Automatic joint classification and segmentation of whole cell 3D images

doi:10.1016/j.patcog.2008.08.009

Pattern Recognition

Volume 42, Issue 6, June 2009, Pages 1067-1079

https://doi.org/10.1016/j.patcog.2008.08.009 Get rights and content

Abstract

We present a machine learning tool for automatic texton-based joint classification and segmentation of mitochondria in MNT-1 cells imaged using ion-abrasion scanning electron microscopy (IA-SEM). For diagnosing signatures that may be unique to cellular states such as cancer, automatic tools with minimal user intervention need to be developed for analysis and mining of high-throughput data from these large volume data sets (typically $\sim 2 GB / cell$ ). Challenges for such a tool in 3D electron microscopy arise due to low contrast and signal-to-noise ratios (SNR) inherent to biological imaging. Our approach is based on block-wise classification of images into a trained list of regions. Given manually labeled images, our goal is to learn models that can localize novel instances of the regions in test datasets. Since datasets obtained using electron microscopes are intrinsically noisy, we improve the SNR of the data for automatic segmentation by implementing a 2D texture-preserving filter on each slice of the 3D dataset. We investigate texton-based region features in this work. Classification is performed by $k$ -nearest neighbor ( $k$ -NN) classifier, support vector machines (SVMs), adaptive boosting (AdaBoost) and histogram matching using a NN classifier. In addition, we study the computational complexity vs. segmentation accuracy tradeoff of these classifiers. Segmentation results demonstrate that our approach using minimal training data performs close to semi-automatic methods using the variational level-set method and manual segmentation carried out by an experienced user. Using our method, which we show to have minimal user intervention and high classification accuracy, we investigate quantitative parameters such as volume of the cytoplasm occupied by mitochondria, differences between the surface area of inner and outer membranes and mean mitochondrial width which are quantities potentially relevant to distinguishing cancer cells from normal cells. To test the accuracy of our approach, these quantities are compared against manually computed counterparts. We also demonstrate extension of these methods to segment 3D images obtained using electron tomography.

Introduction

In this paper, we present a machine learning tool for automatic texton-based joint classification and segmentation of mitochondria in MNT-1 cells imaged using ion-abrasion scanning electron microscopy (IA-SEM) [1], [2], [3]. IA-SEM is a novel technique that can be used to image whole mammalian cells at nanometer resolution. Our tool aims at obtaining a quantitative description of features in the image such as mitochondria.

Object recognition can be defined as the localization of object instances within an image. As mentioned in Ref. [4], object localization is to obtain class models that are invariant enough to incorporate naturally occurring intra-class variations and yet discriminative enough to distinguish between different classes. Both object and region localization usually are realized in practice using texton-based models that are developed in the context of texture recognition [5], [6]. “Texton” and “visual word” have been used interchangeably in computer vision literature, which refers to clusters of filter responses in a high-dimensional space.

Particularly, textures are modeled as joint distribution of filter responses, where the distribution is represented as a frequency histogram of filter response cluster centers (textons) and the texture models are learnt from training images [4], [5], [7], [8]. Many texton-based and object category classification approaches have been applied to image databases, for e.g. CUReT, UIUCTex, CalTech101, CalTech6 and PASCAL challenge (Ref. [8] and references therein), scene understanding [9], semantic image retrieval, web search and interactive image editing [4]. The classification algorithm is affected by parameters such as choice of filter bank and rotational invariance [4], [5], [8], the number of training images (Ref. [5] for a survey) and the parameters of the classifiers (for e.g., choice of k in k-NN and kernel width in SVMs, etc.). Texture recognition has been investigated under various factors such as lighting, viewing conditions [5], [6] brightness and color [10]. In Ref. [11], the authors separate texture from untextured regions, where contours are treated in the intervening contour framework and textures are analyzed using textons. They propose a graph theoretic framework of normalized cuts to find partitions of the image. Texture and object category classifications are studied using different key point detectors and descriptors, as well as different kernels and classifiers by Schmid et al. [8]. Classification using support vector machines (SVMs) have also been extensively applied to object categorization [8], texture [12] and medical applications [13], [14]. A segmentation approach for retinal vessels based on textons has been investigated in Ref. [15]. Recently, multi-class boosting techniques have been applied to region-based classification of aerial imagery [9].

