Search

Introduction: We used two novel computer-assisted imaging applications that incorporate nuclear shape and texture analysis to differentiate Gleason grade patterns 3 and 4 of H&Ein a tissue microarray (TMA). The two approaches utilized a variational adaptive segmentation scheme (AdACM) (Rutgers) and the cardinal multiridgelet transform (CMRT) based texture analysis (University of Pittsburgh). Methods: The TMA include 0.6 mm cores from blocks that included Gleason patterns of CaP radical prostatectomy cases from 40 patients. The TMA comprised 13 Gleason patterns 6 (3+3), 8 pattern 7 (3+4), 7 pattern 8 (4+4) and 5 pattern 9 (4+5) Gleason grade (GG) patterns. The AdACM uses 200 × 200 pixel images extracted for nuclear & architectural textural features. All features were based on mathematical energy terms with a shape prior in a multi-level set formulation. The CMRT is an image transform that is translation- and rotation-invariant, and provides a means of displaying texture contents for selection of discriminative features. With a 3-scale decomposition, it gives 8 ridgelet packets and 16 statistical texture features in a 256 x 256 pixel image patch that are averaged over 9 non-overlapping patches in a given image for separation of GG patterns 3 and 4. Results: The CMRT study utilized 3 of 16 texture features extracted from images and trained a Gaussian kernel support vector machine (SVM), a leave-one-out cross-validation showed an accuracy of 94% and AUC of 0.97. For AdACM, nuclear features and architectural features were extracted. Linear Discriminant Analysis (LDA) was employed for feature selection and best features from all three classes could discriminate Gleason grade 3 vs. 4 with an accuracy of 86% via a SVM classifier. Conclusions: Unique quantitative imaging of tissue nuclear shape and texture features can accurately differentiate Gleason patterns 3 and 4. Our research will next evaluate CaP outcomes. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 4061. doi:1538-7445.AM2012-4061

Background Dimensionality reduction (DR) enables the construction of a lower dimensional space (embedding) from a higher dimensional feature space while preserving object-class discriminability. However several popular DR approaches suffer from sensitivity to choice of parameters and/or presence of noise in the data. In this paper, we present a novel DR technique known as consensus embedding that aims to overcome these problems by generating and combining multiple low-dimensional embeddings, hence exploiting the variance among them in a manner similar to ensemble classifier schemes such as Bagging. We demonstrate theoretical properties of consensus embedding which show that it will result in a single stable embedding solution that preserves information more accurately as compared to any individual embedding (generated via DR schemes such as Principal Component Analysis, Graph Embedding, or Locally Linear Embedding). Intelligent sub-sampling (via mean-shift) and code parallelization are utilized to provide for an efficient implementation of the scheme. Results Applications of consensus embedding are shown in the context of classification and clustering as applied to: (1) image partitioning of white matter and gray matter on 10 different synthetic brain MRI images corrupted with 18 different combinations of noise and bias field inhomogeneity, (2) classification of 4 high-dimensional gene-expression datasets, (3) cancer detection (at a pixel-level) on 16 image slices obtained from 2 different high-resolution prostate MRI datasets. In over 200 different experiments concerning classification and segmentation of biomedical data, consensus embedding was found to consistently outperform both linear and non-linear DR methods within all applications considered. Conclusions We have presented a novel framework termed consensus embedding which leverages ensemble classification theory within dimensionality reduction, allowing for application to a wide range of high-dimensional biomedical data classification and segmentation problems. Our generalizable framework allows for improved representation and classification in the context of both imaging and non-imaging data. The algorithm offers a promising solution to problems that currently plague DR methods, and may allow for extension to other areas of biomedical data analysis.

Commodity graphics hardware has become a cost-effective parallel platform to solve m any general computational problems. In medical imaging and more so in digital pathology, segmentation of multiple structures on high-resolution images, is often a complex and computationally expensive task. Shape-based level set segmentation has recently emerged as a natural solution to segmenting overlapping and occluded objects. However the flexibility of the level set method has traditionally resulted in long computation times and therefore might have limited clinical utility. The processing times even for moderately sized images could run into several hours of computation time. Hence there is a clear need to accelerate these segmentations schemes. In this paper, we present a parallel implementation of a computationally heavy segmentation scheme on a graphical processing unit (GPU). The segmentation scheme incorporates level sets with shape priors to segment multiple overlapping nuclei from very large digital pathology images. We report a speedup of 19× compared to multithreaded C and MATLAB-based implementations of the same scheme, albeit with slight reduction in accuracy. Our GPU-based segmentation scheme was rigorously and quantitatively evaluated for the problem of nuclei segmentation and overlap resolution on digitized histopathology images corresponding to breast and prostate biopsy tissue specimens.

