Histological image classification using biologically interpretable shape-based features
© Kothari et al.; licensee BioMed Central Ltd. 2013
Received: 15 March 2012
Accepted: 20 February 2013
Published: 13 March 2013
Automatic cancer diagnostic systems based on histological image classification are important for improving therapeutic decisions. Previous studies propose textural and morphological features for such systems. These features capture patterns in histological images that are useful for both cancer grading and subtyping. However, because many of these features lack a clear biological interpretation, pathologists may be reluctant to adopt these features for clinical diagnosis.
We examine the utility of biologically interpretable shape-based features for classification of histological renal tumor images. Using Fourier shape descriptors, we extract shape-based features that capture the distribution of stain-enhanced cellular and tissue structures in each image and evaluate these features using a multi-class prediction model. We compare the predictive performance of the shape-based diagnostic model to that of traditional models, i.e., using textural, morphological and topological features.
The shape-based model, with an average accuracy of 77%, outperforms or complements traditional models. We identify the most informative shapes for each renal tumor subtype from the top-selected features. Results suggest that these shapes are not only accurate diagnostic features, but also correlate with known biological characteristics of renal tumors.
Shape-based analysis of histological renal tumor images accurately classifies disease subtypes and reveals biologically insightful discriminatory features. This method for shape-based analysis can be extended to other histological datasets to aid pathologists in diagnostic and therapeutic decisions.
We develop an automatic histological image classification system that uses biologically interpretable shape-based features. These features capture the distribution of shape patterns, described by Fourier shape descriptors, in different stains of a histological image. We use this system to classify hematoxylin and eosin (H&E) stained renal tumor images and assess its classification performance by comparing it to methods based on textural, morphological, and topological features.
Over the last decade, several automatic or automated systems have been developed to aid histological cancer diagnosis and to reduce subjectivity. All of these systems attempt to mimic pathologists by extracting features from histological images. Some important features include color, nuclear shape, fractal, textural gray-level co-occurrence matrices (GLCM), wavelets, and topological, among others [4, 5]. Several diagnostic systems for renal cell carcinoma (RCC) are good examples of the utility of these features. For example, Chaudry et al. proposed a system using textural and morphological features with automated region-of-interest selection for RCC subtype classification [6, 7]. Waheed et al. performed a similar analysis but included fractal as well as textural and morphological features . Choi et al. extended the morphological analysis to three-dimensional nuclei and applied their system to RCC grading . In addition to morphological features, Francois et al. used cell kinetic features in their RCC grading system . Finally, Raza et al. used a scale invariant feature transform (SIFT) method to classify RCC subtypes . Despite the success of these systems in terms of diagnostic accuracy, widespread use of these systems is limited by a lack of feature interpretability. Some researchers have provided visual interpretation of features. For example, some topological features have been related to the amount of differentiation in varying cancer grades . In contrast, pathologists may not be receptive to, or confident in, features such as wavelet or fractal representations of images because they are not easy to interpret biologically. Moreover, most existing systems exploit morphological properties of nuclear shapes and ignore cytoplasmic and glandular structures despite evidence of their utility . Thus, methods based on a holistic view of shapes and colors may more accurately reflect the process by which a pathologist interprets a renal tumor image .
Fourier shape descriptors, described by Kuhl and Giardina  have been reported to be very useful as shape descriptors. They are highly robust to high frequency noise because of their ability to reject higher harmonic shape descriptors. Researchers have used Fourier shape descriptors for various medical imaging applications, including shape-based vertebral image retrieval , and classification of breast tumors . The medical images involved in these studies typically have definite shapes with consistent landmarks. In addition, researchers have used Fourier shape descriptors for analyzing the shapes of nuclear structures [17–19]. Histological images, however, lack such landmarks and they tend to exhibit multiple highly variable shapes. As such, it is difficult to compare histological images using common techniques such as template matching with an image atlas  or using shape-based similarity measures after registration of the shapes in a histological image . Therefore, in order to characterize and compare histological images in terms of shapes, we quantify the distribution of shape patterns in an image using Fourier shape descriptors.