Melanin is produced in small organelles known as melanosomes that are transferred by an yet unknown mechanism from melanocytes to the surrounding keratinocytes [16]. Melanogenesis is the process through which the pigmented melanin is synthesized in melanocytes [17]. Numerous diseases leading to abnormal pigmentation in which patients have compromised immune system have been documented [16]. The knowledge of intracellular distribution of melanosomes at different stages of biogenesis remains an unanswered question and is being addressed using IA-SEM [1], [2]. Mitochondria are also prominent features in cellular interior, and their size, shape and spatial distribution may have some relevance to the overall state of the cell. Our goal is to automatically segment out mitochondria in MNT-1 cells and extract quantities such as the internal volume occupied by mitochondria, the difference between surface areas of inner and outer membranes, and the mean mitochondrial width that can be indicative of the state of the cell.

Texton-based segmentation is challenging in this type of work due to the large size of the data, low contrast and SNR, appearance geometry and viewpoint variation. To improve the SNR, we have implemented a preprocessing step using variational filtering. This method uses an adaptive fidelity term that applies different levels of denoising in different regions [18]. This denoising method preserves the textures and smooth regions are better denoised in terms of signal-to-noise ratio (SNR) compared to nonlinear diffusion [19] and total variation methods [20] that preserve sharpness and location of the edges, and in some cases enhance them, but many small details and textured regions of the image are ignored.

For diagnosing signatures that may be unique to cellular states such as cancer, large volume data sets (typically $\sim 2 GB / cell$ ) need to be analyzed. Various semi-automatic segmentation methods have been developed in the past for biological applications (see chapter 30 of Ref. [21]) that require parameter setting or some sort of a priori information and hence would be a major bottleneck for high-throughput data mining and analysis. Our goal here is to develop a machine learning tool that has minimal user intervention for high-throughput imaging. In our approach, user intervention is limited to choosing the parameters just for the training set (including the parameters for texture preserving filter) which is then retained for subsequent test datasets. We have borrowed texton-based models developed in the context of texture [5] and object [4] classification. Given manually labeled images that contain both segmentation maps and rough annotations of regions, our supervised learning algorithm learns the model such that it can localize novel instances of the regions on test datasets. The segmentation obtained by the proposed methods is compared against the well-studied level-set approach in image segmentation community and manual segmentation by an experienced user.

Binary classification of mitochondria is executed by exploiting region-based texture features. Block-wise classification is performed by $k$ -NN classifier, SVMs, AdaBoost and histogram matching using a NN classifier. Segmentation results demonstrate that the proposed approach using minimal training data perform close to semi-automatic method carried out using variational level-set method and manual segmentation by an experienced user. Our aim is to identify quantitative parameters that describe the mitochondrial structures which can be realized using simple and computationally efficient techniques. Therefore, we study computational complexity in terms of memory and timing requirements versus segmentation accuracy for all the classifiers investigated in this work. Experimental results show that the proposed learning-based algorithm is computationally efficient on large datasets, has minimal user intervention and achieves high classification accuracy. We then make quantitative estimates of measures such as mitochondrial volume and differential surface area between outer and inner membranes with enough numbers to provide statistically meaningful results. To test the accuracy of our approach the quantities are compared against manually computed counterparts. In the end, we test the robustness of automatic texton-based segmentation procedure using a very different type of 3D image dataset obtained using cryo-electron tomography of Doxil, a liposomal formulation of doxorubicin. On tomograms of liposomal doxorubicin formulations (Doxil), an anticancer drug, imaged at cryogenic temperatures, we study the radii and volume distribution which is of great interest in cancer treatment.

The rest of the paper is organized as follows. Data collection techniques are briefly summarized in Section 2. Details of preprocessing, semi-automatic segmentation based on level-sets and automatic segmentation using machine learning approaches are presented in Section 3. Experimental results and quantitative analysis are presented in Section 4. Conclusion and future work are drawn in Section 5.

Section snippets

Data collection techniques

Conventional optical imaging is limited to a resolution of $\sim 0.2 μ m$ microns in the very best cases. Heymann et al. [1], [2] have recently reported a new strategy for 3D imaging of whole cells at resolutions about an order of magnitude higher than what can be achieved with light microscopy (see Ref. [1] for details). Use of this method, termed as IA-SEM, allows imaging of the subcellular architecture of mammalian cells and reveals the locations and shape of membrane-bound organelles such as

Preprocessing of datasets

Our intention is to automatically classify and segment out the mitochondrial structures in MNT-1 cells (melanoma cells) from the rest of the cytoplasm that have a distinct texture. Datasets obtained using electron microscopes are intrinsically noisy due to various contributing factors during the imaging and reconstruction procedures [22]. Denoising algorithms based on nonlinear diffusion [19] and total variation methods [20] preserve sharpness and location of the edges, and in some cases

