Robert F. Wagner

Laboratory measurements of sensitometry, MTF, veiling glare, Wiener spectrum and DQE for image intensifier tubes

Several different test protocols were used to measure the sensitometric response, the modulation transfer function (MTF), the veiling glare and the Wiener spectrum of two x-ray image intensifier (II) tubes. These data provided the means to calculate summary measures of imaging performance, i.e., the noise equivalent quanta, NEQ, or the detective quantum efficiency, DQE, as a function of spatial frequency. Results are presented to show the differences between an older versus a newer generation II tube, i.e., DQE equals 0.3 +/- 0.03 and 0.6 +/- 0.06 at 0.5 lp/mm, respectively. The eventual goal of this work is to achieve some consensus on methodology for these measurements and to validate the use of algorithmic observers in quantitating imaging performance in the clinical environment.

Application of pattern recognition techniques in ultrasound tissue characterization

An approach for the application of multivariate pattern recognition techniques to detection of diffuse and focal disease using acoustic data is reviewed. Supervised and unsupervised techniques are implemented to design the best ultrasonic tissue signature for a given task from a set of measurements. The performances of both techniques are evaluated and compared using several methods. However, it is desirable to utilize a technique that quantitatively detects and displays the heterogeneity of an ultrasound image. It is shown that, for a particular task, choosing features with physical significance will make the classification of the data more robust. It is also shown that the success of combining supervised and unsupervised techniques using such features extends beyond discrimination of one class of data from the other and that the approach can be used to grade and the variations in the same tissue type.<<ETX>>

Rayleigh task performance in tomographic reconstructions: comparison of human and machine performance

We have previously described how imaging systems and image reconstruction algorithms can be evaluated based on the ability of machine and human observers to perform a binary- discrimination task using the resulting images. Machine observers used in these investigations have been based on approximations to the ideal observer of Bayesian statistical decision theory. The present work is an evaluation of tomographic images reconstructed from a small number of views using the Cambridge Maximum Entropy software, MEMSYS 3. We compare the performance of machine and human viewers for the Rayleigh resolution task. Our results indicate that for both humans and machines a broad latitude exists in the choice of the parameter (alpha) that determines the smoothness of the reconstructions. We find human efficiency relative to the best machine observer to be approximately constant across the range of (alpha) values studied. The close correspondence between human and machine performance that we have now obtained over a variety of tasks indicate that our evaluation of imaging systems based on machine observers has relevance when the images are intended for human use.

Progress In Signal And Texture Discrimination In Medical Imaging

We begin with a review of the concepts of first, second, and higher order statistics and the ability of human observers to extract textural information of these orders from images. This ability has been found to be very high for first order statistics and very low for third and higher order statistics. We next explore some classes of second order statistics where the human observer is greatly outperformed by machine analysis and explain this within the texton theory of Julesz. Example images from phase-sensitive detection systems such as medical ultrasound are then presented. The signal detection theory used previously to study the detectability of first order changes in images is generalized to analyze the detectability and classification of textural changes within an image. We conclude that second order statistical properties contain a wealth of unused information that can be easily extracted both for system performance evaluation and for classification of tissue textural changes.

Task performance on constrained reconstructions: human observer performance compared with suboptimal Bayesian performance

We have previously described how imaging systems and image reconstruction algorithms can be evaluated on the basis of how well binary-discrimination tasks can be performed by a machine algorithm that `views the reconstructions. Algorithms used in these investigations have been based on approximations to the ideal observer of Bayesian statistical decision theory. The present work examines the performance of an extended family of such algorithmic observers viewing tomographic images reconstructed from a small number of views using the Cambridge Maximum Entropy software, MEMSYS 3. We investigate the effects on the performance of these observers due to varying the parameter (alpha) ; this parameter controls the stopping point of the iterative reconstruction technique and effectively determines the smoothness of the reconstruction. For the detection task considered here, performance is maximum at the lowest values of (alpha) studied; these values are encountered as one moves toward the limit of maximum likelihood estimation while maintaining the positivity constraint intrinsic to entropic priors. A breakdown in the validity of a Gaussian approximation used by one of the machine algorithms (the posterior probability) was observed in this region. Measurements on human observers performing the same task show that they perform comparably to the best machine observers in the region of highest machine scores, i.e., smallest values of (alpha) . For increasing values of (alpha) , both human and machine observer performance degrade. The falloff in human performance is more rapid than that of the machine observer at the largest values of (alpha) (lowest performance) studied. This behavior is common to all such studies of the so-called psychometric function.