We perform this study on hematoxylin and eosin (H&E) stained histological RGB image datasets acquired from renal tumor samples of patients. In this study, we use two separately acquired datasets: dataset A and dataset B. Both datasets consist of photomicrographs of deidentified renal tumor specimens, derived from human patients. Research was conducted in compliance with the Helsinki Declaration. Tumor specimens were obtained through protocols approved by the Emory University Institutional Review Board, in which patients provided informed consent for residual tumor tissue to be stored in a university tissue bank. Administrators of the tissue bank provided deidentified tissues and associated clinical data (scrubbed of personal health identifiers), to the investigators of this research project. The IRB protocols pertaining to this research project are Emory IRB00045858/1214-2003 and 255–2002. Refer to Figures 1a-d and Figures 1e-h for samples of images in dataset A and dataset B, respectively. After acquisition at constant magnification, a clinician selected 1600 × 1200-pixel portions from whole-slide images and a pathologist assigned each image to a renal tumor subtype. Dataset A contains 48 images with 12 images of each subtype while dataset B has 55 images including 20 chromophobe (CH), 17 clear cell (CC), 13 papillary (PA), and 5 oncocytoma (ON) subtypes. Dataset B has samples with nuclear grade varying from 1 to 4. In total, we analyze 103 renal tumor H&E images.
Automatic color segmentation of the renal tumor images requires an additional reference dataset. The reference dataset need not be the same tissue type. However, the staining protocol should be the same as that of the renal tumor images. We use an H&E stained dataset of 50 randomly selected ovarian cancer images from the NIH Cancer Genome Atlas (TCGA) repository . We use 1024 × 1024-pixel cropped portions of the original slide images. As references, these images are segmented by an expert user with the aid of a user-interactive system . We then use these color-segmented reference images to automatically segment the renal tumor images as described in the following section.
Automatic color segmentation
H&E staining of a renal tumor histological image enhances three colors: blue-purple, white, and pink. These colors correspond to specific cellular structures. Basophilic structures containing nucleic acids—ribosome and nuclei—tend to stain blue-purple; eosinophilic intra- and extracellular proteins in cytoplasmic regions tend to stain bright pink; empty spaces the lumen of glands do not stain and tend to be white. In order to isolate shapes corresponding to these cellular structures, we segment the three colors of every image using an automatic color segmentation method .
The 10 segmentation labels for each pixel (one for each ovarian reference image) are combined using a voting scheme (Figure 3, Step 2). Voting chooses the segmentation label most frequently assigned to a pixel as its preliminary label.
After segmentation, we extract a binary mask for each stain and apply morphological operations to the binary mask to connect broken boundaries and separate overlapping objects. Namely, we dilate objects in the nuclear mask with a circular structural element with a two-pixel radius and erode objects in the cytoplasmic and glandular masks with a circular structural element with a three-pixel radius. Finally, from all binary masks, we remove small noisy regions with area less than five pixels and extract outer boundaries of the remaining connected objects for further analysis.
Discretization of shape descriptors
Generate a binary mask for each color in the histological image. We use three colors for H&E stained RCC images: blue (nuclear), white (no-stain/glandular), and pink (cytoplasmic).
Extract contours for all shapes in a mask after connected component analysis.
Extract axis lengths for Fourier ellipses ( and ) for the first 10 harmonics (n). This will give us 2*10 variables for each shape.
For each harmonic (n), axis type (c), and mask (m), perform a binning procedure (Figure 8). We generate 20 histograms for each mask. We use 15 bins and a range determined by and as previously described.
Combine histogram frequency from the three masks to generate a list of 900 shape-based features
There are a number of advantages in using discretization rather than Euclidian distance to compare images. First, the axes of shapes that are similar, but perhaps not identical, fall into the same histogram bin. Similar histogram frequencies can be interpreted as a similarity of shapes between images. Second, bins sensitive to noise or outlier shapes in any sample will be rejected during feature selection. Finally, discriminating features can be components corresponding to multiple types of shapes rather than components corresponding to the most prominent characteristic shape.