Introduction: Spatially invariant vector quantization (SIVQ) is a texture and color-based image matching algorithm that queries the image space through the use of ring vectors. In prior studies, the selection of one or more optimal vectors for a particular feature of interest required a manual process, with the user initially stochastically selecting candidate vectors and subsequently testing them upon other regions of the image to verify the vector′s sensitivity and specificity properties (typically by reviewing a resultant heat map). In carrying out the prior efforts, the SIVQ algorithm was noted to exhibit highly scalable computational properties, where each region of analysis can take place independently of others, making a compelling case for the exploration of its deployment on high-throughput computing platforms, with the hypothesis that such an exercise will result in performance gains that scale linearly with increasing processor count. Methods: An automated process was developed for the selection of optimal ring vectors to serve as the predicate matching operator in defining histopathological features of interest. Briefly, candidate vectors were generated from every possible coordinate origin within a user-defined vector selection area (VSA) and subsequently compared against user-identified positive and negative "ground truth" regions on the same image. Each vector from the VSA was assessed for its goodness-of-fit to both the positive and negative areas via the use of the receiver operating characteristic (ROC) transfer function, with each assessment resulting in an associated area-under-the-curve (AUC) figure of merit. Results: Use of the above-mentioned automated vector selection process was demonstrated in two cases of use: First, to identify malignant colonic epithelium, and second, to identify soft tissue sarcoma. For both examples, a very satisfactory optimized vector was identified, as defined by the AUC metric. Finally, as an additional effort directed towards attaining high-throughput capability for the SIVQ algorithm, we demonstrated the successful incorporation of it with the MATrix LABoratory (MATLAB TM ) application interface. Conclusion: The SIVQ algorithm is suitable for automated vector selection settings and high throughput computation.

Background: Measurement of estrogen receptor (ER) expression is a routine part of clinical evaluation of individual breast cancers and is used to guide treatment. However not all ER+ breast cancers show equal benefit from hormonal or other treatment. The RT-PCR based Oncotype Dx assay, has been recently shown to robustly stratify early stage ER+ breast cancer and identify those tumors that will have low recurrence rates when treated with adjuvant hormonal therapy alone. Interestingly, standard pathologic grading, based on visual analysis of tumor morphology by trained pathologists, has a strong correlation with Oncotype Dx recurrence scores; low recurrence tumors are mostly low grade, and high recurrence tumors are mostly high grade. However, a major problem with use of pathologist-assigned histologic grade as a prognostic tool is the lack of reproducibility of histological grading between different pathologists. Computer aided image analysis and machine learning techniques offer a way to obtain highly reproducible image-based classification of ER+ breast cancer. These analyses can be performed on digital images of routinely obtained breast cancer histology and be incorporated into a prognostic assay.Materials and Methods: High resolution digital images were obtained for a series of ER+ breast cancers for which associated Oncotype Dx Assay results were available. Regions of invasive breast cancer were identified and then processed by computer-assisted image analysis methods. An automated nuclear detection scheme based on Expectation Maximization algorithm was used to identify cancer cell nuclei, followed by graph-based feature extraction using the Voronoi graph, Delaunay Triangulation and Minimum Spanning Trees, and dimensionality reduction using Graph Embedding with Support Vector Machine based classification. The final manifold generated by GE was unwrapped into a linear space and a Euclidean distance metric was used to generate a single score (Image Based Risk Score- (IbRiS)) for each sample. Correlations between IbRS, clinical features such as grade, and Oncotype Dx Recurrance Score were determined.Results: Unsupervised analysis of image-based features of high resolution digital images of ER+ breast cancer histology leads to natural separation of tumors, with low grade tumors separating from high grade tumors. There is also a robust separation of tumors with high Oncotype Dx Recurrance Scores from tumors with low Recurrance Scores.Discussion: Unsupervised analysis of high resolution image-based features can stratify ER+ breast cancers in a fashion that correlates well with a gene-expression based prognostic assay, Oncotype Dx. These data suggest that tumors with distinct gene-expression profiles also have distinct image-based features that can be measured by computer-aided image analysis and used to build prognostic and predictive assays. Citation Information: Cancer Res 2009;69(24 Suppl):Abstract nr 3046.

A method and apparatus for automatically detecting breast tumors and lesions in images, including ultrasound, digital and analog mammograms, and MRI images, is provided. An image of a breast is acquired. The image is filtered and contrast of the image is enhanced. Intensity and texture classifiers are applied to each pixel in the image, the classifiers indicative of the probability of the pixel corresponding to a tumor. A seed point is identified within the image, and a region of interest is grown around the seed point. Directional gradients are calculated for each pixel of the image. Boundary points of the region of interest are identified. The boundary points are passed as inputs to a deformable model. The deformable model processes the boundary points to indicate the presence or absence of a tumor.