Experimental results

In our experiments, we consider two datasets of mitochondria in MNT-1 cells imaged using IA-SEM. Both the datasets were of low contrast and SNR. These datasets were denoised using texture preserving using the following parameters in order to enhance the texture part of the 3D dataset. In our experiments, $τ = 6$ , and $μ = 0.04$ , such that $τ μ < 0.25$ . We set $σ = 1.5$ (standard deviation of the Gaussian), filter size to be 10, $λ = 5$ , and $ν = {- 2, 0, 2}$ . We define initial level-set function such that it points

Conclusion and future work

In this paper, we investigated automatic techniques for joint classification and segmentation and quantitative analysis of mitochondria in mammalian cells imaged using ion-abrasion scanning electron microscope. Given manually labeled images, a machine learning approach was formulated that exploits texture features, and joint image block-wise classification and segmentation of novel instances of the regions on diverse test datasets was performed by histogram matching using a nearest neighbor

References (37)

J.A. Heymann et al.
Site-specific 3D imaging of cells and tissues with a dual beam microscope
Journal of Structual Biology
(2006)
G.R. Rosania
Mitochondria give cells a tan
Chem. Bio.
(2005)
L. Rudin et al.
Nonlinear total variation based noise removal algorithms
Phys. D: Nonlinear Phenom.
(1992)
S.A. Abraham et al.
Liposomal formulations of doxorubicin
J. Heymann et al.
3D imaging of melanoma cells using automated ion abrasion scanning electron microscopy
Microscopy and Microanalysis
(2007)
J.A. Heymann, D. Shi, S. Kim, D. Bliss, L.S. Jacqueline, S. Subramaniam, Nanometer resolution 3D imaging of mammalian...
J. Winn, A. Criminisi, T. Minka, Object Categorization by Learned Universal Visual Dictionary, in: Proceedings of the...
M. Varma et al.
A statistical approach to texture classification from single images
Int. J. Comp. Vision
(2005)
T. Leung et al.
Representing and recognizing the visual appearance of materials using three-dimensional textons
Int. J. Comp. Vision
(2001)
O.G. Cula et al.
Compact representation of bidirectional texture functions
Proc. Comput. Vision Pattern Recognition
(2001)

J. Zhang et al.

Local features and kernels for classification of texture and object categories: a comprehensive study

Int. J. Comp. Vision

(2007)

J. Kaufhold et al.

Recognition and segmentation of scene content using region-based classification

D.R. Martin et al.

Learning to detect natural image boundaries using brightness color and texture cues

IEEE Trans. Pattern Anal. Mach. Intell.

(2002)

J. Malik et al.

Contour and texture analysis for image segmentation

Int. J. Comp. Vision

(2001)

H. Zhang et al.

SVM-KNN: discriminative nearest neighbor classification for visual category recognition

O. Tuzel, L. Yang, P. Meer, D.J. Foran, Classification of hematologic malignancies using texton signatures, Pattern...

E.N. Issam et al.

A support vector machine approach for detection of microcalcifications

IEEE Trans. Med. Imaging

(2002)

D.A. Adjeroh et al.

Texton-based segmentation of retinal vessels

J. Opt. Soc. Am. A

(2007)

Cited by (42)

