Arriving at the Bureau of Radiological Health in 1972, Bob Wagner was thrust into the Bureau's quandary over how to
quantify the imaging benefit associated with the radiation dose cost of medical imaging procedures. In short order he
had set up the framework for FDA imaging research for the next 36 years. Bob played a key role in these early years in
assisting in the founding of the SPIE Medical Imaging series of meetings, in measuring and organizing round robin
comparisons of imaging measurements of the fundamental physical quantities required for performance evaluation, and
in developing the framework for how these measurements could be combined to provide meaningful assessment figures
of merit. He worked assiduously to counter both those who claimed that radiology was an art not a science and those
who made extravagant claims for the dose reduction/image quality benefits of their particular variety of image
capture/image processing system. In the process he became one of the founding fathers and key participants in the
medical image performance assessment community as represented today at SPIE Medical Imaging 2009.
Complex-valued weights are used in the first layer of a feed forward neural network to produce a `transform' neural network. This network was applied to a phase-uncertain sine wave detection task against a Gaussian white noise background. When compared with results of a human observer study on this task by Burgess et al., performance of the transform network was found to be nearly equal to that of an ideal observer and far superior to that of the human observers. Performance was found to be dramatically affected by initial values of the weights, which is explained in terms of concepts from statistical decision theory.
Neural networks are applied to a Rayleigh discrimination task for simulated limited-view computed tomography with maximum-entropy reconstruction. Network performance is compared to that obtained using the best machine approximation to the ideal observer found in an earlier investigation. Results obtained on 2D subimage inputs are compared with those for 1D inputs and presented previously at this conference. Back-propagation neural networks significantly outperform the `best' standard nonadaptive linear machine observer and also the intuitively appealing `matched filter' obtained by averaging over the images in a large training data set. In addition, the back-propagation neural network operating on 2D subimages performs significantly better than that limited to 1D inputs. Finally, improved performance on this Rayleigh task is found for nonlinear (over linear, that is, simple perceptron) neural network decision strategies.
Neural networks are applied to the Rayleigh discrimination task. Network performance is compared to results obtained previously using human viewers, and to the best machine approximation to the ideal observer found in an earlier investigation. We find that simple preprocessing of the input image, in this case by projection, greatly improves network convergence and only results obtained on projections are presented here. It is shown that back propagation neural networks significantly outperform a standard nonadaptive linear machine also operating on the projections. In addition, this back propagation neural network performs competitively to a nonadaptive machine that uses the complete two-dimensional information, even though some relevant information is destroyed in the projection process. Finally, improved performance on this Rayleigh task is found for nonlinear (over linear) neural network decision strategies.
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.
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.
Neural networks were applied to the task of detecting simulated low contrast lesions in limited-view reconstruction tomography images. Results were compared with those for theoretically derived machine observers and for human observers. Preliminary results indicated improved neural network performance for the small data set on which human observer data had been obtained, but further results for a larger data set give performance generally inferior to the best machine observer.
Neural network based analysis of ultrasound image data was carried out on liver scans of normal subjects and those diagnosed with diffuse liver disease. In a previous study, ultrasound images from a group of normal volunteers, Gaucher's disease patients, and hepatitis patients were obtained by Garra et al., who used classical statistical methods to distinguish from among these three classes. In the present work, neural network classifiers were employed with the same image features found useful in the previous study for this task. Both standard backpropagation neural networks and a recently developed biologically-inspired network called Dystal were used. Classification performance as measured by the area under a receiver operating characteristic curve was generally excellent for the back propagation networks and was roughly comparable to that of classical statistical discriminators tested on the same data set and documented in the earlier study. Performance of the Dystal network was significantly inferior; however, this may be due to the choice of network parameter. Potential methods for enhancing network performance was identified.
A back propagation neural network was used to compress simulated nuclear medicine liver images with and without simulated lesions. The network operated on the Gabor representation of the image, in order to take advantage of the apparent similarity between that representation and the natural image processing of the human visual system. The quality of the compression scheme was assessed objectively by comparing the original images to the compressed/reconstructed images through calculation of an index shown to track with human observers for this class of image, the Hotelling trace. Task performance was measured pre- and post-compression for the task of classifying normal versus abnormal livers. Compression of even 2:1 was found to result in significant performance degradation in comparison with other means of compression, but produced a visually pleasing image.
Compression methods based on Gabor functions are implemented for simulated nuclear medicine liver images with and without simulated lesions. The quality of the compression schemes are assessed objectively by comparing the original images to the compressed images through calculation of the Hotelling trace, an index shown to track with human observers for images from this modality.1 For compression based on thresholding the complex Gabor coefficients, better than 2: 1 compression is obtained without significant degradation in image quality.
The problem of image assessment is examined for several cases of parameter uncertainty. Several ideal and
sub-ideal observers are considered and figures of merit (FOM) for describing their performance are
considered. Advantages and disadvantages of these FOMs are enumerated. The spectrum of noise
equivalent quanta, NEQ(f), appears to be the most useful for evaluating a broad class of practical problems
since the performance of the best-linear as well as the ideal non-linear observers considered here is
monotonic with its components. However, more work is required within this context to quantify the effects
of the "null space," or regions in object space for which NEQ = 0. These regions derive from incomplete
measurement sets and may lead to severe image degrading artifacts that are not adequately covered by any
FOMs considered here.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.