Traditional features in computer-aided diagnosis include texture, morphological, topological, and nuclear. In order to compare shape-based features to these traditional features, we extract additional features from histological renal tumor images.
For texture, we have two sets of features: Gray-Level Co-occurrence Matrix (GLCM) and wavelet. For GLCM features, we extract a 16 × 16 GLCM matrix for each gray-scale tissue image with 16 quantization levels . Using this matrix, we extract 13 texture properties including contrast, correlation, energy (angular second moment), entropy, homogeneity (inverse difference moment), variance, sum average, sum variance, sum entropy, difference variance, difference entropy, and two information measures for correlation. These features are reported to successfully capture texture properties of the image and are very useful in automated cancer grading [12, 27, 28].
For wavelet features, we perform three-level wavelet (db6) packet decomposition  of the gray-level tissue image and extract energy and entropy  of 84 coefficient matrices (level 1, 2 and 3), producing 168 features. Wavelet features capture texture properties of an image.
For morphological features, we use color-GLCM, a method proposed by Chaudry et al. to classify renal tumor subtypes. This method generates a four-level gray-scale image from four color stains in H&E-stained images . The four colors resulting from H&E-stained images (blue, white, pink, and red) correspond to segmented regions of nuclei, lumen, cytoplasm, and red blood cells. We then extract a 4 × 4 GLCM matrix for the gray-scale image. We extract 21 features from this matrix including 16 elements of the 4 × 4 GLCM matrix, contrast, correlation, energy (angular second moment), entropy, and homogeneity (inverse difference moment). These features capture morphological features of the image such as stain area and stain co-occurrence properties.
For topological features, we use a graph-based method. Several researchers have proposed graph-based features to capture the distribution of patterns in an image. Biologically, these features capture the amount of differentiation (related to cancer grade) in a histological image. We morphologically erode our nuclear mask to separate nuclear clusters and use their centroids (nuclear centers) for this analysis. First, we create a Voronoi diagram from these centers and then calculate area and perimeter of each region and all side-lengths. We then calculate mean, minimum, maximum, and disorder of the distribution to produce 12 features . The disorder, D, of a distribution, r, is given by , where σ r and μ r are standard deviation and mean of r, respectively . Second, we calculate the area and side lengths of the Delaunay triangles and extract statistics similar to those of the Voronoi diagram to produce eight more features. Last, we calculate side lengths of the minimum spanning tree and extract the same statistics to produce four more features. In total, we extract 24 topological features.
For nuclear features, we extract nuclear count and elliptical-shape properties, which have proven to be useful for renal carcinoma subtyping and grading . For segmenting nuclear clusters, we use an edge-based method with three steps: concavity detection, straight-line segmentation, and ellipse fitting . We describe each elliptical nucleus using area, major-axis length, minor-axis length, and eccentricity. We then calculate mean, minimum, maximum and disorder of the distribution of these descriptors to produce 16 features. In total, including nuclear count, we extract 17 nuclear features.
We combine the GLCM (13 features), color-GLCM (21), wavelet (168), topological (24), and nuclear (17) features to produce a set of 243 “Combined Traditional” features. Finally, we combinethe “Combined Traditional” (243) and “Shape” (900) features to a produce a set of 1143 “All” features.
Feature selection and classification
For validation, we combine datasets A and B, then randomly split them into two new training and testing datasets with balanced sampling from both datasets. We perform a three-fold split, in which two folds form the training set while one fold forms the testing set. Each fold acts as a testing set once, resulting in three training–testing sets. We perform 10 iterations of this split to estimate the variance in performance. Thus, there are 30 training–testing sets in the external cross-validation (CV) that produces the final classification accuracy. For each of the 30 training sets, we perform an additional three-fold, 10 iterations of CV to choose an optimal set of classifier and feature selection parameters. This forms the internal CV of a nested CV (Figure 8).
