Near-infrared fluorescent digital pathology for the automation of disease diagnosis and biomarker assessment.

Hematoxylin-eosin (H&E) staining of tissue has been the mainstay of pathology for more than a century. However, the learning curve for H&E tissue interpretation is long, whereas intra- and interobserver variability remain high. Computer-assisted image analysis of H&E sections holds promise for increased throughput and decreased variability but has yet to demonstrate significant improvement in diagnostic accuracy. Addition of biomarkers to H&E staining can improve diagnostic accuracy; however, coregistration of immunohistochemical staining with H&E is problematic as immunostaining is completed on slides that are at best 4 µm apart. Simultaneous H&E and immunostaining would alleviate coregistration problems; however, current opaque pigments used for immunostaining obscure H&E. In this study, we demonstrate that diagnostic information provided by two or more independent wavelengths of near-infrared (NIR) fluorescence leave the H&E stain unchanged while enabling computer-assisted diagnosis and assessment of human disease. Using prostate cancer as a model system, we introduce NIR digital pathology and demonstrate its utility along the spectrum from prostate biopsy to whole mount analysis of H&E-stained tissue.

Global error minimization in image mosaicing using graph connectivity and its applications in microscopy.

Several applications such as multiprojector displays and microscopy require the mosaicing of images (tiles) acquired by a camera as it traverses an unknown trajectory in 3D space. A homography relates the image coordinates of a point in each tile to those of a reference tile provided the 3D scene is planar. Our approach in such applications is to first perform pairwise alignment of the tiles that have imaged common regions in order to recover a homography relating the tile pair. We then find the global set of homographies relating each individual tile to a reference tile such that the homographies relating all tile pairs are kept as consistent as possible. Using these global homographies, one can generate a mosaic of the entire scene. We derive a general analytical solution for the global homographies by representing the pair-wise homographies on a connectivity graph. Our solution can accommodate imprecise prior information regarding the global homographies whenever such information is available. We also derive equations for the special case of translation estimation of an X-Y microscopy stage used in histology imaging and present examples of stitched microscopy slices of specimens obtained after radical prostatectomy or prostate biopsy. In addition, we demonstrate the superiority of our approach over tree-structured approaches for global error minimization.

Ideal AFROC and FROC observers.

Detection of multiple lesions in images is a medically important task and free-response receiver operating characteristic (FROC) analyses and its variants, such as alternative FROC (AFROC) analyses, are commonly used to quantify performance in such tasks. However, ideal observers that optimize FROC or AFROC performance metrics have not yet been formulated in the general case. If available, such ideal observers may turn out to be valuable for imaging system optimization and in the design of computer aided diagnosis techniques for lesion detection in medical images. In this paper, we derive ideal AFROC and FROC observers. They are ideal in that they maximize, amongst all decision strategies, the area, or any partial area, under the associated AFROC or FROC curve. Calculation of observer performance for these ideal observers is computationally quite complex. We can reduce this complexity by considering forms of these observers that use false positive reports derived from signal-absent images only. We also consider a Bayes risk analysis for the multiple-signal detection task with an appropriate definition of costs. A general decision strategy that minimizes Bayes risk is derived. With particular cost constraints, this general decision strategy reduces to the decision strategy associated with the ideal AFROC or FROC observer.

COMPUTER-AIDED GLEASON GRADING OF PROSTATE CANCER HISTOPATHOLOGICAL IMAGES USING TEXTON FORESTS.

The Gleason score is the single most important prognostic indicator for prostate cancer candidates and plays a significant role in treatment planning. Histopathological imaging of prostate tissue samples provides the gold standard for obtaining the Gleason score, but the manual assignment of Gleason grades is a labor-intensive and error-prone process. We have developed a texture classification system for automatic and reproducible Gleason grading. Our system characterizes the texture in images belonging to a tumor grade by clustering extracted filter responses at each pixel into textons (basic texture elements). We have used random forests to cluster the filter responses into textons followed by the spatial pyramid match kernel in conjunction with an SVM classifier. We have demonstrated the efficacy of our system in distinguishing between Gleason grades 3 and 4.

Quantification of brain maturation and growth patterns in C57BL/6J mice via computational neuroanatomy of diffusion tensor images.

Diffusion Tensor magnetic resonance imaging and computational neuroanatomy are used to quantify postnatal developmental patterns of C57BL/6J mouse brain. Changes in neuronal organization and myelination occurring as the brain matures into adulthood are examined, and a normative baseline is developed, against which transgenic mice may be compared in genotype-phenotype studies. In early postnatal days, gray matter-based cortical and hippocampal structures exhibit high water diffusion anisotropy, presumably reflecting the radial neuronal organization. Anisotropy drops rapidly within a week, indicating that the underlying brain tissue becomes more isotropic in orientation, possibly due to formation of a complex randomly intertwined web of dendrites. Gradual white matter anisotropy increase implies progressively more organized axonal pathways, likely reflecting the myelination of axons forming tightly packed fiber bundles. In contrast to the spatially complex pattern of tissue maturation, volumetric growth is somewhat uniform, with the cortex and the cerebellum exhibiting slightly more pronounced growth. Temporally, structural growth rates demonstrate an initial rapid volumetric increase in most structures, gradually tapering off to a steady state by about 20 days. Fiber maturation reaches steady state in about 10 days for the cortex, to 30-40 days for the corpus callosum, the hippocampus, and the internal and external capsules.