Prewhitening matched filter: practical implementation, SNR estimation and bias reduction

In the clinical setting, image quality is most commonly evaluated by the visual observation of images of test objects and/or phantoms. Because of the uncertainties in such results (either large variance or bias or both), more precise quantitative measures based on statistical decision theory should be investigated. A series of simulations and experiments were conducted to investigate the statistical properties, i.e., the bias and variance, of the estimate of the square of the SNR of the ideal observer (SNRPWMF). Several methods of bias reduction were compared including one due to Fukunaga and Hayes. Good agreement was obtained between the results of simulations and the theoretical predictions for the bias and variance. The different methods of bias reduction have the same applicability for both ideal and quasi-ideal observers for the series of SKE/BKE tasks investigated in the present study. This work also provides some new avenues for additional investigation. First, the techniques can lead to protocols for making the evaluation of imaging system performance with a limited number of sample images, which is an important issue for any clinical implementation. Second, since selective spatial frequency channels can be used in estimating the SNRPWMF, the method has potential utility for imaging tasks beyond SKE/BKE tasks such as those with clinically relevant backgrounds but possessing stationary statistics.

Combining a few diagnostic tests or features

There are many current trends toward combining diagnostic tests and features in medical imaging. For this reason we have been exploring the stucture of the finite-training-sample bias and variance that one encounters in pilot or feasibility studies within this paradigm. Here we report on the case of the simple linear Bayesian classifier in a space of a few dimensions (two through fifteen). The results argue for the importance of estimating these effects in clinical studies, perhaps through the use of resampling techniques.

Supervised Pattern Recognition Techniques In Quantitative Diagnostic Ultrasound

The methods of statistical pattern recognition are well suited to the problems of in vivo ultrasonic tissue characterization. This paper describes supervised pattern recognition methods for selecting features for tissue classification, calculating decision boundaries within the selected feature space, and evaluating the performance. We address the considerations of dimensionality and feature size which are important in classification problems where the underlying probability distributions are not completely known. Examples are given for the detection of diffuse liver disease in the clinical environment.

Effects of finite sample size and correlated/noisy input features on neural network pattern classification

In many areas of practical interest, for example medical decision making problems, input data for training and testing neural networks are severely limited in number, are corrupted by noise, and may be highly correlated. In this study we examine these factors by investigating network performance on a simulated Gaussian data set with known first and second order statistics. Following the work of Wagner et al. for statistical (likelihood- ratio) classifiers, we study how the addition of noisy/correlated features affects the performance of neural network classifiers. Results are similar to that of the previous study, demonstrating that for small data sets, additional noisy/correlated features in fact degrade network performance. In addition, the use of sophisticated statistical techniques including the jackknife, Fukunaga-Hayes group jackknife, and bootstrap to estimate performance variation and remove small-sample bias are examined and found to offer significant advantages.

Application of parametric spectral estimation to medical ultrasound and magnetic resonance spectroscopy

This paper addresses variances of estimates of power spectral densities of radio-frequency (RF) signals generated with ultrasound and magnetic resonance spectroscopy. The spectral estimation methods studied involved autoregressive (AR) and moving average (MA) models. With experimental ultrasonic data, the power spectral density estimate obtained using the MA model exhibited an appreciable reduction in vanance compared with the squared modulus of the FFT. With magnetic resonance spectroscopic data, the AR spectral estimate was comparable to, but not significantly better than, the squared modulus of the FFT.