time-consuming task. MRI scanners produce three dimensional images by generating multiple two-dimensional cross-sections (slices), and the radiologist or medical expert has to go through the 3D dataset slice by slice for selecting the most accurate contours from which the relevant regions or volumes are carefully generated . However, if the person drawing the contours is not a medical expert or radiologist, it will most likely generate poor segmentation results. The task of delineating the fat pads slice by slice sometimes limits the expert’s view and control, and generates jaggy regions. Needless to say, manual fat pads segmentation is also operator dependent and the selected volume is subject to large intra- observer variability.
Segmentation algorithms can be categorized into region-based, boundary-based, or model-based approaches. Region-based approaches may be categorized as being based on thresholding, clustering, or region growing. They do not work very well for inhomogeneous data sets. Boundary-based approaches only work in case of clearly extractable boundaries, while model-based approaches only work if the shape of the to-be-segmented object is well-defined or if a large training dataset in the form of an atlas is available. As, to our knowledge, no existing approaches exist that specifically target fat pad segmentation from MR images, we evaluated our method by comparing against commonly used, generally applicable segmentation algorithms. Since fat pads are not well-defined objects and we do not have large training datasets, we can only compare against region-based and boundary-based approaches. Prominent representatives of the three groups of region-based approaches are multiOtsu thresholding , fuzzy c-means clustering , seeded region growing . Worth mentioning boundary-based segmentation algorithms are watershed transform , and level sets segmentation .
The Study of Health in Pomerania (SHIP) , a cohort study conducted in Northeast Germany, provides us the test datasets. In the SHIP study, more than 3 400 participants aged 20 to 89 years participated. The test datasets are isotropic, three dimensional T1-weighted head MRI with an image size of 176×256×176 voxels, where each voxel has an isotropic volume of 1 mm 3. These head MR images contain only the upper airways and its surrounding tissues which are considered to contain the most essential information to study the OSA syndrome in the frame of an epidemiological study. Upper airway imaging is a best way to examine the pharyngeal airway and its surrounding soft tissues. Magnetic resonance scanning is considered an ideal modality for soft tissues, since it can accurately quantify pharyngeal fat pads in axial, sagittal and coronal planes without radiation .
Thirty individual head MRI datasets were randomly selected for our experiments and tests. Among the 30 datasets, 10 were selected randomly for generating ground truths by manual segmentation by two observers under the supervision of an experienced radiologist.
Para-pharyngeal fat pads segmentation
Retropalatal oropharynx extraction
Texture analysis describes the characteristics of image regions by their texture content. Texture analysis can be useful when a region in an image is more characterized by its texture than by gray levels. We attempt to quantify the intuitive qualities such as smooth, homogeneous, or random as a function of the intensity variations. Homogeneous regions have no variation, smooth regions are slowly changing and as such have small local variations, while random parts exhibit largest variations. Hence, standard statistical measures (standard deviation, entropy, homogeneity, average, variance etc.) are used to characterize the texture of a region in our image. They indicate the local variability of voxel gray levels in the image.
We applied texture analysis to discard the axial slices having relatively smooth regions and low contrast in the process of obtaining initial 3D objects for fat pads. In low contrast axial slices, we have region leakage problem for the fat pads which causes strong artifacts in the shape of fat pads. These artifacts lead to uncertain variations in the shape features of the fat pad candidates. Omission of the low contrast axial slices during the seed selection process does not have any major effect on the segmentation process because we just need a good starting axial slice as a starting point for our segmentation algorithm. In this way, we achieve more consistency in terms of shape features.
Connected component analysis
After having reduced the number of slices to be analyzed in the texture analysis, a median filter of kernel size 3×3×5 is applied to minimize the salt and pepper noise  from the head MRI. Then, an anisotropic diffusion filter  is applied to smooth the image while keeping the object boundaries intact.
Having bright regions on axial slices, we connect them using connected component analysis in order to build 3D fat pad candidates. We use a 26-neighborhood relationship in a 3D connected component analysis. As a result to the connected component analysis on bright regions, we obtain a considerably small number of 3D regions that serve as fat pad candidates.
Object-based image analysis
State-of-the-art segmentation algorithms mainly rely on a voxel level to draw the contours of organs. However, when operating exclusively on a voxel level, only a limited number of features and properties can be utilized to separate the fat pads from surrounding structures and tissues. Thus, we overcome the limitations of voxel-based processing by examining the image as atomic regions (called objects) in their local semantic context . Atomic regions introduce additional features that are not available on a voxel level such as intensity statistics, orientation vectors, and shape descriptors.