Combined Traditional (1:6:243)
Shape and All (5:5:180)
We choose the feature size step such that the total number of feature sizes is approximately 40. For Shape and All features we also consider number of harmonics (n = 2 to 10) as a feature selection parameter. We choose the simplest model with a CV accuracy within one standard deviation of the best performing model . In choosing the simplest model, we give preference to the linear SVM kernel over the radial SVM kernel and lower values of gamma for the radial SVM kernel, SVM cost, number of harmonics, and feature size.
where k’ is the transformed feature k, μ k and σ k are the mean and standard deviation of feature k over all samples in the training dataset, respectively.
Results and discussion
Shape-based features discriminate renal tumor histological images
Predictive performance of shape-based features
Inner CV accuracy
External CV accuracy
0.77 ± 0.03
CH vs. CC
0.83 ± 0.03
0.83 ± 0.05
CH vs. ON
0.83 ± 0.02
0.84 ± 0.04
CH vs. PA
0.97 ± 0.01
0.96 ± 0.02
CC vs. ON
0.90 ± 0.02
0.90 ± 0.07
CC vs. PA
0.96 ± 0.01
0.95 ± 0.04
ON vs. PA
0.94 ± 0.01
0.93 ± 0.04
Frequently selected model parameters for each binary comparison
CH vs. CC
CH vs. ON
CH vs. PA
CC vs. ON
CC vs. PA
ON vs. PA
CC and PA are the most prevalent subtypes of RCC and are generally the easiest for pathologists to visually identify. Consequently, discriminating shape-based features for these classes are easy to identify, resulting in high classification performance. One exception, however, is the CH vs. CC comparison. CH is known to exhibit some CC properties such as clear cytoplasm. As a result, the prominent feature for the CC subtype is sometimes not sufficient for accurate classification of CC and CH. Moreover, the ON renal tumor subtype is histologically and genetically very similar to the CH RCC subtype, despite the fact that ON is a benign tumor whereas CH is a carcinoma . This similarity explains the moderate performance of the CH and ON binary classifier.
Shape-based features out-perform or complement traditional histological features
Classification accuracy of features in external CV*
0.57 ± 0.04
0.67 ± 0.02
0.52 ± 0.06
0.50 ± 0.03
0.66 ± 0.03
0.79 ± 0.04a
0.77 ± 0.03b
0.78 ± 0.03c
CH vs. CC
0.75 ± 0.05
0.77 ± 0.05
0.74 ± 0.05
0.74 ± 0.05
0.76 ± 0.06
0.81 ± 0.03
0.83 ± 0.05
0.82 ± 0.05
CH vs. ON
0.76 ± 0.05
0.68 ± 0.06
0.67 ± 0.05
0.72 ± 0.05
0.79 ± 0.05
0.86 ± 0.05
0.84 ± 0.04
0.88 ± 0.04
CH vs. PA
0.85 ± 0.04
0.95 ± 0.02
0.86 ± 0.05
0.80 ± 0.04
0.91 ± 0.04
0.94 ± 0.02
0.96 ± 0.02
0.96 ± 0.03
CC vs. ON
0.74 ± 0.06
0.78 ± 0.06
0.63 ± 0.03
0.77 ± 0.07
0.93 ± 0.04
0.93 ± 0.04
0.90 ± 0.07
0.91 ± 0.05
CC vs. PA
0.78 ± 0.06
0.97 ± 0.04
0.69 ± 0.09
0.59 ± 0.07
0.76 ± 0.07
0.95 ± 0.05
0.95 ± 0.04
0.97 ± 0.03
ON vs. PA
0.74 ± 0.07
0.86 ± 0.06
0.74 ± 0.07
0.65 ± 0.04
0.96 ± 0.03
0.97 ± 0.03
0.93 ± 0.04
0.92 ± 0.04
Shape-based features are biologically interpretable
Histopathological features of the CC subtype include clear cytoplasm, compact alveolar, tubular, and cystic architecture leading to distinct cell membranes . Comparing CC to PA and ON, we see that clear cytoplasm (no-stain/glandular (white) mask region, outlined with green) is the primary distinguishing characteristic that is noticeably less frequent in PA and ON. On the other hand, because CH images tend to also exhibit halos resembling clear cytoplasm, the distinguishing features between CC and CH are distinct cell membranes (small cytoplasmic (pink) mask areas outlined with green between larger no-stain/glandular (white) mask areas) that are more frequent in CC compared to CH. Similarity in halos and clear cytoplasm shapes is possibly the reason for low accuracy in the CH vs. CC binary classification.