Mitochondria morphometry in 3D datasets obtained from mouse brains with serial block-face scanning electron microscopy
2023, Methods in Cell Biology
The dysfunction of mitochondria is linked with many diseases. In the nervous system, evidence of their implication in neurodegenerative disease is growing. Mitochondria health is assessed by their impact on cellular metabolism but alterations in their morphologies and locations in the cells can also be markers of dysfunctions. Light microscopy techniques allow us to look at mitochondria in vivo in cells or tissue. But in the case of the nervous system, in order to assess the precise location of mitochondria in different cell types and neuronal compartments (cell bodies, dendrites or axons), electron microscopy is required. While the percentage of volume occupied by mitochondria can be assessed on 2D images, alterations in length, branching, and interactions with other organelles require three-dimensional (3D) segmentation of mitochondria in volumes imaged at ultrastructural level. Nowadays three-dimensional volume electron microscopy (vEM) imaging techniques such as serial block face scanning electron microscopy (SBF-SEM) enable us to image 3D volumes of tissue at ultrastructural level and can be done routinely. Segmentation of all the neuropil is also successfully achieved at a large scale in the nervous system. Here, we show a workflow based on open access resources, which allows us to image, segment, and analyze mitochondria in 3D volumes of regions of interest in the mouse brain. Taking advantage of recent developments, e.g., pre-trained models for mitochondria, we speed up the reconstruction and analysis. We also critically assess the impact on the results of the different reconstruction methods chosen and the level of manual corrections invested.
Fear memory-associated synaptic and mitochondrial changes revealed by deep learning-based processing of electron microscopy data
2022, Cell Reports
Serial section electron microscopy (ssEM) can provide comprehensive 3D ultrastructural information of the brain with exceptional computational cost. Targeted reconstruction of subcellular structures from ssEM datasets is less computationally demanding but still highly informative. We thus developed a region-CNN-based deep learning method to identify, segment, and reconstruct synapses and mitochondria to explore the structural plasticity of synapses and mitochondria in the auditory cortex of mice subjected to fear conditioning. Upon reconstructing over 135,000 mitochondria and 160,000 synapses, we find that fear conditioning significantly increases the number of mitochondria but decreases their size and promotes formation of multi-contact synapses, comprising a single axonal bouton and multiple postsynaptic sites from different dendrites. Modeling indicates that such multi-contact configuration increases the information storage capacity of new synapses by over 50%. With high accuracy and speed in reconstruction, our method yields structural and functional insight into cellular plasticity associated with fear learning.
Few-shot hypercolumn-based mitochondria segmentation in cardiac and outer hair cells in focused ion beam-scanning electron microscopy (FIB-SEM) data
2019, Pattern Recognition Letters
We present a novel AI-based approach to the few-shot automated segmentation of mitochondria in large-scale electron microscopy images. Our framework leverages convolutional features from a pre-trained deep multilayer convolutional neural network, such as VGG-16. We then train a binary gradient boosting classifier on the resulting high-dimensional feature hypercolumns. We extract VGG-16 features from the first four convolutional blocks and apply bilinear upsampling to resize the obtained maps to the input image size. This procedure yields a 2688-dimensional feature hypercolumn for each pixel in a 224 × 224 input image. We then apply L₁-regularized logistic regression for supervised active feature selection to reduce dependencies among the features, to reduce overfitting, as well as to speed-up gradient boosting-based training. During inference we block process 1728 × 2022 large microscopy images. Our experiments show that in such a formulation of transfer learning our processing pipeline is able to achieve high-accuracy results on very challenging datasets containing a large number of irregularly shaped mitochondria in cardiac and outer hair cells. Our proposed few-shot training approach gives competitive performance with the state-of-the-art using far less training data.
Machine learning for locating organic matter and pores in scanning electron microscopy images of organic-rich shales
2019, Fuel
Citation Excerpt :
Various segmentation methods have been proposed to process SEM images to derive the information at nanoscale. Narasimha et al. [25] tested kNN, SVMs and Adaboost methods on SEM mammalian cells images in segmentation of mitochondria using texton-based region features. Segmentation results demonstrate the machine learning method can perform close to manual segmentation carried out by an experienced user.
For purposes of locating kerogen/organic matter and pores in SEM images of shale samples, we tested an automated SEM-image segmentation workflow involving feature extraction followed by machine learning. The proposed segmentation workflow is an alternative to threshold-based and object-based segmentation. For each pixel in the SEM image, sixteen features are generated and then fed to a random forest classifier to determine the presence of four rock components, namely (1) pore/crack, (2) rock matrix including clay, calcite and quartz, (3) pyrite, and (4) organic/kerogen components. Extraction of the 16 features for each pixel in a 2058 (pixel) by 2606 (pixel) SEM image requires approximately 5 s on a conventional desktop configuration. The most important features for the desired segmentation are wavelet transforms, Gaussian blur and difference of Gaussians. The random-forest classifier was trained using only 705, 2074, 17373, and 15,000 pixels for the pore/crack, organic/kerogen, rock matrix, and pyrite components, respectively. Training of the classifier was accomplished in approximately 7 s. The proposed segmentation methodology can segment one 2058 (pixel) by 2606 (pixel) SEM image in approximately 30 s. For purposes of validation of the segmentation method, 5121, 4725, 4815, and 4775 pixels were used for the pore/crack, organic/kerogen, rock matrix, and pyrite components, respectively, where 1 pixel covers 100 nm². Despite the limited number of pixels used for training, the performance of the trained classifier, quantified in terms of overall F1 score, on the validation dataset was higher than 0.9. The newly developed method is significantly more reliable and robust in comparison to the popular histogram thresholding and object-based segmentation methods, especially for the matrix and pyrite components in shale samples.
On the projection similarity in line grouping
2015, Pattern Recognition Letters
Citation Excerpt :
This fact accounts for the vast range of mitochondrial shapes and textures and challenges a unified approach to localization and segmentation. Presently the segmentation of mitochondria is performed manually [16], though automatic classification has also been recently reported in [14,7]. In many cases transmission electron microscope (TEM) images of mitochondria show characteristic quasi-linear structural elements (Fig. 1(c)–(d)) which can be converted to the proper line segments.
This paper is concerned with the grouping of elementary line segments which are comprised of pairwise projected entities. In a dynamic and densely cluttered scene we consider a feature-driven recognition of objects with predominantly linear or quasi-linear structural elements. The motivation arises from the field of biological imaging such as the detection of mitochondria in a complex subcellular environment. Subsequent line extraction operations result in a set of line segments with different lengths, density and orientations. We observe that a distinct criterion to distinguish such a salient group of line segments from the background can be formulated as Projectivity. We introduce a new similarity measure Projection-to-Distance Ratio which combines the proximity and the amount of spanned orthogonal projections between two line segments. Further, we perform investigations on the Euclidean properties of the proposed similarity measure. We construct the similarity matrix and show that it translates into an indefinite pseudo-covariance matrix. In order to test the introduced similarity measure we examine the applicability of NN (nearest neighbor) and non-NN clustering methods for the grouping of line segments.
Unsupervised mitochondria segmentation using recursive spectral clustering and adaptive similarity models
2013, Journal of Structural Biology
Citation Excerpt :
Due to their central role in cellular bioenergetics, apoptosis and autophagy, identification and classification of mitochondria is an area of increasing biomedical importance. Narasimha et al. (2009), for example, propose an automatic texton-based joint classification and segmentation of mitochondria by utilizing k-nearest neighbor (k-NN) classifier, support vector machines (SVMs), adaptive boosting (AdaBoost) and histogram matching. Another texture-based approach is reported in Vitaladevuni et al. (2008) where authors consider the detection of mitochondria in TEM images of brain tissue and apply the Gentle-Boost classifier.
The unsupervised segmentation method proposed in the current study follows the evolutional ability of human vision to extrapolate significant structures in an image. In this work we adopt the perceptual grouping strategy by selecting the spectral clustering framework, which is known to capture perceptual organization features, as well as by developing similarity models according to Gestaltic laws of visual segregation. Our proposed framework applies but is not limited to the detection of cells and organelles in microscopic images and attempts to provide an effective alternative to presently dominating manual segmentation and tissue classification practice. The main theoretical contribution of our work resides in the formulation of robust similarity models which automatically adapt to the statistical structure of the biological domain and return optimal performance in pixel classification tasks under the wide variety of distributional assumptions.