Linear discriminant analysis (LDA) results for all experiments. Table S2. Diagonal linear discriminant analysis (DLDA) results for all experiments. Table S3. Support vector machine (SVM) results for all experiments. Table S4. NaĂŻve Bayes results for all experiments. Table S5. Quadratic discriminant analysis (QDA) results for all experiments. (DOCX 35.9 kb)

The recent explosion in availability of gene and protein expression data for cancer detection has necessitated the development of sophisticated machine learning tools for high dimensional data analysis. Previous attempts at gene expression analysis have typically used a linear dimensionality reduction method such as Principal Components Analysis (PCA). Linear dimensionality reduction methods do not however account for the inherent nonlinearity within the data. The motivation behind this work is to demonstrate that nonlinear dimensionality reduction methods are more adept at capturing the nonlinearity within the data compared to linear methods, and hence would result in better classification and potentially aid in the visualization and identification of new data classes. Consequently, in this paper, we empirically compare the performance of 3 commonly used linear versus 3 nonlinear dimensionality reduction techniques from the perspective of (a) distinguishing objects belonging to cancer and non-cancer classes and (b) new class discovery in high dimensional gene and protein expression studies for different types of cancer. Quantitative evaluation using a support vector machine and a decision tree classifier revealed statistically significant improvement in classification accuracy by using nonlinear dimensionality reduction methods compared to linear methods.

supplementary methods. (DOCX 12.3 kb)

In this work, we analyze and evaluate different strategies for comparing Feature Selection (FS) schemes on High Dimensional (HD) biomedical datasets (e.g. gene and protein expression studies) with a small sample size (SSS). Additionally, we define a new feature, Robustness, specifically for comparing the ability of an FS scheme to be invariant to changes in its training data. While classifier accuracy has been the de facto method for evaluating FS schemes, on account of the curse of dimensionality problem, it might not always be the appropriate measure for HD/SSS datasets. SSS lends the dataset a higher probability of containing data that is not representative of the true distribution of the whole population. However, an ideal FS scheme must be robust enough to produce the same results each time there are changes to the training data. In this study, we employed the robustness performance measure in conjunction with classifier accuracy (measured via the K-Nearest Neighbor and Random Forest classifiers) to quantitatively compare five different FS schemes (T-test, F-test, Kolmogorov-Smirnov Test, Wilks Lambda Test and Wilcoxon Rand Sum Test) on 5 HD/SSS gene and protein expression datasets corresponding to ovarian cancer, lung cancer, bone lesions, celiac disease, and coronary heart disease. Of the five FS schemes compared, the Wilcoxon Rand Sum Test was found to outperform other FS schemes in terms of classification accuracy and robustness. Our results suggest that both classifier accuracy and robustness should be considered when deciding on the appropriate FS scheme for HD/SSS datasets.

supplementary methods. (DOCX 12.3 kb)

TN-specific and HER2+-specific response prediction. Table S6. Feature Discovery: HER2+ and TN tumors. Additional feature discovery experiments were performed for the subsets of HER2+ patients only (n=28) and TN patients only (n=19). Due to the limited sample size, only 5 top radiomic features were identified for each subtype. Table S7. Validation: HER2+ and TN tumors. The HER2+ and TN-specific signatures were evaluated by three-fold cross-validation of a DLDA classifier within the corresponding subset of patients and assessed by average AUC. Also shown are the results of using the top 5 features identified in the combined TN/HER2+ group for a DLDA classifier in a cross-validation setting within the combined TN/HER2+ group (n=47), HER2+ only group (n=28), and TN only group (n=19). (DOCX 17.2 kb)

Supplementary methods. (DOCX 12.4 kb)

In spite of the high global incidence of prostate cancer, limited computer-aided tools to assist in its detection exist and that too only for ultrasound images. In this work we present a novel feature ensemble scheme for combining different 3D texture features for automated segmentation of prostatic adenocarcinoma from 4T MR images. The first step of our methodology comprises of a feature extraction module to extract 3D statistical, gradient and Gabor texture features at multiple scales and orientations and generate the corresponding Feature Scenes. Every voxel in each of the Feature Scenes is assigned a likelihood of malignancy using a Bayesian inference module. These results are then combined using a novel weighted linear combination scheme; weights being determined by minimization of a cost function. The method was found to be optimal compared to other popular ensemble methods such as Boosting, Majority Rule, Product Rule and Averaging in terms of Sensitivity and Positive Predictive Value (PPV). In fact, our feature ensemble scheme also outperformed an expert radiologist in terms of Sensitivity. An interesting result from the comparison of the different feature ensembles was that Boosting performs poorly on MR data that has been corrected for background in homogeneity.

