In this study, we developed an optimal filter which maximizes the perceived SNR based on a statistical decision theory model for a human observer. This filter serves as a prewhitening filter that compensates for both the image noise Wiener spectrum and the observer's visual system response, thus allowing the observer to perform matched-filtering in a white noise background during signal detection. However, we found that the filtered image has to be displayed with a strong contrast enhancement factor in order to reduce the effects of observer's internal noise and the display system noise. The use of a large windowing factor results in an image exceeding the dynamic range of a display system. Due to this limitation, it appears to be difficult to implement the optimal statistical filter (OSF) effectively in a practical digital radiographic imaging system. Therefore, we examined alternative filters by using series approximation of the OSF. The perceived SNR's of the filtered images predicted by the statistical decision theory model indicate that these filters in combination with a moderate windowing factor can improve the detectability of signals over that achieved by the windowing technique alone. We discuss the theoretical basis for the development of these new filters and the results of our calculations. Examples of simple test object images processed by the filters are shown. The potential usefulness and limitations of the various image processing methods in practical settings are discussed.

An ideal observer is a mythical detector optimized to give the best performance for a given signal detection or discrimination task. "Best" is defined here to mean that the detector is designed to yield the lowest possible Bayesian risk. The signal-to-noise ratio can be measured at the output of an ideal detector, which we will call SNRideal, and used as a figure of merit for a medical imaging system. For SNRideal to be a meaningful metric for medical images it must predict the ability of a human observer to perform the same detection or discrimination task. Images have been generated with equal ideal-observer SNR's but different autocorrelation functions to test the applicability of ideal-observer SNR for predicting human observer performance for images of different noise textures.

A study has been made of threshold detection of disks in Computerized Tomography (CT) images. Low contrast objects with and without sharp edges (polystyrene cylinders and cones) have been constructed and placed in a phantom. The phantom is filled with various water/alcohol mixtures to vary contrast, and imaged with a modern CT Scanner. The scan dose, slice thickness, reconstruction algorithm and display parameters have been varied to demonstrate their effect, on delectability and to understand how low contrast detestability might be optimized. Preliminary results of detection experiments are presented.

The effects of the total inherent digital imaging system shift variance on the perception of image details were investigated. These effects are particularly important in clinical digital imaging systems when the sampling aperture is comparable to or greater than the resolution dimensions (width of the point spread function) of the input images. The total system shift variance of a commercial digital imaging system, System One, DigiRad, Inc. (Palo Alto, CA), was quantified using a multiple high contrast bar pattern image. The square wave response functions (SWRF) of several nominally randomly selected locations on the displayed images were obtained and linearized by the logit transformation. The Coltman equation was then used to calculate the corresponding modulation transfer functions. Contrast detail (CD) diagrams were obtained using images of Burger-Rose phantoms for different relative positions of the input images with respect to the sampling pixel matrix. The effects of the shift variance on the CD diagrams were found to depend not only on the relative location of the image with respect to the pixel matrix but also on the input image quality. Further studies need to be performed to assess the clinical importance of the inherent shift variance on clinical digital imaging systems.

A study was performed to characterize the response of five different film screen combina-tions for potential application in cephalometric radiographic studies. These film screen combinations included Kodak T-Mat G, T-Mat L and XL-films in combination with Lanex Regular and Lanex Medium screens plus XL-film with a Par Speed screen. The response of the film screen combinations evaluated were characterized by the use of contrast detail response functions. These functions were determined with a minimum of six observers, trained and untrained, for each of the image receptor combinations evaluated. The images for evaluation in this study were generated using a low contrast "Rose Burger" phantom with optical densities from 0.5 to more than 1.0 for all of the film screen combinations evaluated. Also included are the relative dose-density data for the film screen combination images evaluated.

This experiment studied the detectability of lesions superimposed at specified locations on CT images of water phantoms. Each set of images contained 120 pedestals (a constant CT number added to circular areas of the water-phantom images), 60 of which also had a circular lesion profile of a given size and contrast superimposed on their centers. Lesion detectability (estimated from observers' ROC curves) was measured for 24 separate sets of images, representing four different levels of contrast for each of six lesion sizes (3.0 to 16.0 mm in diameter). Direct calculations of each lesion's "signal-to-noise ratio" (SNR), from the physical CT values :in each set of images, predicted its detectability for observers, and there was no systematic change in the linear relation between detectability and lesion SNR as a function of lesion size. The levels of contrast required to produce either a given level of lesion SNR or a given level of detectability were approximately linear functions of the lesion's diameter on a log-log plot, and parallel to the contrast-detail curve that we measured independently with the same image format. The slope of these curves (about -1.0) is consistent with those measured by other investigators, and is inconsistent with the signal detection predictions that assume an ideal ramp function for the CT noise power spectrum. These results suggest that the ideal noise power spectrum may not be realized in actual CT scanners, and that the actual deviations at low spatial frequencies affect the detectability of large lesions.