Features of the PA subtype include scanty eosinophilic cytoplasm and a papillary (i.e., finger-like) pattern of growth resulting in long, complex clusters of nuclei . In all comparisons with the PA subtype, complex clusters of nuclei are the dominant distinguishing feature and are generally more prominent in PA (nuclear (blue) mask areas outlined with yellow). The frequency of nuclear shapes in ON appears to be similar to that of PA. However, the nuclear clusters in PA are generally larger and more irregular due to the clustering, resulting in different Fourier shape axes values.
Histopathological features of the CH subtype include wrinkled nuclei with perinuclear halos . When comparing CH to PA or ON, our feature extraction and selection method identifies these halos (no-stain/glandular (white) mask areas, outlined with blue). In addition, single nuclei become dominant when comparing CH to PA.
Histopathological features of the ON subtype include granular cytoplasm with round nuclei, usually arranged in compact nests or microcysts . These round nuclei appear to be dominant in ON when compared to other subtypes. It can be observed that dominant features for both CH and ON are present in the opposite subtype as well. Hence, the difficulty in distinguishing the two subtypes.
Limitations and computational complexity of shape-based features
Some limitations of shape-based features for histological image classification depend on the specific biological application. Shape-based features may not be suitable for cases in which the primary discriminating features are not based on shapes. For example, in cancer grading applications, topological and texture properties may be more useful than shape-based features. Moreover, as we have seen the results of Table 3 and Figure 11, shape-based features may not capture all of the important distinguishing information. For example, in the case of the CH vs. ON endpoint, the addition of texture and wavelet features to shape-based features increases prediction performance by 4%. In addition, for the CC vs. PA endpoint, inclusion of the GLCM texture features increases prediction performance by 2%. Thus, shape-based features are limited to clinical prediction applications that are inherently shape-based, but, in such cases, may be used to complement other non-shape-based features.
The computational complexity of shape-based features is higher than those of traditional histological feature extraction and analysis methods, but should not prevent implementation in a clinical setting. To convert a RGB histological image (1600x1200 pixel portions) into 900 shape-based features (Figure 7), a desktop computer (Intel Xeon E5405 quad-core processor, 20 GB RAM) requires an average of 74.96 seconds. Compared to some histological image features, this processing time is high. However, the processing time depends on the number of harmonics used for representation and the number of shapes in an image. We have reported the processing time for extracting features from the first ten harmonics. However, in practice, we have observed that all optimized models use less than five harmonics. Optimization of these parameters to identify a predictive model can be time consuming depending on the size of the training set. However, in a clinical setting, such a model would only need to be optimized once, and then periodically updated with new patient data. In a clinical scenario, a pathologist that requires a histological diagnosis for a patient would submit a few image samples from a tissue biopsy to a pre-optimized prediction system. Computational time for processing and predicting based on these image samples would be negligible compared to time required for biopsy, image acquisition, and consultation with a pathologist.
We presented a novel methodology for automatic clinical prediction of renal tumor subtypes using shape-based features. These shape-based features describe the distribution of shapes extracted from three dominant H&E stain colors in renal tumor histopathological images. We evaluated the four-class prediction performance of shape-based classification models using 10 iterations of three-fold nested CV. The overall classification accuracy of 77% (average external CV accuracy) is favorable compared to previous methods that use traditional textural, morphological, and wavelet-based features. Moreover, results indicate that combining shape-based features with traditional histological image features can improve prediction performance. The biological significance of the characteristic shapes identified by our algorithm suggests that this automatic diagnostic system mimics the diagnostic criteria of pathologists. We applied this methodology to renal tumor subtype prediction. However, the methodology may be extended to any histological image classification problem that traditionally depends on visual shape analysis by a pathologist. Moreover, these shape-based features may be coupled with other image features to achieve higher diagnostic accuracy.