View all citing articles on Scopus

About the Author—RAJESH NARASIMHA received the Bachelor's degree in electronics and communication engineering from Mangalore University, Mangalore, India, in 1999, the Master's degree in electrical engineering from the Rochester Institute of Technology (RIT), Rochester, NY, in 2002, and the Ph.D. degree in electrical and computer engineering from the Georgia Institute of Technology, Atlanta, in 2007. From 1999 to 2000, he was a Research Associate with the Electronics and Communication Engineering Department. Indian Institute of Science (IISc), Bangalore working on speech recognition for Ericsson, Inc. From August to December 2002, he was a Research Associate with RIT, and collaborated with Xerox Corporation, Webster, Rochester, on content-based video summarization and browsing applications. From 2006 to 2007, he was a Predoctoral Fellow at the National Cancer Research Institute, NIH, Bethesda, MD, where he worked on applications of machine learning to biology. He is the recipient of an SPIE scholarship from 2003 to 2004 and the NIH Predoctoral Fellowship from 2006 to 2007. Currently he is a Member of the Technical Staff at the DSPS R&D Center, Texas Instruments, Incorporated. He is a member of SPIE and IEEE. His research interests span broad areas of signal, image, and video processing; machine learning applications to scalable data mining from large and complex datasets; and statistical modeling and data analysis.

About the Author—HUA OUYANG is currently pursuing his Ph.D. degree at the College of Computing at the Georgia Institute of Technology. His research interests are machine learning applications for biology.