The sensitivity of the human visual system to light intensity is time-frequency dependent, hence the perception of temporal changes in intensity depends on their frequency content. This can have great influence on the perceived signal to noise ratio of a dynamic image display. Our purpose was to develop a filter which would reduce noise frequencies to which we are more sensitive without affecting the useful information in the series. We developed a linear filter which creates a new series of images by adding three subsequent images multiplied by factors a , and a respectively. The optimal ratio 0/a for the reduction of the disturbing frequencies of noise depends on the speed of projection of the series. Because the optimisation is matched with the time properties of the human visual system, the effect of the filter can only be fully appreciated when the filtered series is presented on a movie display. The resulting effect is psychophysical and differs from observer to observer because of differences in individual visual systems.

In most previous studies of observer performance it has been assumed that the observer had complete a priori information about signal parameters and tasks have been performed using spatially constant backgrounds. Results will be presented for more complex tasks including identification of signals from an orthogonal set, signal edge location accuracy, detectability of negative contrast discs and detection of discs on sinusoidal and square wave luminance backgrounds. The effect of digital amplitude quantization will be shown to be predictable by a simple theory. It will also be shown that the standard assumption of a constant observer internal noise is incorrect. It has been found that internal noise is proportional to display noise when the display noise is easily visible. It is proposed that human performance is best described by a sub-optimal Bayesian decision strategy based on maximum a posteriori probabilities.

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.

A figure of merit for characterizing the performance of medical imaging systems is suggested. It depends on statistical properties of the object and noise and on the transfer matrix specifying the imaging system.

From the strict mathematical viewpoint, it is impossible to fully achieve the goal of digital image processing, which is to determine an unknown function of two dimensions from a finite number of discrete measurements linearly related to it. However, the necessity to display image data in a form that is visually useful to an observer supersedes such mathematically correct admonitions. Engineering defines the technological limits of what kind of image processing can be done and how the resulting image can be displayed. The appeal and usefulness of the final image to the human eye pertains to aesthetics. Effective image processing necessitates unification of mathematical theory, practical implementation, and artistic display.

By relating an arbitrary x-ray projection to several projections of the same object produced from a small array of source positions bearing a known circular geometric relationship to each other, it is possible to synthesize approximately an arbitrary projection not contained in the known data set. This investigation explores the underlying theory and applies it to radiographic images of diagnostic interest in dentistry.

A filtered backprojection reconstruction method suggested by Knutsson et al (1,2) has been applied to circular tomographic DSA (digital subtraction angiography). We have applied this reconstruction technique, termed "ectomography," to phantom data as well as preliminary in vivo studies employing a dog model. The properties of this technique are illustrated using limited angle (20 degree) circular tomography. Both eight and 16 projections were studied with and without subtraction. The infocus depth of field is markedly increased with this technique. This increased depth of field appears to be extremely useful for displaying dilute objects, such as opacified vasculature. However, both noise and reconstruction artefact are increased. The present status of this work and the future directions for reducing these artefacts are discussed.

Two "point-by-point" iteration procedures, namely, Iterative Least Square Technique (ILST) and Simultaneous Iterative Reconstructive Technique (SIRT) were applied to classical circular tomographic reconstruction. The technique of tomosynthetic DSA was used in forming the tomographic images. Reconstructions of a dog's renal and neck anatomy are presented.

Tomographic DSA (digital subtraction angiography) can be used to improve the image quality that results from intravenous angiographic studies of relatively stationary arterial anatomy. While DSA removes much of the non-opacified anatomy, tomographic blurring reduces both the severity of patient motion artefacts and the confusion introduced by overlapping vascular anatomy. For this purpose a conventional longitudinal tomography device to which an image intensifier and television has been added can be used. However, such an apparatus is inadequate for cardiac imaging due to the slow speed of the tomographic motion. A tomographic system consisting of a rotating focal spot x-ray tube and an image intensifier, modified to allow electronic image scanning, is proposed. After this device is constructed it will be possible to acquire tomographic images of the beating heart in as little as .005-.010 seconds. When combined with image subtraction it is anticipated that the quality of intravenous coronary angiograms will be improved in much the same way that tomographic DSA improves image quality in many of the other arteries of the body.