Renal cell carcinoma
Minimum redundancy maximum relevance
Directed acyclic graph
Gray-level co-occurrence matrix
Directed acyclic graph
Linear discriminant analysis
Support vector machine.
We thank Dr. Todd Stokes and Dr. Mitch Parry for their valuable comments and suggestions. This research has been supported by grants from NIH (Bioengineering Research Partnership R01CA108468, P20GM072069, and CCNE U54CA119338), Georgia Cancer Coalition, Hewlett Packard, and Microsoft Research.
- Siegel R, Ward E, Brawley O, Jemal A: Cancer statistics, 2011. CA Cancer J Clin. 2011, 61 (4): 212-236. 10.3322/caac.20121.View ArticlePubMedGoogle Scholar
- Teloken PE, Thompson RH, Tickoo SK, Cronin A, Savage C, Reuter VE, Russo P: Prognostic Impact of Histological Subtype on Surgically Treated Localized Renal Cell Carcinoma. J Urol. 2009, 182 (5): 2132-2136. 10.1016/j.juro.2009.07.019.View ArticlePubMedPubMed CentralGoogle Scholar
- Eble J, Sauter G, Epstein J, Sesterhenn I: Pathology and genetics of tumours of the urinary system and male genital organs. 2004, Lyon: IARC press LyonGoogle Scholar
- Demir C, Yener B: Automated cancer diagnosis based on histopathological images: a systematic survey. 2005, Tech Rep: Rensselaer Polytechnic InstituteGoogle Scholar
- Gurcan MN, Boucheron LE, Can A, Madabhushi A, Rajpoot NM, Yener B: Histopathological image analysis: A review. Biomed Eng, IEEE Rev. 2009, 2: 147-171.View ArticleGoogle Scholar
- Chaudry Q, Raza SH, Sharma Y, Young AN, Wang MD: Improving renal cell carcinoma classification by automatic region of interest selection. BioInformatics and BioEngineering, 2008 BIBE 2008 8th IEEE International Conference on: 2008. 2008, Athens, Greece: IEEE, 1-6.View ArticleGoogle Scholar
- Chaudry Q, Raza SH, Young AN, Wang MD: Automated Renal Cell Carcinoma Subtype Classification Using Morphological, Textural and Wavelets Based Features. J Signal Process Syst. 2009, 55 (1): 15-23. 10.1007/s11265-008-0214-6.View ArticleGoogle Scholar
- Waheed S, Moffitt RA, Chaudry Q, Young AN, Wang MD: Computer Aided Histopathological Classification of Cancer Subtypes. Bioinformatics and Bioengineering, 2007 BIBE 2007 Proceedings of the 7th IEEE International Conference on: 2007. 2007, Boston, United States: IEEE, 503-508.Google Scholar
- Choi HJ, Choi HK: Grading of renal cell carcinoma by 3D morphological analysis of cell nuclei. Comput Biol Med. 2007, 37 (9): 1334-1341. 10.1016/j.compbiomed.2006.12.008.View ArticlePubMedGoogle Scholar
- François C, Moreno C, Teitelbaum J, Bigras G, Salmon I, Danguy A, Brugal G, van Velthoven R, Kiss R, Decaestecker C: Improving accuracy in the grading of renal cell carcinoma by combining the quantitative description of chromatin pattern with the quantitative determination of cell kinetic parameters. Cytometry B Clin Cytom. 2000, 42 (1): 18-26. 10.1002/(SICI)1097-0320(20000215)42:1<18::AID-CYTO4>3.0.CO;2-S.View ArticleGoogle Scholar
- Raza SH, Sharma Y, Chaudry Q, Young AN, Wang MD: Automated classification of renal cell carcinoma subtypes using scale invariant feature transform. Engineering in Medicine and Biology Society, 2009 EMBC 2009 Annual International Conference of the IEEE: 3–6 Sept. 2009 2009. 2009, Minneapolis, United States: IEEE, 6687-6690.View ArticleGoogle Scholar