Aperture optimization in emission imaging using ideal observers for joint detection and localization.

For the familiar 2-class detection problem (signal present/absent), ideal observers have been applied to optimization of pinhole and collimator parameters in planar emission imaging. Given photon noise and background and signal variabilities, such experiments show how to optimize an aperture to maximize detectability of the signal. Here, we consider a fundamentally different, more realistic task in which the observer is required to both detect and localize a signal. The signal is embedded in a variable background and is known except for location. We inquire whether the addition of a localization requirement changes conclusions on aperture optimization. We have previously formulated an ideal observer for this joint detection/localization task, and here apply it to the classic problem of determining an optimal pinhole diameter in a planar emission imaging system. We conclude that as search tolerance on localization decreases, the optimal pinhole diameter shrinks from that required by detection alone, and, in addition, task performance becomes more sensitive to fluctuations about the optimal pinhole diameter. As in the case for detection only, the optimal pinhole diameter shrinks as the amount of background variability grows and, in addition, conspicuity limits can be observed. Unlike the case for detection only, our task leads to a finite aperture size in the absence of background variability. For both tasks, the inclusion of background variability yields a finite aperture size.

On analyzing diffusion tensor images by identifying manifold structure using isomaps.

This paper addresses the problem of statistical analysis of diffusion tensor magnetic resonance images (DT-MRI). DT-MRI cannot be analyzed by commonly used linear methods, due to the inherent nonlinearity of tensors, which are restricted to lie on a nonlinear submanifold of the space in which they are defined, namely R6. We estimate this submanifold using the Isomap manifold learning technique and perform tensor calculations using geodesic distances along this manifold. Multivariate statistics used in group analyses also use geodesic distances between tensors, thereby warranting that proper estimates of means and covariances are obtained via calculations restricted to the proper subspace of R6. Experimental results on data with known ground truth show that the proposed statistical analysis method properly captures statistical relationships among tensor image data, and it identifies group differences. Comparisons with standard statistical analyses that rely on Euclidean, rather than geodesic distances, are also discussed.

Kernel-based manifold learning for statistical analysis of diffusion tensor images.

Diffusion tensor imaging (DTI) is an important modality to study white matter structure in brain images and voxel-based group-wise statistical analysis of DTI is an integral component in most biomedical applications of DTI. Voxel-based DTI analysis should ideally satisfy two desiderata: (1) it should obtain a good characterization of the statistical distribution of the tensors under consideration at a given voxel, which typically lie on a non-linear submanifold of R6, and (2) it should find an optimal way to identify statistical differences between two groups of tensor measurements, e.g., as in comparative studies between normal and diseased populations. In this paper, extending previous work on the application of manifold learning techniques to DTI, we shall present a kernel-based approach to voxel-wise statistical analysis of DTI data that satisfies both these desiderata. Using both simulated and real data, we shall show that kernel principal component analysis (kPCA) can effectively learn the probability density of the tensors under consideration and that kernel Fisher discriminant analysis (kFDA) can find good features that can optimally discriminate between groups. We shall also present results from an application of kFDA to a DTI dataset obtained as part of a clinical study of schizophrenia.

Rapid Optimization of SPECT Scatter Correction Using Model LROC Observers.

The problem we address is the optimization and comparison of window-based scatter correction (SC) methods in SPECT for maximum a posteriori reconstructions. While sophisticated reconstruction-based SC methods are available, the commonly used window-based SC methods are fast, easy to use, and perform reasonably well. Rather than subtracting a scatter estimate from the measured sinogram and then reconstructing, we use an ensemble approach and model the mean scatter sinogram in the likelihood function. This mean scatter sinogram estimate, computed from satellite window data, is itself inexact (noisy). Therefore two sources of noise, that due to Poisson noise of unscattered photons and that due to the model error in the scatter estimate, are propagated into the reconstruction. The optimization and comparison is driven by a figure of merit, the area under the LROC curve (ALROC) that gauges performance in a signal detection plus localization task. We use model observers to perform the task. This usually entails laborious generation of many sample reconstructions, but in this work, we instead develop a theoretical approach that allows one to rapidly compute ALROC given known information about the imaging system and the scatter correction scheme. A critical step in the theory approach is to predict additional (above that due to to the propagated Poisson noise of the primary photons) contributions to the reconstructed image covariance due to scatter (model error) noise. Simulations show that our theory method yields, for a range of search tolerances, LROC curves and ALROC values in close agreement to that obtained using model observer responses obtained from sample reconstruction methods. This opens the door to rapid comparison of different window-based SC methods and to optimizing the parameters (including window placement and size, scatter sinogram smoothing kernel) of the SC method.