Within our work, we define the image regions as possible candidates and extract the descriptive features of objects using the methodology of object-based image analysis introduced by Homeyer et al. . They assign to objects a set of features regarding their shape, orientation profile, and intensity statistics. We define features that we consider potentially helpful in describing fat pads and discriminating them from other candidates. Hence, we define features for 3D objects that describe their intensity profile, their relative and absolute position, and their shape.
To describe the intensity profile, we compute different statistical measures of intensity, including minimum, maximum, and average values as well as median, and upper and lower quartiles.
Sheet-like objects have a flatness close to one (for infinitely wide, infinitely thin sheet). Wire-like structures have flatness close to zero, just like spherical objects.
All these descriptive features form a multidimensional feature space, where each object corresponds to a multi-dimensional point in that feature space. We normalized the numerical values of the features to have a uniform range. Now, the task is to explore the multidimensional feature space of useful features in order to determine the decisive contribution of features.
Visualization of feature space
Hence, we visualize the descriptive features collectively, which form a multidimensional feature space, to find suitable weights of feature values for separating our candidates. Dimensionality reduction method is utilized to map multidimensional data to a low-dimensional 2D space for a visual exploration of the data distribution. Principal component analysis (PCA)  or multi-dimensional scaling  are the most commonly used dimensionality reduction methods. The selection of an optimal method generally depends upon preservation of distances, computational efficiency, and cluster preservation . Since we aim to develop an intuitive interactive system, linear projection is the best available choice, as they are economical in terms of computational cost and do not introduce non-linear distortion, which makes them intuitive.
The linear projection of the n-dimensional feature space Q onto a 2-dimensional projection space U can be implemented by using a 2×n projection matrix P. The columns of the projection matrix represent the coordinates of the basis vector images mapped from Q to U space. Since the origin of the projection space coincides with the origin of feature space, the columns of linear projection matrix are considered as axes of a star-coordinates widget.
Our interactive visual analysis system allows us to manipulate the projection by varying the positions of the end-points of basis vectors. The projection matrix is then recomputed to project the feature space. Moving the end-point of a basis vector to the origin cancels out the effect of selected feature by leading a corresponding column of zero elements in the projection matrix.
We visualize the training datasets, where we know the class of each sample, to find a configuration of the star-coordinates widget which can visually decouple the fat pads class from the background class. We started with a default configuration of the star-coordinates widget where the uniformly distributed axes of the widget represent the all possible features. We use two colors for the projected samples as shown in Fig. 6, where each color corresponds to a different class. We determine the influence of each individual feature on projected space by varying the length and angle of its corresponding basis vector in the star-coordinates widget. We eliminate the feature from the default configuration of the widget if it affects both classes in a same way and does not help us to visually decouple both classes. In this way, we are left with few efficacious features which distinguish both classes vividly.
On average, for our data as described above there are twelve samples in each dataset while only two samples are classified as (left and right) fat pads. To evaluate the performance of our classifier in practice, holdout validation method is employed for the thirty datasets. We partitioned our datasets into two sets of ten datasets for training and twenty datasets for testing. We built the classifier using the training set and evaluated it using the testing set. We run multiple rounds of handout validation, and the validation results are averaged over the rounds to reduce the variability. We achieved excellent results for our supervised fat pad classifier which detected all fat pad samples accurately in our datasets with no false positives and no false negatives. The perfectly designed classifier laid the foundation for the complete automation of our segmentation algorithm.
Our classified fat pad objects include some false surrounding voxels due to intensity inhomogeneity and low contrast of MR images. More precisely, the connected component step leaked into surrounding regions, which we need to cut off in this last step. We used only intensity clustering for generating the fat pad candidates in the connected component analysis step to keep the algorithm simple and fast which created extra artifacts. Therefore, we need to remove these redundant voxels and regions from fat pad objects in a refinement step. This refinement is performed on a slice-wise basis in axial direction.