- Doyle S, Agner S, Madabhushi A, Feldman M, Tomaszewski J: Automated grading of breast cancer histopathology using spectral clustering with textural and architectural image features. Biomedical Imaging: From Nano to Macro, 2008 ISBI 2008 5th IEEE International Symposium on: 2008. 2008, Paris, France: IEEE, 496-499.View ArticleGoogle Scholar
- Sertel O, Kong J, Catalyurek UV, Lozanski G, Saltz JH, Gurcan MN: Histopathological image analysis using model-based intermediate representations and color texture: Follicular lymphoma grading. J Signal ProcessSyst. 2009, 55 (1): 169-183. 10.1007/s11265-008-0201-y.View ArticleGoogle Scholar
- Kuhl F, Giardina C: Elliptic Fourier features of a closed contour. Comput Graph Image Process. 1982, 18 (3): 236-258. 10.1016/0146-664X(82)90034-X.View ArticleGoogle Scholar
- Lee D, Antani S, Long L: Similarity measurement using polygon curve representation and fourier descriptors for shape-based vertebral image retrieval. In: SPIE Medical Imaging. 2003, 2003: 1283-1291.Google Scholar
- Rangayyan R, El-Faramawy N, Desautels J, Alim O: Measures of acutance and shape for classification of breast tumors. IEEE Trans Med Imaging. 1997, 16 (6): 799-810. 10.1109/42.650876.View ArticlePubMedGoogle Scholar
- Cukierski W, Nandy K, Gudla P, Meaburn K, Misteli T, Foran D, Lockett S: Ranked retrieval of segmented nuclei for objective assessment of cancer gene repositioning. BMC Bioinforma. 2012, 13 (1): 232-10.1186/1471-2105-13-232.View ArticleGoogle Scholar
- Yang L, Tuzel O, Chen W, Meer P, Salaru G, Goodell LA, Foran DJ: PathMiner: a Web-based tool for computer-assisted diagnostics in pathology. IEEE Trans Inf Technol Biomed. 2009, 13: 291-299.View ArticlePubMedPubMed CentralGoogle Scholar
- Comaniciu D, Meer P: Cell Image Segmentation for Diagnostic Pathology. Advanced Algorithmic Approaches to Medical Image Segmentation. Edited by: Suri JS, Setarehdan SK, Singh S. 2002, London: Springer, 541-558.View ArticleGoogle Scholar
- Lao Z, Shen D, Xue Z, Karacali B, Resnick S, Davatzikos C: Morphological classification of brains via high-dimensional shape transformations and machine learning methods. Neuroimage. 2004, 21 (1): 46-57. 10.1016/j.neuroimage.2003.09.027.View ArticlePubMedGoogle Scholar
- Berg AC, Berg TL, Malik J: Shape matching and object recognition using low distortion correspondences. Computer Vision and Pattern Recognition, 2005 CVPR 2005 IEEE Computer Society Conference on: 2005 2005. 2005, San Diego, United States: IEEE, 26-33. Volume 21.Google Scholar
- McLendon R, Friedman A, Bigner D, Van Meir E, Brat D, Mastrogianakis G, Olson J, Mikkelsen T, Lehman N, Aldape K: Cancer Genome Atlas Research Network. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature. 2008, 455 (7216): 1061-1068. 10.1038/nature07385.View ArticleGoogle Scholar
- Kothari S, Phan JH, Moffitt RA, Stokes TH, Hassberger SE, Chaudry Q, Young AN, Wang MD: Automatic batch-invariant color segmentation of histological cancer images. Biomedical Imaging: From Nano to Macro, 2011 IEEE International Symposium on: March 30 2011-April 2 2011 2011. 2011, Chicago, United States: IEEE, 657-660.Google Scholar