An overview of fast convergent ordered-subsets reconstruction methods for emission tomography based on the incremental EM algorithm.

Statistical reconstruction has become popular in emission computed tomography but suffers slow convergence (to the MAP or ML solution). Methods proposed to address this problem include the fast but non-convergent OSEM and the convergent RAMLA [1] for the ML case, and the convergent BSREM [2], relaxed OS-SPS and modified BSREM [3] for the MAP case. The convergent algorithms required a user-determined relaxation schedule. We proposed fast convergent OS reconstruction algorithms for both ML and MAP cases, called COSEM (Complete-data OSEM), which avoid the use of a relaxation schedule while maintaining convergence. COSEM is a form of incremental EM algorithm. Here, we provide a derivation of our COSEM algorithms and demonstrate COSEM using simulations. At early iterations, COSEM-ML is typically slower than RAMLA, and COSEM-MAP is typically slower than optimized BSREM while remaining much faster than conventional MAP-EM. We discuss how COSEM may be modified to overcome these limitations.

Decision strategies that maximize the area under the LROC curve.

For the 2-class detection problem (signal absent/present), the likelihood ratio is an ideal observer in that it minimizes Bayes risk for arbitrary costs and it maximizes the area under the receiver operating characteristic (ROC) curve [AUC]. The AUC-optimizing property makes it a valuable tool in imaging system optimization. If one considered a different task, namely, joint detection and localization of the signal, then it would be similarly valuable to have a decision strategy that optimized a relevant scalar figure of merit. We are interested in quantifying performance on decision tasks involving location uncertainty using the localization ROC (LROC) methodology. Therefore, we derive decision strategies that maximize the area under the LROC curve, A(LROC). We show that these decision strategies minimize Bayes risk under certain reasonable cost constraints. The detection-localization task is modeled as a decision problem in three increasingly realistic ways. In the first two models, we treat location as a discrete parameter having finitely many values resulting in an (L + 1) class classification problem. In our first simple model, we do not include search tolerance effects and in the second, more general, model, we do. In the third and most general model, we treat location as a continuous parameter and also include search tolerance effects. In all cases, the essential proof that the observer maximizes A(LROC) is obtained with a modified version of the Neyman-Pearson lemma. A separate form of proof is used to show that in all three cases, the decision strategy minimizes the Bayes risk under certain reasonable cost constraints.

Fast LROC analysis of Bayesian reconstructed emission tomographic images using model observers.

Lesion detection and localization is an important task in emission computed tomography. Detection and localization performance with signal location uncertainty may be summarized by a scalar figure of merit, the area under the localization receiver operating characteristic (LROC) curve, A(LROC). We consider model observers to compute A(LROC) for two-dimensional maximum a posteriori (MAP) reconstructions. Model observers may be used to rapidly prototype studies that use human observers. We address the case background-known-exactly (BKE) and signal known except for location. Our A(LROC) calculation makes use of theoretical expressions for the mean and covariance of the reconstruction and, unlike conventional methods that also use model observers, does not require computation of a large number of sample reconstructions. We validate the results of the procedure by comparison to A(LROC) obtained using a gold-standard Monte Carlo method employing a large set of reconstructed noise samples. Under reasonable simulation conditions, our theoretical calculation is about one to two orders of magnitude faster than the conventional Monte Carlo method.

An accelerated convergent ordered subsets algorithm for emission tomography.

We propose an algorithm, E-COSEM (enhanced complete-data ordered subsets expectation-maximization), for fast maximum likelihood (ML) reconstruction in emission tomography. E-COSEM is founded on an incremental EM approach. Unlike the familiar OSEM (ordered subsets EM) algorithm which is not convergent, we show that E-COSEM converges to the ML solution. Alternatives to the OSEM include RAMLA, and for the related maximum a posteriori (MAP) problem, the BSREM and OS-SPS algorithms. These are fast and convergent, but require ajudicious choice of a user-specified relaxation schedule. E-COSEM itself uses a sequence of iteration-dependent parameters (very roughly akin to relaxation parameters) to control a tradeoff between a greedy, fast but non-convergent update and a slower but convergent update. These parameters are computed automatically at each iteration and require no user specification. For the ML case, our simulations show that E-COSEM is nearly as fast as RAMLA.

Rapid Computation of LROC Figures of Merit Using Numerical Observers (for SPECT/PET Reconstruction).

The assessment of PET and SPECT image reconstructions by image quality metrics is typically time consuming, even if methods employing model observers and samples of reconstructions are used to replace human testing. We consider a detection task where the background is known exactly and the signal is known except for location. We develop theoretical formulae to rapidly evaluate two relevant figures of merit, the area under the LROC curve and the probability of correct localization. The formulae can accommodate different forms of model observer. The theory hinges on the fact that we are able to rapidly compute the mean and covariance of the reconstruction. For four forms of model observer, the theoretical expressions are validated by Monte Carlo studies for the case of MAP (maximum a posteriori) reconstruction. The theory method affords a 10(2) - 10(3) speedup relative to methods in which model observers are applied to sample reconstructions.