We start refining the fat pad object from the axial slice where fat pads tend to be convex. From our a priori domain knowledge, fat pads have close-to-convex shape in general. After finding the best axial slice, we measure the statistics (center of gravity, area, average intensity, and intensity variance) of the fat pad object for processing in other axial slices. First, we move in the downward and then in the upward direction to process all axial slices of fat pad objects. For each axial slice, a specific area around the center of gravity of the adjacent processed axial slice is selected which is still comparatively much smaller than the entire axial slice. Then iterative MultiOtsu thresholding  is applied to separate fat pads for statistical analysis. In statistical analysis, different features (center of gravity, area, average intensity, and intensity variance) of delineated regions are computed, and regions like blood vessels and mandible tissues are discarded depending upon a defined criterion. Only those regions are considered as fat pad regions which do not show a variation of more than 30% from the adjacent slice features. It helps us to remove blood vessels (more bright in intensity) and mandible tissues (far away from the center of fat pad) from fat pad objects.
After applying all these refinement steps and procedures, we are left with fat pad objects where all redundant and extra regions have been excluded. In the end, we perform a morphological closing operation  to fill the holes and cavities and to remove the boundary irregularities.
Evaluation and comparison of our segmentation results with observer2 against observer1
We also measured the inter-observer reliability  in terms of kappa (κ) statistics for our two manually segmented ground truths. First, we calculated observed agreement (O) equal to 0.98 for our ground truths. The observed agreement has the disadvantage that some agreement would exist even if both observers simply guessed the result. We found κ=0.8 in our case which indicates a substantial agreement according to the guidelines of Landis and Koch .
Comparison of our segmentation results with generally available segmentation methods
It can be seen that no single algorithm manages to achieve accurate results. Seeded region growing algorithm and level sets segmentation method need user interaction to select the seed points for the segmentation of the object. The seeded region growing method faces severe region leakage problem in low contrast slices of the data set. MultiOtsu thresholding and fuzzy c-means algorithm could not produce better results when applied on complete image. Moreover, multiOtsu thresholding is expensive in terms of time. The watershed transform also faces limitations like over-segmentation and under-segmentation for different slices of a dataset, and could not produce sufficient results. For all the algorithms, preprocessing is applied on the image before the application of the segmentation method. The region growing algorithm faces the problem of region leakage due to the low contrast of the data sets. Other algorithms fail due to the inhomogeneous nature of fat pads. As a comparison to them, our proposed method produces desirable results, does not involve any user interaction and takes less than one minute for each subject.
A fast and fully automatic segmentation technique has been presented to segregate para-pharyngeal fat pads from T1-weighted head MR images. The complete pipeline includes texture analysis, connected component analysis, object-based image analysis, and supervised classifier using interactive visual analysis tool. The algorithm is fast, as the whole processing consumes less than a minute for a single MRI dataset. The method has been tested on thirty random datasets. The proposed algorithm generates sufficiently reliable and accurate results and has potential to be applied for the study of large data in epidemiological studies such as SHIP to understand the pathogenesis of the obstructive sleep apnea syndrome. Our algorithm could also be readily integrated in clinical routine.
Our idea of defining a supervised classifier using the interactive visual analysis tool can be employed to develop a classifier for the segmentation of other organs and tissues. The core idea of our segmentation algorithm can be applied to solve many medical segmentation problems. Initially generating coarse objects as fat pad candidates using intensity clustering improves the speed of our algorithm and then, refining the fat pad object increases the accuracy of our segmentation method.
As future work, we plan to extract different features and properties of segmented fat pads and visualize them to understand the pathogenesis of OSA. Furthermore, we are also interested in studying the role of fat pads in the treatment of OSA.
The proposed segmentation algorithm delineates para-pharyngeal fat pads automatically without any user interaction. The algorithm is fast enough to be applied in a cohort study to investigate OSA. Different image features like fat pads’ volume, and distance between left and right fat pads are measured from the segmented fat pads. These image features will be further analyzed to determine their associations with OSA.
False positive volume fraction
Magnetic resonance imaging
Object-based image analysis
Obstructive sleep apnea
Principal component analysis
Study of health in Pomerania
True positive volume fraction
We are grateful to the Visualization and Computer Graphics Laboratory research group of Jacobs University, Bremen, Germany, for valuable suggestions and remarks. Finally, we are really thankful to our medical experts’ team and other members from the University of Greifswald, Germany, for their collaboration in this project.
DAAD provided the funding to carry out this research to find the pathogenesis of obstructive sleep apnea public health problem. SHIP cohort study is a part of the Research Network of Community Medicine, which is funded by the Federal State of Mecklenburg Vorpommern and the University Medicine Greifswald, Germany.