The Monte Carlo method with three input parameters, photon energy, scatter material, and material thickness, is used to calculate the scatter fraction and the point spread function of the scatter (PSF). Four scatter materials, C, Ca, H2O, and Al are tested using 50, 70, 90, and 110 keV monoenergetic sources. From this simulation, two conclusions can be made: (1) the PSF of scatter can be approximated as an exponential function, and (2) the scatter fraction shows energy dependence when expressed as a function of the product of linear attenuation coefficient and material thickness.

An algorithm for reducing the scatter component in digitally acquired radiographic images is introduced. Using two images acquired in different geometries, the scatter component is calculated on a pixel by pixel basis and used to correct an original image. This process is expedited by using an array processor for computation. The theoretical basis of the algorithm is developed, and experimental results on simple Lucite phantom images are reported.

A general analysis is made of the influence of stochastic amplifying and scattering mechanisms on the transfer of signal modulation and photon noise in imaging processes. In this way we quantify the spatial-frequency dependence of signal and noise as they propagate through a multistage imaging system. Whereas by definition the signal structure (or modulation) is transferred via the MTF, the input photon noise is effectively unmodulated signal and as such bypasses the MTF. However, stochastic amplification of photon noise by one stage of an imaging process may produce noise structure that constitutes an effective signal spectrum to the next. Thus, in general, it is necessary to cascade these two components of the noise spectrum separately at each stage.

Though intensifying screens reduce patient exposure, they also spread a point source of x-rays into a radial distribution of emitted light. The point spread function of the screen can be approximately decomposed into a portion due to scattering of the emitted light and a portion due to reabsorption of secondary x-rays. A random walk model is introduced to explain the dependence of light spread on scattering length and on the depth in the screen where the light is emitted. Monte Carlo simulations are used to find both the spatial spread of secondary x-rays and the fraction of photons that results from reabsorption of secondaries. These secondary x-ray parameters depend on (1} incident x-ray spectrum, (2) phosphor composition, (3) phosphor loading, and (4) screen thickness. These dependences are shown (for 60 - 120 kVp x-rays) for conventional phosphors, for Se, and for CsI. Finally, the magnitudes of secondary x-ray spread and of light spread are compared.

An experimental technique has been developed for characterizing the complex light absorption and transmission (radiative transfer) properties of any phosphor-binder composite in terms of the mean free photon scattering length (Xs), the mean free path before photon capture (A ), and the average cosine of the scattering angle <cos * >. These properties are determinedcfor arbitrary phosphor crystal size, phosphor and binder refractive indices, phosphor-binder volume ratio, phosphor emission wavelength, and binder absorption spectrum. The technique requires the measurement of total reflectance and transmission, using an integrating sphere, of a self-supporting screen series differing only in thickness. The experimental data is then used to find As,A , and <cos IP > by means of a nonlinear least squares fit to the 'enhanced' "6-flux" radiative transfer model developed by Richards and Mudgett. The radiation transfer data can then be used in Monte Carlo methods or Swank's neutron diffusion theory models to help predict speed, MTF and noise properties of practical X-ray screens constructed from the same phosphor-binder composite. We have found that, for a given binder and binder-phosphor volume ratio, scattering (A a) is primarily determined by mechanical factors such as phosphor crystal size, crystal habit, and packing density. Absorption (A ) is primarily determined by the phosphor's emission wavelength relative to the binder's spectral absorption. We also confirm Swank's results that light scatter within the screen reduces image spread and increases image sharpness.

The detective quantum efficiency (DQE) of radiographic screen-film systems is limited at low exposures by the threshold response of silver halide grains. Model calculations predict improvements in DQE at low x-ray exposures with a uniform (light) bias exposure to the film. The DQE gains predicted are compared with those established in the literature for the case where the signal exposure is also a direct light exposure. An optimum bias exposure has been calculated that balances the increase in grain sensitivity against the increase in fog to maximize DQE for the limit of very small exposures.

Models have been developed for the DQE of a screen-film system that, although based on a number of simplifying assumptions, include most of the significant parameters. At the same time an increasing number of experimental measurements are available for the complete DQE characteristics as a function of exposure and spatial frequency for practical screen-film systems. A systematic set of DQE measurements is presented here, which we examine from the viewpoint of current models. This reveals both areas of agreement and aspects that require further theory and experimentation.

