Previously, several instances of variability in the output of pattern classifiers that have the same Receiver Operating
Characteristic (ROC) curve have been observed. We present a theoretical framework for understanding some sources of
this variability, which result in classifiers with monotonically related outputs. We restrict our analysis to pattern
classifiers that discriminate between two linearly separable classes. We show that variability in the output of pattern
classifiers can arise due to differences in the functional mappings between their inputs and outputs. We further identify
some practical situations wherein such variability in the output of such pattern classifiers arises. These include situations
in which there are differences in (a) the datasets employed for training and evaluation of classifiers, (b) the a priori
probabilities of the two classes, or (c) the stochastic processes employed for training the different pattern classifiers.
Previously, we proposed a technique based on the matching of the histograms of differently distributed classifier output
to reduce the variability in their diagnostic performance and their output values. Here, we prove theoretically and
demonstrate empirically on simulated data, that for monotonically related classifier outputs, this technique successfully
learns the true monotonic transformation function that exists between different pattern classifier outputs.
Proc. SPIE. 6515, Medical Imaging 2007: Image Perception, Observer Performance, and Technology Assessment
KEYWORDS: Computer aided diagnosis and therapy, Data modeling, Calibration, Image processing, Computing systems, Receivers, Medical imaging, Artificial neural networks, Image enhancement, Probability theory
In this study we developed an effective novel method for reducing the variability in the output of different artificial neural network (ANN) configurations that have the same overall performance as measured by the area under their receiver operating characteristic (ROC) curves. This variability can lead to inaccuracies in the interpretation of results when the outputs are employed as classification predictors. We extended a method previously proposed to reduce the variability in the performance of a classifier with data sets from different institutions to the outputs of ANN configurations. Our approach is based on histogram shaping of the outputs of all ANN configurations to resemble the output histogram of a baseline ANN configuration. We tested the effectiveness of the technique using synthetic data generated from two two-dimensional isotropic Gaussian distributions and 100 ANN configurations. The proposed output calibration technique significantly reduced the median standard deviation of the ANN outputs from 0.010 before calibration to 0.006 after calibration. The standard deviation of the sensitivity of the 100 ANN configurations at the same decision threshold reduced significantly from 0.005 before calibration to 0.003 after calibration. Similarly the standard deviation of their specificity values decreased significantly from 0.016 before calibration to 0.006 after calibration.
KEYWORDS: Facial recognition systems, Principal component analysis, 3D acquisition, Detection and tracking algorithms, Data modeling, 3D modeling, Feature selection, Algorithm development, Performance modeling, 3D image processing
We propose a novel method to improve the performance of existing three dimensional (3D) human face recognition algorithms that employ Euclidean distances between facial fiducial points as features. We further investigate a novel 3D face recognition algorithm that employs geodesic and Euclidean distances between facial fiducial points. We demonstrate that this algorithm is robust to changes in facial expression. Geodesic and Euclidean distances were calculated between pairs of 25 facial fiducial points. For the proposed algorithm, geodesic distances and 'global curvature' characteristics, defined as the ratio of geodesic to Euclidean distance between a pairs of points, were employed as features. The most discriminatory features were selected using stepwise linear discriminant analysis (LDA). These were projected onto 11 LDA directions, and face models were matched using the Euclidean distance metric. With a gallery set containing one image each of 105 subjects and a probe set containing 663 images of the same subjects, the algorithm produced EER=1.4% and a rank 1 RR=98.64%. It performed significantly better than existing algorithms based on principal component analysis and LDA applied to face range images. Its verification performance for expressive faces was also significantly better than an algorithm that employed Euclidean distances between facial fiducial points as features.
The purpose of this study was to investigate approaches for combining information from the MLO and CC mammographic views for Computer-aided Diagnosis (CADx) algorithms. Feature level and classifier output level combinations were explored. Linear discriminant analysis (LDA) with step-wise feature selection from a set of Haralick's texture features was used to develop classifiers for distinguishing between benign and malignant mammographic lesions. The effect of correlation between features from the two views on the performance of classifiers was investigated. The single view models included: (a) an LDA model with stepwise selection based on the MLO view only (MLO-Only) and similarly (b) a CC-Only LDA model. The feature-level combination models included: (a) LDA based on concatenation of feature sets selected independently from the two views (FEAT_CON), (b) LDA based on the concatenated feature sets along with the corresponding value of each feature from the opposite view (FEAT_COR_CON) if the correlation was below a threshold, (c) LDA based on the average of the MLO and CC feature values (FEAT_AVG). The classifier output level combination models investigated included: (a) average of the outputs of the MLO-Only and CC-Only classifiers (OUTPUT_AVG), (b) maximum of the outputs of the MLO-Only and CC-Only classifiers (OUTPUT_MAX), (c) minimum of the outputs of the MLO-Only and CC-Only classifiers (OUTPUT_MIN), (d) a second level LDA classifier on the outputs of the MLO-Only and CC-Only classifiers (OUTPUT_LDA), (e) product of the output values of the two classifiers (OUTPUT_PROD). The performance of the models was assessed and compared using the ROC methodology to determine if combination models performed better than the single-view models.
This work focuses on the development of materials and growth techniques suitable for future spintronic device applications. Metal-organic chemical vapor deposition (MOCVD) was used to grow high-quality epitaxial films of varying thickness and manganese doping levels by introducing bis-cyclopentadienyl as the manganese source. High-resolution X-ray diffraction indicates that no macroscopic second phases are formed during growth, and Mn containing films are similar in crystalline quality to undoped films Atomic force microscopy revealed a 2-dimensional MOCVD step-flow growth pattern in the Mn-incorporated samples. The mean surface roughnesses of optimally grown Ga1-xMnxN films were almost identical to that from the as-grown template layers, with no change in growth mechanism or morphology. Various annealing steps were applied to some of the samples to reduce compensating defects and to
investigate the effects of post processing on the structural, magnetic and opto-electronic properties. SQUID measurements showed an apparent ferromagnetic hysteresis behavior which persisted to room temperature. An optical absorption band around 1.5 eV was observed via transmission studies. This band is assigned to the internal Mn3+ transition between the 5E and the partially filled 5T2 levels of the 5D state. The broadening of the absorption band is
introduced by the high Mn concentration. Recharging of the Mn3+ to Mn2+ was found to effectively suppress these transitions resulting in a reduction of the magnetization. The structural quality, and the presence of Mn2+ ions were
confirmed by EPR spectroscopy, meanwhile no Mn-Mn interactions indicative of clustering were observed. The absence of doping-induced strain in Ga1-xMnxN was observed by Raman spectroscopy.