- Persoon E, Fu K: Shape discrimination using Fourier descriptors. IEEE Trans Syst Man Cybern. 1977, 7 (3): 170-179.View ArticleGoogle Scholar
- Wong W, Shih F, Liu J: Shape-based image retrieval using support vector machines, Fourier descriptors and self-organizing maps. Inform Sci. 2007, 177 (8): 1878-1891. 10.1016/j.ins.2006.10.008.View ArticleGoogle Scholar
- McGill R, Tukey J, Larsen W: Variations of box plots. Am Stat. 1978, 32 (1): 12-16.Google Scholar
- Haralick R, Shanmugam K, Dinstein I: Textural features for image classification. IEEE Trans Syst Man Cybern. 1973, 3 (6): 610-621.View ArticleGoogle Scholar
- Tae-Yun K, Hyun-Ju C, Soon-Joo C, Heung-Kook C: Study on texture analysis of renal cell carcinoma nuclei based on the Fuhrman grading system. Enterprise networking and Computing in Healthcare Industry, 2005 HEALTHCOM 2005 Proceedings of 7th International Workshop on: 2005. 2005, 384-387.View ArticleGoogle Scholar
- Laine A, Fan J: Texture classification by wavelet packet signatures. IEEE Trans Pattern Anal Mach Intell. 1993, 15 (11): 1186-1191. 10.1109/34.244679.View ArticleGoogle Scholar
- Jafari-Khouzani K, Soltanian-Zadeh H: Multiwavelet grading of pathological images of prostate. IEEE Trans Biomed Eng. 2003, 50 (6): 697-704. 10.1109/TBME.2003.812194.View ArticlePubMedGoogle Scholar
- Sudbø J, Marcelpoil R, Reith A: New algorithms based on the Voronoi Diagram applied in a pilot study on normal mucosa and carcinomas. Anal Cell Pathol. 2000, 21 (2): 71-86.View ArticlePubMedPubMed CentralGoogle Scholar
- Kothari S, Phan JH, Young AN, Wang MD: Histological Image Feature Mining Reveals Emergent Diagnostic Properties for Renal Cancer. Bioinformatics and Biomedicine (BIBM), 2011 IEEE International Conference on: 2011. 2011, Atlanta, United States: IEEE, 422-425.View ArticleGoogle Scholar
- Kothari S, Chaudry Q, Wang MD: Extraction of informative cell features by segmentation of densely clustered tissue images. Engineering in Medicine and Biology Society, 2009 Annual International Conference of the IEEE: 3–6 Sept. 2009 2009. 2009, Minneapolis, United States: IEEE, 6706-6709.View ArticleGoogle Scholar
- Platt J, Cristianini N, Shawe-Taylor J: Large margin DAGs for multiclass classification. Adv Neural Inf Process Syst. 2000, 12 (3): 547-553.Google Scholar
- Boser B, Guyon I, Vapnik V: training algorithm for optimal margin classifiers. 1992, New York, NY, USA: ACM, 144-152.Google Scholar
- Chang CC, Lin CJ: LIBSVM: A library for support vector machines. ACM Trans Intell Syst Technol (TIST). 2011, 2 (3): 27-Google Scholar
- Hastie T, Tibshirani R, Friedman JH: The elements of statistical learning: data mining, inference, and prediction. 2009, Verlag: SpringerView ArticleGoogle Scholar
- Ding C, Peng H: Minimum redundancy feature selection from microarray gene expression data. J Bioinforma Comput Biol. 2005, 3 (2): 185-10.1142/S0219720005001004.View ArticleGoogle Scholar
- Sakai Y, Watanabe S, Matsukuma S: Chromophobe renal cell carcinoma showing oncocytoma-like hyalinized and edematous stroma: a case report and review of the literature. Urol Oncol. 2004, 22 (6): 461-464. 10.1016/j.urolonc.2004.03.015.View ArticlePubMedGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1471-2342/13/9/prepub
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.