Availability of data and materials
The dataset used in the article is available on request to the Ernst-Moritz-Arndt-University of Greifswald Germany.
MS designed and implemented the main segmentation pipeline and drafted the manuscript. TC implemented the OBIA section of the pipeline. TI carried out the study of pharyngeal structures and soft tissues. VM helped to visually analyze the multi-dimensional feature space to design the classifier. HV acquired and analyzed the MR images. LL conceived of the study, and participated in its design and coordination and helped to draft the manuscript. All authors read and approved the final manuscript.
The authors declare that they have no competing interests.
Consent for publication
Ethics approval and consent to participate
Ethics approval with waiver of patient informed consent is obtained from the Project Manager of the Institute for Community Medicine, Ernst-Moritz-Arndt-University of Greifswald Germany.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- Sutherland K, Lee RWW, Phillips CL, Dungan G, Yee BJ, Magnussen JS, Grunstein RR, Cistulli PA. Effect of weight loss on upper airway size and facial fat in men with obstructive sleep apnoea. 2011; 66(9):797–803. doi:10.1136/thx.2010.151613.
- Pack AI. Sleep Apnea: Pathogenesis, Diagnosis and Treatment. New York: CRC Press; 2002.View ArticleGoogle Scholar
- Lowe AA, Fleetham JA. Two- and three-dimensional analyses of tongue, airway, and soft palate size. Atlas of the Difficult Airway In: Norton ML, Brown ACD, editors. Mosby-Year Book, St. Louis: 1991. p. 74–82.
- Young T, Peppard PE, Taheri S. Excess weight and sleep-disordered breathing. J Appl Physiol. 2005; 99(4):1592–9. doi:10.1152/japplphysiol.00587.2005.View ArticlePubMedGoogle Scholar
- Grunstein RR, Stenlöf K, Hedner JA, Peltonen M, Karason K, Sjöström L. Two year reduction in sleep apnea symptoms and associated diabetes incidence after weight loss in severe obesity. Sleep. 2007; 30(6):703.PubMedPubMed CentralGoogle Scholar
- Schwartz AR, Patil SP, Laffan AM, Polotsky V, Schneider H, Smith PL. Obesity and obstructive sleep apnea: pathogenic mechanisms and therapeutic approaches. Proc Am Thorac Soc. 2008; 5(2):185–92.View ArticlePubMedPubMed CentralGoogle Scholar
- Schwab RJ, Pasirstein M, Pierson R, Mackley A, Hachadoorian R, Arens R, Maislin G, Pack AI. Identification of upper airway anatomic risk factors for obstructive sleep apnea with volumetric magnetic resonance imaging. Am J Respir Crit Care Med. 2003; 168(5):522–30.View ArticlePubMedGoogle Scholar
- Watanabe T, Isono S, Tanaka A, Tanzawa H, Nishino T. Contribution of body habitus and craniofacial characteristics to segmental closing pressures of the passive pharynx in patients with sleep-disordered breathing. Am J Respir Crit Care Med. 2002; 165(2):260–5.View ArticlePubMedGoogle Scholar
- Ivanovska T, Dober J, Laqua R, Hegenscheid K, Völzke H. Pharynx segmentation from MRI data for analysis of sleep related disoders In: Bebis G, editor. Advances in Visual Computing. 2nd edn. New York: Springer: 2013. p. 20–9.Google Scholar
- Shahid MLUR, Chitiboi T, Ivanovska T, Molchanov V, Völzke H, Hahn HK, Linsen L. Automatic pharynx segmentation from MRI data for obstructive sleep apnea analysis. In: Proceedings of the 10th International Conference on Computer Vision Theory and Applications: 2015. p. 599–608. doi:10.5220/0005315905990608.
- Ivanovska T, Laqua R, Shahid ML, Linsen L, Hegenscheid K, Völzke H. Automatic pharynx segmentation from MRI data for analysis of sleep related disorders. Int J Artif Intell Tools. 2015; 24(04):1550018. doi:10.1142/s0218213015500189.View ArticleGoogle Scholar
- Liu J, Udupa JK, Odhnera D, McDonough JM, Arens R. System for upper airway segmentation and measurement with mr imaging and fuzzy connectedness. Acad Radiol. 2003; 10(1):13–24.View ArticlePubMedGoogle Scholar
- Ivanovska T, Buttke E, Laqua R, Völzke H, Beule A. Automatic trachea segmentation and evaluation from MRI data using intensity pre-clustering and graph cuts. In: Image and Signal Processing and Analysis (ISPA), 2011 7th International Symposium On. Dubrovnik: 2011. p. 513–8. IEEE.