We apply an existing model for the relationship governing the DQE characteristics of a screen-film system to the calculation of absolute values of the noise-equivalent number of quanta (NEQ). The roles of the screen and film are investigated separately, and the influence of interactive properties such as screen amplification and grain size is explored. Hence factors limiting the NEQ of existing screen-film systems are identified.

The information content of radiographs depends on the signal-to-noise ratio (SNR) of all the image areas. By means of an experimental device, developed by Oldelft, SNR can be directly measured. Data obtained are analyzed and presented for several radiographic film-screen combinations.

After some theoretical remarks on multienergy radiology, a simulation program able to perform the following tasks is described: image production and acquisition, image processing and test pattern simulation. Using a test pattern based on geometric structures with schematic shape of a thorax, dual and triple energy are investigated. Pictures showing material look-alike and removal, in presence of various noise sources (photon noise, detector electronic noise, quantization noise) are presented. Simulation program provides useful indications which will be employed in the experimental stage we are starting in our laboratory. Finally some considerations on multienergy digital radiology are discussed.

A technique for the measurement of optical trasfer function of digital radiographic systems, making use of a commercially available test pattern is presented and discussed.

Multichannel lithium-drifted silicon (Si(Li)) semiconductor detectors are well suited to line-scan medical imaging systems. They have excellent linearity, high efficiency, large dynamic range and good stability. A major limitation in many medical imaging systems using image intensifiers is the veiling-glare associated with the large differences in intensity between different parts of the image field. Images acquired with linear Si(Li) detectors should suffer less from veiling-glare, primarily because they are line-scan images, but also because the Si(Li) detector responds to X-rays directly rather than to the light produced in a scintillator. The effective veiling-glare has been measured in a linear, 64-channel Si(Li) detector of thickness 5 mm and center-to-center spacing between adjacent sensitive areas of 0.5 mm. Each elemental contact was 6 mm high and 0.4 mm wide. The measurement was made in two different ways, both employing a 33 keV beam of synchrotron X-rays at the Stanford Synchrotron Radiation Laboratory. In one method, a lead phantom in the shape of an isosceles triangle was illuminated across its width with a fan beam of 33 keV X-rays 0.5 mm high by 20 mm wide. The X-ray flux passing by the phantom was measured in each channel of the Si(Li) detector. In the second method, a highly collimated beam of 33 keV X-rays (0.5 mm high by 0.025 mm wide) was scanned across 15 channels of the detector in 0.01 mm steps. The measured intensity in each channel was recorded at each position of the beam and the photon flux fell to less than 0.1% at a distance of 3 mm from the edge of the beam. The results of both measurements were in mutual agreement. Both were compared also to calculated results based on the known total X-ray cross sections for Compton scattering and photoelectric absorption. To within the experimental error, the observed and calculated results were found to be generally in good agreement.

We present an X ray measurement methodology with some preliminary X ray sensitivity characteristics for a prototype digital radiography system using amorphous selenium as the primary image receptor. As an imaging modality this experimental electrostatic system has the potential to replace film in existing general diagnostic radiography procedures. The imaging plate consists of a 360 micron layer of amorphous selenium deposited on an aluminumoxide substrate. An initial plate charge of 1400 volts (3.9 volts per micron) was exposed to X ray spectra produced with 50, 70, and 90 kVp with total filtration of 3 mm aluminum and 9 cm lucite. After this exposure the plate was scanned by a bank of electrometer probes at a distance of 100 microns. Sensitometric comparisons were then made to a conventional calcium-tungstate film-screen combination. The sensiometric response of the system is shown to be linear with an almost four fold increase in exposure latitude.

Conventional ultrasound echography (B-scan) displays the detected envelope of the echoes that originate in a tissue probed by an ultrasound beam. Since the main echoes originate at organ boundaries, these images mainly display contours. During the envelope detection process most of the information that is related to phase (and then to frequency) has been lost. Now this information can be related to physical parameters of the tissue, such as the atte-nuation the ultrasound beam experiences in it. And these parameters have proved to be good pathology discrimi-nants for some diffuse disease that are very difficulty diagnosed using other methods. Tissue characterization aims at restoring this lost information. Various methods, based on sophisticated signal processing techniques such as time-frequency distributions, are available to estimate attenuation. Methods using either the so-called spectral centroid or the intensity of the signal in narrow frequency bands are presented. The hypotheses they require and their capability to estimate true attenuation are discussed. A method to correct these estimations for the depth-dependent effect of acoustic diffraction is developed. This method is based on the derivation of the ultrasonic field in the case of unfocused transducers and on calibration measurements. After the agreement between the theoretical derivation and experimental calibration has been shown the method is extended to focused transducers and its effectiveness is demonstrated. As a by-product of these computations, new imaging modalities are presented. They involve as well spectral-centroid images as narrow band images. The first display an estimate of the instant frequency of the ultrasonic signal. They seem to provide more information about the inner structure of biological tissue than conventional images do. The latter enhance the frequency diversity of ultrasonic signals and provide a powerful tool for the reduction of the speckle-like noise that conventional reflectivity images evidence.