About the Author—ALEXANDER GRAY After completing Bachelor's degrees in Applied Mathematics (concentration in Computational Statistics) and Computer Science from UC Berkeley, spending Summers at the Santa Fe Institute and Los Alamos National Laboratory, among other places, Alexander Gray worked in the Machine Learning Systems Group of NASA's Jet Propulsion Laboratory, then completed his Ph.D. in Computer Science and a Postdoctoral at Carnegie Mellon University supervised by Prof. Andrew Moore. Since August 2005 he has been an Assistant Professor in the College of Computing at Georgia Tech, within the new Interactive and Intelligent Computing Division and affiliated with the even newer Computational Science and Engineering Division. He is a Member of the Center for the Study of Systems Biology, Center for Robotics and Intelligent Machines, Center for Experimental Research in Computer Systems, and the Graphics, Visualization and Usability Center, and affiliated with the Industrial and Technology Statistics Center.

About the Author—STEVEN W. MCLAUGHLIN received the B.S. degree from Northwestern University, Evanston, IL, in 1985, the M.S.E. degree from Princeton University, Princeton, NJ, in 1986, and the Ph.D. degree from the University of Michigan, Ann Arbor, in 1992, all in electrical engineering. From 1992 to 1996, he was the Electrical Engineering faculty at the Rochester Institute of Technology, Rochester, NY. He joined the School of Electrical and Computer Engineering at the Georgia Institute of Science and Technology, Atlanta, in 1996, where he is Ken Byers Professor of Electrical and Computer Engineering and Vice Provost for International Initiatives. His research interests are in the general area of communications and information theory. His research group has ongoing projects in the areas of turbo, LDPC , and constrained codes for magnetic and optical recording; forward error correction (FEC) and equalization for wireless and optical networks; quantum key distribution, wireless and radio-frequency identification (RFID) security; and theory of error-control coding. He has published more than 200 papers in journals and conferences and holds 26 US patents. Dr. McLaughlin received the Presidential Early Career Award for Scientists and Engineers (PECASE) where he was cited “for leadership in the development of high-capacity, nonbinary optical recording formats.” He received (with Dr. David Warland at UC-Davis) the Information Storage Industries Consortium Technical Achievement Award in 2002 for “pioneering work in the development of multilevel optical disk storage technology”. In 2005, he was President of the IEEE Information Theory Society. He was an Associate Editor for Coding Techniques for the IEEE TRANSACTIONS ON INFORMATION THEORY and was Publications Editor for that journal from 1995 to 1999. He co-edited (with S.Verdu) Information Theory: 50 Years of Discovery (Wiley/IEEE Press, 1999). He has also served on the IEEE Publications Activities Board (1998–2001) and is a former Secretary of the IEEE Atlanta Section (2000).

About the Author—SRIRAM SUBRAMANIAM obtained his Ph.D. in Physical Chemistry from Stanford University in 1987 and carried out postdoctoral work in the Departments of Chemistry and Biology at the Massachusetts Institute of Technology. In 1992, he joined the faculty at the Johns Hopkins University School of Medicine as an Assistant Professor, and was promoted to Associate Professor in 1998. He was appointed as the Chief of the Biophysics Section in the Laboratory of Biochemistry in 1998, and subsequently Chief of the Biophysics Section in the Laboratory of Cell Biology in 2003. He continues to maintain a visiting faculty appointment with the Johns Hopkins University School of Medicine.

View full text

Automatic joint classification and segmentation of whole cell 3D images

Abstract

Introduction

Section snippets

Data collection techniques

Preprocessing of datasets

Experimental results

Conclusion and future work

Journal of Structual Biology

Chem. Bio.

Phys. D: Nonlinear Phenom.

3D imaging of melanoma cells using automated ion abrasion scanning electron microscopy

Microscopy and Microanalysis

A statistical approach to texture classification from single images

Int. J. Comp. Vision

Representing and recognizing the visual appearance of materials using three-dimensional textons

Int. J. Comp. Vision

Compact representation of bidirectional texture functions

Proc. Comput. Vision Pattern Recognition

Local features and kernels for classification of texture and object categories: a comprehensive study

Int. J. Comp. Vision

Recognition and segmentation of scene content using region-based classification

Learning to detect natural image boundaries using brightness color and texture cues

IEEE Trans. Pattern Anal. Mach. Intell.

Contour and texture analysis for image segmentation

Int. J. Comp. Vision

SVM-KNN: discriminative nearest neighbor classification for visual category recognition

A support vector machine approach for detection of microcalcifications

IEEE Trans. Med. Imaging

Texton-based segmentation of retinal vessels

J. Opt. Soc. Am. A