- Gordillo N, Montseny E, Sobrevilla P. State of the art survey on MRI brain tumor segmentation. Magn Reson Imaging. 2013; 31(8):1426–38. doi:10.1016/j.mri.2013.05.002.View ArticlePubMedGoogle Scholar
- Udupa JK, LaBlanc VR, Schmidt H, Imielinska C, Saha PK, Grevera GJ, Zhuge Y, Currie L, Molholt P, Jin Y. Methodology for evaluating image-segmentation algorithms. In: Medical Imaging 2002. San Diego: International Society for Optics and Photonics: 2002. p. 266–77.Google Scholar
- Liao PS, Chen TS, Chung PC. A fast algorithm for multilevel thresholding. J Inf Sci Eng. 2001; 17(5):713–27.Google Scholar
- Bezdek JC, Ehrlich R, Full W. Fcm: The fuzzy c-means clustering algorithm. Comput Geosci. 1984; 10(2–3):191–203.View ArticleGoogle Scholar
- Adams R, Bischof L. Seeded region growing. IEEE Trans Pattern Anal Mach Intell. 1994; 16(6):641–7. doi:10.1109/34.295913.View ArticleGoogle Scholar
- Beucher S, Meyer F. The morphological approach to segmentation: the watershed transformation. Opt Engineering-New York-Marcel Dekker Inc. 1992; 34:433–3.Google Scholar
- Paragios N, Deriche R. Geodesic active contours and level sets for the detection and tracking of moving objects. IEEE Trans Pattern Anal Mach Intell. 2000; 22(3):266–80.View ArticleGoogle Scholar
- Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet. 1986; 327(8476):307–10.View ArticleGoogle Scholar
- Völzke H, Alte D, Schmidt CO, Radke D, Lorbeer R, Friedrich N, Aumann N, Lau K, Piontek M, Born G, et al. Cohort profile: the study of health in pomerania. Int J Epidemiol. 2011;294–307.
- Welch KC, Foster GD, Ritter CT, Wadden TA, Arens R, Maislin G, Schwab RJ. A novel volumetric magnetic resonance imaging paradigm to study upper airway anatomy. Sleep, New York. 2002; 25(5):532–42.Google Scholar
- Gonzalez RC, Woods RE. Digital image processing: Pearson prentice hall. Upper Saddle River, NJ; 2008.
- Perona P, Malik J. Scale-space and edge detection using anisotropic diffusion. Pattern Anal Mach Intell IEEE Trans. 1990; 12(7):629–39.View ArticleGoogle Scholar
- Shapiro LG, Stockman GC. Computer Vision. New Jersey, Prentice-Hall; 2001, pp. 279–325. ISBN 0-13-030796-3.
- Homeyer A, Schwier M, Hahn HK. A generic concept for object-based image analysis. In: VISAPP (2): 2010. p. 530–3.
- Yang M, Kpalma K, Ronsin J, et al. A survey of shape feature extraction techniques. Pattern Recogn. 2008;43–90.
- Burger W, Burge MJ. Principles of Digital Image Processing. London: Springer; 2009.Google Scholar
- Jolliffe I. Principal Component Analysis. New Jersey: Wiley Online Library; 2005.View ArticleGoogle Scholar
- Borg I, Groenen PJ. Modern Multidimensional Scaling: Theory and Applications. New York: Springer; 2005.Google Scholar
- Molchanov V, Chitiboi T, Linsen L. Visual analysis of medical image segmentation feature space for interactive supervised classification. In: Eurographics Workshop on Visual Computing for Biology and Medicine: 2015. p. 11–19. doi:10.2312/vcbm.20151204.
- Rousseeuw PJ. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J Comput Appl Math. 1987; 20:53–65.View ArticleGoogle Scholar
- Vincent L, Soille P. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Trans Pattern Anal Mach Intell. 1991; 6:583–98.View ArticleGoogle Scholar
- Soille P. Morphological Image Processing: Principles and Applications. New York: Springer; 1999.View ArticleGoogle Scholar
- Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977; 33(1):159–74.View ArticlePubMedGoogle Scholar