Six real time units were compared in this study. Three units were mechanical sector scanners operating at 5. MHz or 7.5 MHz. Two phased arrays (one 3.5 MHz unit with dynamic focussing, and the other a fixed focus 5. MHz array) and a 5. MHz annular array were also tested. Two test objects were imaged: an anthropomorphic breast phantom and a gray scale test object with renal calculi, Thirty-seven images were evaluated by three radiologists specializing in ultrasound. They concluded that; 1) Even the most expensive 3.5 MHz dynamically focussed phased array could not compete with a 7.5 MHz or 5 MHz mechanical sector for imaging small structures in the 1. to 5. cm near the transducer; 2) Mechanical sector real time units whose focal zone is adjusted with a water bag do not produce images equivalent to the images produced by a phased array with dynamic focussing; 3) The cost of the unit is no guide to the quality of the image; 4) Sector images have more digital scan convertor mapping errors than rectangular images. This applies to both phased arrays and mechanical sectors.

The data acquisition rate in medical ultrasonic imaging devices is limited by the acoustic propagation velocity in the tissues. Typically in such machines the image lines are produced sequentially one line per transmitted pulse. A parallel processing scheme has been implemented which enables the data acquisition rate to increase by a factor of four through the simultaneous acquisition of four B-mode image lines from each individual broadened transmit pulse. The higher data rate can be used to increase the image frame rate, to produce independent images that can be averaged in the image frame to reduce noise, or to produce a conventional image at standard video frame rates while reducing patient exposure. Alternatively, the field of view can be increased over that of a normal scan without sacrificing frame rate. These advantages are achieved with little reduction in the measured resolution. The design and performance of this device are described. A sample in vivo image is included.

Magnetic resonance images from 0.35T and 0.5T imaging systems are processed and evaluated in order to visualize and quantify blood flow. The velocity of moving substances is either measured by local and temporal tracking or by motion dependent encoding and decoding. With the first method the passage time through an imaged slice of known thickness may be measured. The signal intensity at an offset position also indicates how far a bolus has advanced under consideration of its velocity distribution. The second principle depends on MR gradient field encoding.

With a minimally complete set of acquired patient data, computational techniques can be used to extract intrinsic quantities and reconstruct magnetic resonance (MR) images. The effects of various sequences can be modeled for the purpose of understanding the functional behavior of the NMR response to a variety ot changing parameters, as well as for the generation of new images without the need of actual acquisition. Maps are constructed which plot the signal difference between two tissues as a function of NMR parameters. These maps can be used as a guide to acquisition parameters for best contrast. Noise in calculated images resulting from the propagation or noise in the calculational process can be accounted for by displaying signal difference-to-noise as a function of NMR parameters. The effects of magnetic field strength changes on relaxation times are modeled using empirical data. These models are used to predict the effects on tissue contrast using difference maps and calculated images that have been extrapolated to new values of field strength.

In preceding works, a theoretical model has been introduced for the transmission function of a biological cell, in which the physical effects of the three transmission functions for the cytoplasm, the nucleus and the cellular membrane have been considered and studied1. Since biological cells are generally objects having small phase variations, we introduce this specific characteristic in the proposed theoretical model, in the present work, Its Fourier transform is numerically analyzed.

The ophthalmic image processor (IS 2000) was developed by the authors to provide a eans of acquiring images and applying technology to aid in the diagnosis of various disorders of the eye. Nelson'- has previously reported on the system design considerations, the hardware configuration and the system operation. The IS 2000 system has undergone nine months of clinical trials, acquiring images of the cornea, lens, retina and the optic nerve. Various analysis routines have been developed and tested to aid in the diagnosis of diseases of the eye. Clinical trials were begun in May of 1984. During that time a number of small modifications have been implemented in the system and analysis routines developed or modified in the intervening time.