External beam radiation treatment (EBRT) is a popular method for treating prostate cancer (CaP) involving destroying tumor cells with ionizing radiation. Following EBRT, biochemical failure has been linked with disease recurrence. However, there is a need for methods for evaluating early treatment related changes to allow for an early intervention in case of incomplete disease response. One method for looking at treatment evaluation is to detect changes in MRI markers on a voxel-by-voxel basis following treatment. Changes in MRI markers may be correlated with disease recurrence and complete or partial response. In order to facilitate voxel-by-voxel imaging related treatment changes, and also to evaluate morphologic changes in the gland post treatment, the pre- and post-radiated MRI must first be brought into spatial alignment via image registration. However, EBRT induces changes in the prostate volume and distortion to the internal anatomy of the prostate following radiation treatment. The internal substructures of the prostate, the central gland (CG) and peripheral zone (PZ), may respond to radiation differently, and their resulting shapes may change drastically. Biomechanical models of the prostate that have been previously proposed tend to focus on how external forces affect the surface of the prostate (not the internals), and assume that the prostate is a volume-preserving entity. In this work we present DoCD, a biomechanical model for automatically registering pre-, post-EBRT MRI with the aim of expressly modeling the (1) changes in volume, and (2) changes to the CG and PZ. DoCD was applied to a cohort of 30 patients and achieved a root mean square error of 2.994 mm, which was statistically significantly better a traditional biomechanical model which did not consider changes to the internal anatomy of the prostate (mean of 5.071 mm).

Supplementary methods. (DOCX 12.4 kb)

Magnetic Resonance Imaging (MRI) is known to be significantly affected by a number of acquisition artifacts, such as intensity non-standardness, bias field, and Gaussian noise. These artifacts degrade MR image quality significantly, obfuscating anatomical and physiological detail and hence need to be corrected for to facilitate application of computerized analysis techniques such as segmentation, registration, and classification. Specifically, algorithms are required to correct for bias field (intensity inhomogeneity), intensity non-standardness (drift in tissue intensities across patient acquisitions), and Gaussian noise, an artifact that significantly affects and blurs tissue boundaries (resulting in poor gradients). While clearly one needs to correct for all these artifacts, the exact sequence in which all three operations need to be applied in order to maximize MR image quality has not been explored. In this paper, we empirically evaluate the interplay between distinct algorithms for bias field correction (BFC), intensity standardization (IS), and noise filtering (NF) to study the effect of these operations on image quality in the context of 3 Tesla T2-weighted (T2w) prostate MRI. 7 different sequences comprising combinations of BFC, IS, and NF were quantitatively evaluated in terms of the percent coefficient of variation (%CV), a statistic which attempts to quantify the intensity inhomogeneity within a region of interest (prostate). The different combinations were also independently evaluated in the context of a classifier scheme for detection of prostate cancer on high resolution in vivo T2w prostate MRI. A secondary contribution of this work is a novel evaluation measure for quantifying the level of intensity non-standardness, called difference of modes (DoM). Experimental evaluation of the different sequences of operations across 22 patient datasets revealed that the sequence of BFC, followed by NF, and IS provided the best image quality in terms of %CV as well as classifier accuracy. The DoM measure was able to accurately capture the level of intensity non-standardness present in the images resulting from the different sequences of operations.

In this paper we present a system for detecting regions of carcinoma of the prostate (CaP) in H&E stained radical prostatectomy specimens using the color fractal dimension. Color textural information is known to be a valuable characteristic to distinguish CaP from benign tissue. In addition to color information, we know that cancer tends to form contiguous regions. Our system leverages the color staining information of histology as well as spatial dependencies. The color and textural information is first captured using color fractal dimension. To incorporate spatial dependencies, we combine the probability map constructed via color fractal dimension with a novel Markov prior called the Probabilistic Pairwise Markov Model (PPMM). To demonstrate the capability of this CaP detection system, we applied the algorithm to 27 radical prostatectomy specimens from 10 patients. A per pixel evaluation was conducted with ground truth provided by an expert pathologist using only the color fractal feature first, yielding an area under the receiver operator characteristic curve (AUC) curve of 0.790. In conjunction with a Markov prior, the resultant color fractal dimension + Markov random field (MRF) classifier yielded an AUC of 0.831.

