Derek C. Stanford

Approximate Bayes factors for image segmentation: the Pseudolikelihood Information Criterion (PLIC)

We propose a method for choosing the number of colors or true gray levels in an image; this allows fully automatic segmentation of images. Our underlying probability model is a hidden Markov random field. Each number of colors considered is viewed as corresponding to a statistical model for the image, and the resulting models are compared via approximate Bayes factors. The Bayes factors are approximated using BIC (Bayesian Information Criterion), where the required maximized likelihood is approximated by the Qian-Titterington (1991) pseudolikelihood. We call the resulting criterion PLIC (Pseudolikelihood Information Criterion). We also discuss a simpler approximation, MMIC (Marginal Mixture Information Criterion), which is based only on the marginal distribution of pixel values. This turns out to be useful for initialization and it also has moderately good performance by itself when the amount of spatial dependence in an image is low. We apply PLIC and MMIC to a medical image segmentation problem.

Glaucoma screening using the scanning laser polarimeter.

Purpose: To determine the ability of scanning laser polarimetry (GDx Nerve Fiber Analyzer; Laser Diagnostic Technologies, Inc., San Diego, CA) to separate normal eyes and those considered likely to have glaucoma in a public glaucoma screening. Methods: A 2‐day public glaucoma screening program was held at two different institutions. Each subject underwent ophthalmologic examination, Humphrey perimetry (24–2 Fastpac program), and imaging using scanning laser polarimetry (GDx) in each eye for allocation into a diagnostic category: normal, ocular hypertensive, glaucoma suspect, or glaucoma. Results from the normal and glaucoma groups were analyzed, using modulation parameters calculated from a measurement band located 1.8 disc diameters from the disc, and selected parameters provided automatically by GDx software. Receiver operating characteristic curves were used to depict the sensitivity/specificity relationship at different GDx parameter cutoff levels. Results: Of 200 subjects, 197 were classified; 122 were classified as normal, 23 were classified with ocular hypertension, 30 were classified as glaucoma suspects, and 22 were classified with definite glaucoma. Three subjects had ocular diseases other than glaucoma. The maximum area under the receiver operating characteristic curve for modulation parameters was 0.935, and for the GDx software parameters was 0.901. Conclusions: Scanning laser polarimetry may be useful in glaucoma screening.

Model‐based methods for textile fault detection

Addressing the problem of automatic fault detection in woven and dyed fabric, we discuss a number of new statistical model‐based methods and relate them to a first stage of point/local detection and a second stage of extended pattern detection. One model‐based method defines a maximum likelihood binarization of the image. In another model‐based method, we describe a discrete Fourier transform‐based texture analysis technique that is highly effective for woven textiles in discriminating subtle flaw patterns from the pronounced background of repetitive weaving pattern and random clutter. Finally, we describe a model‐based clustering method that can be employed to aggregate perceptual groupings of point and local detections.

Mechanisms of Bone Marrow-Derived Cell Therapy in Ischemic Cardiomyopathy with Left Ventricular Assist Device Bridge to Transplant

BACKGROUND Clinical trials report improvements in function and perfusion with direct injection of bone marrow cells into the hearts of patients with ischemic cardiomyopathy. Preclinical data suggest these cells improve vascular density, which would be expected to decrease fibrosis and inflammation. OBJECTIVES The goal of this study was to test the hypothesis that bone marrow stem cells (CD34+) will improve histological measurements of vascularity, fibrosis, and inflammation in human subjects undergoing left ventricular assist device (LVAD) placement as a bridge to cardiac transplantation. METHODS Subjects with ischemic cardiomyopathy who were scheduled for placement of an LVAD as a bridge to transplantation underwent bone marrow aspiration the day before surgery; the bone marrow was processed into cell fractions (bone marrow mononuclear cells, CD34+, and CD34-). At LVAD implantation, all fractions and a saline control were injected epicardially into predetermined areas and each injection site marked. At the time of transplantation, injected areas were collected. Data were analyzed by paired Student t test comparing the effect of cell fractions injected within each subject. RESULTS Six subjects completed the study. There were no statistically significant differences in complications with the procedure versus control subjects. Histological analysis indicated that myocardium injected with CD34+ cells had decreased density of endothelial cells compared to saline-injected myocardium. There were no significant differences in fibrosis or inflammation between groups; however, density of activated fibroblasts was decreased in both CD34+ and CD34- injected areas. CONCLUSIONS Tissue analysis does not support the hypothesis that bone marrow-derived CD34+ cells promote increased vascular tissue in humans with ischemic cardiomyopathy via direct injection.

Probability maps of sequential Glaucoma-Scope images help identify significant change

PurposeThe purpose of this study was to identify areas of the optic disc showing high variability of repeated depth measurements, and to minimize the effect of baseline variability in interpretation of possible change over time using the Glaucoma-Scope. MethodsSeventy-four eyes from 70 subjects were analyzed with the Glaucoma-Scope. Three images were obtained on each of two separate sessions during the same day. At each location, the mean depth of the three images for each session was calculated to create a “baseline image.” A contour map of standard deviation (SD) values at each topographic location was created for each subject reflecting local variability at different parts of the disc. The contour map and disc photograph were compared to determine what photographic features predicted high variability. A modified two-sample t-test was used at each topographic location to obtain p-values for the likelihood that a difference in mean depth between sessions was attributable to measurement variability alone. ResultsContour plots of SD for most subject eyes showed high variability in steeply sloped areas of the disc and along large blood vessels, with low variability near the cup center. The use of probability plots for significance of depth changes between test sessions automatically accounted for increased pointwise variability. The proportion of topographic locations showing statistically significant change but attributable to chance variation when no true change has occurred approximated the predicted proportion based on our modified t-test model. ConclusionA contour map of standard deviations of depth based on Glaucoma-Scope baseline images can identify areas of the disc with high variability. Statistical methods such as probability maps that account for local variability in the baseline image may be helpful in distinguishing true change from artefactual change over time.

Detection of Optic Disc Changes with Glaucoma-Scope Probability Maps

PurposeTo test whether a statistical method using a probability map could detect true changes in optic disc topography. MethodsThe average of three Glaucoma-Scope images (Ophthalmic Imaging Systems were used for analysis at each of two sessions. A Glaucoma-Scope probability map was constructed for each eye using statistical methods. The proportion of topographic locations with p values less than 0.05 on a modified two-sample t test (p-proportion) and the difference in the mean position of the disc (MPD) from two imaging sessions were calculated. Two pairs of stereoscopic disc photographs for 43 eyes with longitudinal follow-up were evaluated for change by four experienced glaucoma specialists masked to patient clinical information. Clinical change was considered to have occurred when the assessments of at least three of the four specialists were agreed on. The cutoff values for p-proportion and change in MPD that provided 95% specificity were calculated using a separate sample of 69 subjects who had serial images taken at two separate sessions on the same day, and thus showed no clinical change in the optic disc. ResultsThe cutoff values of 95% specificity for the p-proportion and the change in MPD were 18% and 25.1 μm, respectively. Of 43 eyes with longitudinal follow-up, 14 showed definite clinical change. Sensitivity of the p-proportion and change in MPD for detecting this change was 100% and 85.7%, respectively. For all 43 patients with longitudinal follow-up, the percent change in intraocular pressure (IOP) correlated strongly with both the p-proportion and the change in MPD. ConclusionUsing data obtained with the Glaucoma-Scope, a statistical method based on probability mapping can be used to detect true changes in disc topography. The p-proportion was more sensitive than change in MPD in detecting clinical change in the study eyes. This statistical methodology may also be applicable for interpretation of data obtained with other optic disc analyzers.