An algorithm to determine the relative rotation of two digitized images is presented. The algorithm does not need any pre-processing of the picture elements and allows an effective recursive implementation. A theoretical stability analysis with respect to random disturbances, along with experimental results, is also reported.

DELTAvision is a high performance architecture consisting of standardized building blocks of hardware, software, and communications products combination to solve the PACS problem. A single architecture solves the small and large problem of acquiring, storage (archiving) distribution, database management, and viewing of medical data and images.

Diagnostic imaging has expanded rapidly in recent years as an effective tool in medicine. This growth has been supported by technological advances in sophisticated image acquisition and processing techniques. In particular, image processing may improve diagnostic performance by enhancing the information content or the ease of information extraction from a given image acquisition configuration, or may allow for lower required illuminating dosages and/or the use of smaller amounts of contrast media to obtain a given level of information. One class of enhancement operators transforms image data in a manner dependent on the statistical composition of local data neighborhoods. These operators generally have the effect of manipulating the effective image contrast in a local-area adaptive manner to enhance certain features while suppressing unwanted large scale shading. Such techniques have traditionally been performed off-line by computing or image processing hardware/software on single image frames, requiring many millions of arithmetic operations, with speeds heavily dependent on the size of local areas used and the degree of complexity of the algorithm applied. We describe here a system which enables complex image processing operations to be performed at video frame rates by a relatively inexpensive, very compact (e.g. 3 1/2 in. rack mount) piece of equipment which can be simply inserted in the video stream of an already existing image acquisition facility. This system utilizes a sampled data approach in a modular design which allows for such image enhancement techniques as the contrast enhancement schemes mentioned above, and also includes provision for storage (or temporal averaging) and subtraction of reference data masks. While this system can be applied to real-time adjustable enhancement of static images, it has particular application to fluoroscopy, angiography, endoscopy, etc., or in general to any modality which employs real-time (30 Hz) video display of image data.

An interactive image processing language (CLIPS) has been developed for use in an image processing environment. CLIPS uses a simple syntax with extensive on-line help to allow even the most naive user perform complex image processing tasks. In addition, CLIPS functions as an interpretive language complete with data structures and program control statements. CLIPS statements fall into one of three categories: command, control,and utility statements. Command statements are expressions comprised of intrinsic functions and/or arithmetic operators which act directly on image or user defined data. Some examples of CLIPS intrinsic functions are ROTATE, FILTER AND EXPONENT. Control statements allow a structured programming style through the use of statements such as DO WHILE and IF-THEN - ELSE. Utility statements such as DEFINE, READ, and WRITE, support I/O and user defined data structures. Since CLIPS uses a table driven parser, it is easily adapted to any environment. New commands may be added to CLIPS by writing the procedure in a high level language such as Pascal or FORTRAN and inserting the syntax for that command into the table. However, CLIPS was designed by incorporating most imaging operations into the language as intrinsic functions. CLIPS allows the user to generate new procedures easily with these powerful functions in an interactive or off line fashion using a text editor. The fact that CLIPS can be used to generate complex procedures quickly or perform basic image processing functions interactively makes it a valuable tool in any image processing environment.

In this paper we describe a system for the digitization, display, and storage of radiographic images. The system was set up to support a project to assess the clinical impact of matrix size on teleradiology, defined for our purposes as the electronic recording and transmission of radiographs. The system consists of two separate facilities, one for image digitizing and the other for image display, interpretation, and archiving. The image digitizing laboratory uses a solid state scanner interfaced to a dedicated, general-purpose minicomputer. The display facility consists of a general-purpose, multi-user minicomputer with large main memory and disk capacity and a 1280 by 1024 display subsystem with three separate image memories. The two facilities communicate via magnetic tape. This paper documents the characterization of the image data base used for the project, the design and operational features of the two facilities, technical problems encountered in the digitizing and display of radiographs, and some logistical and managerial considerations involved in generating and maintaining a large digital image archive. Since many of the functions of the system we have set up are common to picture archiving and communication systems, we have, where possible, compared the performance of the system we have developed with the requirements that would have to be satisfied by a system intended for clinical use. We leave for future reports the discussion of the conduct of the image reading experiments and the results of those experiments.

The construction of preformed cranial prostheses for large cranial bony defects is both error prone and time consuming. We discuss a method used for the creation of cranial prostheses from automatically extracted bone contours taken from Computerized Tomographic (CT) scans. Previous methods of prosthesis construction have relied on the making of a mold directly from the region of cranial defect. The use of image processing, bone contour extraction, and three-dimensional display allowed us to create a better fitting prosthesis while reducing patient surgery time. This procedure involves direct bone margin extraction from the digital CT images followed by head model construction from serial plots of the bone margin. Three-dimensional data display is used to verify the integrity of the skull data set prior to model construction. Once created, the model is used to fabricate a custom fitting prosthesis which is then surgically implanted. This procedure is being used with patients in the Maxillofacial Prosthetic Clinic at UCLA and this paper details the technique.