Breast ultrasound (US) in conjunction with digital mammography has come to be regarded as the gold standard for breast cancer diagnosis. While breast US has certain advantages over digital mammography, it suffers from image artifacts such as posterior acoustic shadowing (PAS), presence of which often obfuscates lesion margins. Since classification of lesions as either malignant or benign is largely dictated by lesion's shape and margin characteristics, it is important to distinguish lesion area from PAS. This paper represents the first attempt to extract and identify those image features that can help distinguish between lesion and PAS. Our method- ology comprises of extracting over 100 statistical, gradient, and Gabor features at multiple scales and orientations at every pixel in the breast US image. Adaboost, a powerful feature ensemble technique is used to discriminate between lesions and PAS by combining the different image features. A non-linear dimensionality reduction method called Graph Embedding is then used to visualize separation and inter-class dependencies between lesions and PAS in a lower dimensional space. Results of quantitative evaluation on a database of 45 breast US images indicate that our methodology allows for greater discriminability between the lesion and PAS classes compared to that achievable by any individual image texture or intensity feature. Index Terms— Breast, ultrasound, posterior acoustic shad- owing, AdaBoost, non-linear dimensionality reduction, lesion.

TN-specific and HER2+-specific response prediction. Table S6. Feature Discovery: HER2+ and TN tumors. Additional feature discovery experiments were performed for the subsets of HER2+ patients only (n=28) and TN patients only (n=19). Due to the limited sample size, only 5 top radiomic features were identified for each subtype. Table S7. Validation: HER2+ and TN tumors. The HER2+ and TN-specific signatures were evaluated by three-fold cross-validation of a DLDA classifier within the corresponding subset of patients and assessed by average AUC. Also shown are the results of using the top 5 features identified in the combined TN/HER2+ group for a DLDA classifier in a cross-validation setting within the combined TN/HER2+ group (n=47), HER2+ only group (n=28), and TN only group (n=19). (DOCX 17.2 kb)

Background Automated classification of histopathology involves identification of multiple classes, including benign, cancerous, and confounder categories. The confounder tissue classes can often mimic and share attributes with both the diseased and normal tissue classes, and can be particularly difficult to identify, both manually and by automated classifiers. In the case of prostate cancer, they may be several confounding tissue types present in a biopsy sample, posing as major sources of diagnostic error for pathologists. Two common multi-class approaches are one-shot classification (OSC), where all classes are identified simultaneously, and one-versus-all (OVA), where a “target” class is distinguished from all “non-target” classes. OSC is typically unable to handle discrimination of classes of varying similarity (e.g. with images of prostate atrophy and high grade cancer), while OVA forces several heterogeneous classes into a single “non-target” class. In this work, we present a cascaded (CAS) approach to classifying prostate biopsy tissue samples, where images from different classes are grouped to maximize intra-group homogeneity while maximizing inter-group heterogeneity. Results We apply the CAS approach to categorize 2000 tissue samples taken from 214 patient studies into seven classes: epithelium, stroma, atrophy, prostatic intraepithelial neoplasia (PIN), and prostate cancer Gleason grades 3, 4, and 5. A series of increasingly granular binary classifiers are used to split the different tissue classes until the images have been categorized into a single unique class. Our automatically-extracted image feature set includes architectural features based on location of the nuclei within the tissue sample as well as texture features extracted on a per-pixel level. The CAS strategy yields a positive predictive value (PPV) of 0.86 in classifying the 2000 tissue images into one of 7 classes, compared with the OVA (0.77 PPV) and OSC approaches (0.76 PPV). Conclusions Use of the CAS strategy increases the PPV for a multi-category classification system over two common alternative strategies. In classification problems such as histopathology, where multiple class groups exist with varying degrees of heterogeneity, the CAS system can intelligently assign class labels to objects by performing multiple binary classifications according to domain knowledge.

Esophageal adenocarcinoma arises from Barrett's esophagus, which is the most serious complication of gastroesophageal reflux disease. Strategies for screening involve periodic surveillance and tissue biopsies. A major challenge in such regular examinations is to record and track the disease evolution and re-localization of biopsied sites to provide targeted treatments. In this paper, we extend our original inter-operative relocalization framework to provide a constrained image based search for obtaining the best view-point match to the live view. Within this context we investigate the effect of: the choice of feature descriptors and color-space; filtering of uninformative frames and endoscopic modality, for view-point localization. Our experiments indicate an improvement in the best view-point retrieval rate to [92%,87%] from [73%,76%] (in our previous approach) for NBI and WL., Medical Content-based Retrieval for Clinical Decision Support and Treatment Planning, MICCAI Conference