The coupling of mentation to regional cerebral blood flow (rCBF) has prompted the application of the Xe-133 inhalation method of measuring rCBF in the differential diagnosis of the two most common dementing diseases, Alzheimer's disease and multi-infarct dementia (MID). In this study receiver-operating-characteristic (ROC) curve analysis was used to assess the effectiveness of a 32 detector Xe-133 inhalation system in discriminating between patients with Alzheimer's disease and normal controls, MID patients and normal controls and between patients with Alzheimer's disease and MID. The populations were clinically evaluated as 1) normal (age 63.1 + 13.1, n=23), 2) Alzheimer's disease (age 72.7 + 7.0, n=82), 3) MID (age 76.4 + 7.6, n=27): The mean flow values for all detectors were lowest for the Alzheimer's disease group, larger for the MID group and largest for the normal controls. The dynamic relationship between the correct identifications (true posi-tives) versus incorrect identifications (false positives) per detector for any 2 pairs of clinical groups varies as the cutoff value of flow is changed over the range of experimental blood flow values. Therefore a quantitative characterization of the "decision" or ROC curve (TP vs FP) for each detector and for each pair of clinical groups provides a measure of the overall diagnostic efficacy of the detector. Detectors directed approximately toward the speech, auditory and association cortices were most effective in disciminatinq between each of the dementia groups and the controls. Frontal detectors were diagnostically inefficient. The Xe-133 inhalation system provided virtually no diagnostic power in discriminating between the two forms of dementia, however. Therefore this imaging technology is most useful when assessing the general diagnostic state of dementia (Alz-heimer's disease and MID) from normal cognitive function.

Two methods were studied of estimating automatically the relative volume of local lesions in digital subtractions radiographs. The first method approximates the projected, lesion area by an equivalent circular area, and the second by an equivalent polygonal area. Lesion volume is estimated in both methods as equivalent area times the average gray-level difference between the detected area and the surrounding background. Regression results of the estimated relative volume versus the calibrated size of lesions induced in dry human mandibles showed the polygonal approximation to be superior. This method also permitted successful monitoring of bone remodelling during the healing process of surgically induced lesions in dogs. The quantitative results, as well as the examples from in vivo lesions demonstrate feasibility and clinically relavance of the methodology.

We have been testing a digital beam attenuator (DBA) system for fabricating patient-specific compensating filters to improve image quality in chest radiography. At present, the technique is limited by a 2 hour attenuator fabrication time and a 20 mR exposure used to acquire an initial image from which the attenutor is designed. We now are developing an improved DBA system capable of generating, in a few seconds, patient-specific compensating filters for clinical chest radiography. The initial image will be acquired at less than 1 mR of skin exposure using an intensifying screen viewed by a low light-level television camera. Image processing including scatter and beam-hardening corrections generate a 614x614 matrix of values from which the attenuator is fabricated. The attenutor will be fabricated with individual layers having a 16x16 format using a special purpose dot-matrix printer. Successive layers are shifted by a quarter of a dot width to maintain the 614x614 sampling frequency in the final attenuator. After the attenuator is positioned automatically in the x-ray beam, a final image will be acquired directly on film.

An investigation is made of the possibility of obtaining reliable measurements of tissue density from digital chest radiographs. This involves assessment of the repeatability and uniformity of x-ray exposure using microdensitometry of conventional film and estimation of scattered radiation. The range of variation of film density spatially and over time was found to be within ±0.03 optical density units (O.D.). Scatter fraction measurements using partial absorbers gave results within, on average, 6% of the conventional measurement and was found to be approximately linearly related to mass/unit area. By correcting for film response and scatter, measurement of a variety of densities of foam rubber was found to correlate well with true density, with an average error of 5%. The estimation of regional lung ventilation from inspiration/expiration radiographs is investigated, using spatial warping techniques to facilitate the identification of corresponding areas of the two radiographs.

Three major errors implicit in the determination of ejection fractions from videodensi-tometric analysis of intravenous digital subtraction ventriculograms are errors due to: (1) nonlinear transfer of densitometric information about relative volumes, (2) inadequate background subtraction and (3) overlapping atrial contrast. Nonlinear errors associated with our imaging system were estimated using balloon phantoms in a water bath. These were small for a tube voltage of 90 kV and an iodine concentration of approximately 37 mg/cc commonly used in clinical studies. Background subtraction errors are approached by using phase matched second order mask images temporally close to the contrast images. Atrial overlap errors remain a significant problem. However, videodensitometric ejection frac-tions from intravenous digital studies correlated well with first pass right ventriculo-grams (r = 0.84) and left ventriculograms (r = 0.85).

Computerized Tomography (CT) images reflect the x-ray absorptive properties of the patient. CT images of the chest region, however, are difficult to display. Due to the large density differences between the tissue region of the heart and the air filled region of the lung, the lung appears nearly black if the heart is viewed at normal contrast. Similarly, if the image is adjusted to view the lungs at reasonable contrast, the heart region becomes saturated and featureless. A standard technique for viewing the two regions remaps the darker lung gray-level to values comparable to that of the heart region. The "dark-line artifact" results for those pels of intermediate value between the heart and lung. This apparent line surrounds the lungs and gives the entire image an objectionable piecemeal appearance. A software system, CTIP, written in FORTRAN is discussed which remaps CT image gray-level so that both the lung and heart regions are clearly visible with a more natural anatomical appearance. The system is adaptive to image statistics derived from the gray level histogram and is completely automated. The processed and the standard approach images were compared by twenty-two radiologists with 82% preferring the processed images. All of the respondents agreed that the processed images possessed a more anatomical appearance than the standard approach.

A contrast enhancement digital filtering technique has been explored for improving the visibility of vasculature in nonsubtraction digital angiograms. Experimental results on sample images showed that the average contrast between the vasculature and its surrounding tissue was increased by sixfolds and the signal-to-noise ratio was increased by twofolds. This suggests a great potential of this enhancement filtering technique for digital non-subtraction angiography, bypassing the problem of patient body motion which causes difficulty in image registration in digital subtraction angiography.

Digital Subtraction Angiography (DSA) is commonly used in conjunction with intravenous contrast injection for detection of atherosclerotic disease in arteries outside of the heart. Images of coronary arteries obtained with intravenous DSA have been limited in quality by several important factors. Among these is the confusing background provided by superposed pulmonary veins. Because these opacify just before the coronary arteries, conventional remasking results in substantial loss in coronary artery contrast. This paper presents preliminary work on a processing scheme in which the degree of correlation between the contrast pass curves in individual pixels and a reference region can be used to adaptively suppress pulmonary structures.

The aim of this research is to avoid as much as possible the development of the woman breast cancer by the mean of a preventive systematic analysis of mammographic textures and the automatic determination of a "risk coefficient" for each given texture. In a first step, a standard procedure for obtaining X-ray mammograms is set-up and a specialist classifies the resulting radiographic images in four groups of risk. In a second step, specific and selected textures algorithms using both global and local statistical properties of the images are implemented. A number of X-ray mammograms, given by the hospital, has been studied. One of the resulting important observation is that it seems unapropriate to define a set of distinct classes of risk and that an increasing gravity degree correlated to a continuous evolution of the mammographic textures from the lowest to the highest degree of risk is to be preferred. Finally, a systematic comparison between the human classification and the numerical coefficients provided by the texture analysis is performed : a critical examination of these preliminary results leads us to a constructive discussion concerning the future developments of the proposed method.

Images are taken from a 1023 line video television chain coupled to an x-ray image intensifier. The images were captured on either a 512 x 512 pixel digital field store or summed into a 1024 x 1024 pixel frame store. The resultant images could be copied by using a multiformat camera if desired. All images are obtained using conventional fluoroscopic x-ray exposure rates (25 uR/s to the input of the x-ray image intensifier). The image qualities are compared using contrastdetail diagrams and motion modulation transfer functions. It is found that a reasonable compromise between image quality and motion blurring is obtained for four television frame exposures. It is suggested from the results that pulse progressive readout of the camera would permit an even better compromise. Preliminary work on such a system will be discussed. We also discuss the applicability of intensified x-ray television systems for general digital radiography and the requirements necessary for physician acceptance.

In a cooperative effort with Dupont we have been studying digitized GI images utilizing stupid and smart filters to understand how they might help the radiologist in the reduction of perception error. Specifically, we have dealt with the detection of polyps and diverticulae both by the radiologist and the computer. In this paper we show why digital techniques are needed and describe our progress in implementing them.