A controlled experiment which simulated the standard reporting practices used by radiologists was performed on IRIS to (1) assess the quality of reporting using digital images and (2) evaluate the use of enhancement tools. Reports generated by 5 experienced radiologists using IRIS were compared to reports generated by the same radiologists using analog films on the same emergency department cases at least seven months earlier. In 100 cases, nine findings were missed using IRIS and 14 new findings were identified using IRIS. To make diagnoses, enhancements which caused the image to be darker and increased contrast were most frequently used. There was no systematic relationship between the enhancement tools used and different types of cases.

We have developed a practical digital imaging colposcope for use in research on early detection of cancerous and pre-cancerous tissue in the cervix. Several copies of the system have now been used in a variety of clinical and research environments. Two issues of considerable interest which emerged early in our work involved the roles of color and spatial resolution as they applied to digital imaging colposcopy. In each instance these qualities potentially have a significant impact on the diagnostic efficacy of the system. In order to evaluate the role of these parameters we devised and conducted a receiver operating characteristic (ROC) evaluation of the system. It is apparent from these tests that a spatial resolution of 512 x 480 pixel with 7 or 8 bits of contrast is adequate for the task. The more interesting result arises from the study of the use of color in these examinations; it appears that in general, contrary to the widely held perception of the physicians involved, color apparently provides the clinician with little or no diagnostic information. Indeed, in some instances, access to color seemed to confuse the physician and resulted in an elevated rate of false positives. Results of the ROC tests are presented in this paper along with their implications for further development of this imaging modality.

Computed radiography uses a photostimulable phosphor coated imaging plate which is exposed to x-rays and laser read to form an image. After laser reading, there is a considerable amount of energy remaining on the imaging plate which is not used. This study simulated a change in the laser readout process to utilize more of the energy on the image plate, and potentially improve image quality without changing exposure factors. Images of a contrast detail phantom were made before and after alteration of the readout process and analyzed by both physical and psychophysical means. It was found that there is an increase in the signal-to-noise ratio, when measured with an aperture of the size of a single pixel (linear dimension about 0.15 mm). However there is no change in the signal-to-noise ratio, when measured with apertures of the size of 0.75 mm (5 x 5 pixels) and 1.5 mm (10 x 10 pixels). This agrees with the results of the contrast detail study: after alteration, the observers did not detect smaller objects than they had before the alteration. It appears the imaging plate readout process is fairly optimized.

A major difficulty with currently available display systems is contrast rendition. The number of gray steps that can be displayed by a digital system driving an analog display is determined by the luminance range of the display and the signal-to-noise ratio (Schreiber, 1986). For a good display this might be 128 steps (7 bits). The minimal detectable contrast is determined by the average luminance at the eye and the image noise. Pizer and Chan (1979) using ROC techniques found the "perceived dynamic range" of a "high quality W monitor" to be about 87 steps. In their experiment, the eye was adapted to each luminance level. In a display like the chest with large bright and dark areas (heart and mediastinum vs. lungs) the eye will be adapted to some intermediate level and not optimally coupled to either the bright or the dark area. For this reason gray scale image processing is used for matching correctly the display to the observer' s perceptual system (Kundel, 1986). There have been some attempts to develop "perceptually relevant" gray scales (Johnston et al.,1985). Clinicians do not seem willing to spend the time needed to adjust optimally the controls at a display console. Even when the controls are fairly simple they still prefer images that technicians have recorded on film using some prescription for window and level. Perhaps they do not know how to optimize the contrast in the image and prefer to leave it to someone else or perhaps the controls are not as simple as they look. We have been working on a rule-based system that will select the proper gray scale transfer characteristic (GSTC) for the initial look at the image. It is based on the notion that there is an optimal GSTC for each diagnostic finding on the image and that for most images the diagnostic findings can be predicted from prompts derived directly from the clinical history. If this is the case, radiologists who receive diagnostic prompts should be able to predict the appropriate GSTC. This experiment was done to determine if radiologists can agree about the GSTC that will optimize the video display for particular diagnoses.

This paper raises methodological questions concerning Free-Response Receiver Operating Characteristic (FROC) analysis as used to judge the quality of images in radiology. This paper has three purposes: describing FROC analysis as used in the literature; discussing whether FROC methods achieve their objectives; and identifying several methodological shortcomings in FROC at the present state of the art that must be solved if FROC is to be fully useful.

As Picture Archiving and Communication System (PACS) technology has matured, video image capture has become a common way of capturing digital images from many modalities. While digital interfaces, such as those which use the ACR/NEMA standard, will become more common in the future, and are preferred because of the accuracy of image transfer, video image capture will be the dominant method in the short term, and may continue to be used for some time because of the low cost and high speed often associated with such devices. Currently, virtually all installed systems use methods of digitizing the video signal that is produced for display on the scanner viewing console itself. A series of digital test images have been developed for display on either a GE CT9800 or a GE Signa MRI scanner. These images have been captured with each of five commercially available image capture systems, and the resultant images digitally transferred on floppy disk to a PC1286 computer containing Optimast' image analysis software. Here the images can be displayed in a comparative manner for visual evaluation, in addition to being analyzed statistically. Each of the images have been designed to support certain tests, including noise, accuracy, linearity, gray scale range, stability, slew rate, and pixel alignment. These image capture systems vary widely in these characteristics, in addition to the presence or absence of other artifacts, such as shading and moire pattern. Other accessories such as video distribution amplifiers and noise filters can also add or modify artifacts seen in the captured images, often giving unusual results. Each image is described, together with the tests which were performed using them. One image contains alternating black and white lines, each one pixel wide, after equilibration strips ten pixels wide. While some systems have a slew rate fast enough to track this correctly, others blur it to an average shade of gray, and do not resolve the lines, or give horizontal or vertical streaking. While many of these results are significant from an engineering standpoint alone, there are clinical implications and some anatomy or pathology may not be visualized if an image capture system is used improperly.

The affects of quantization noise in magnetic resonance images (MRI) were studied, and simple modifications are shown to give significant improvement in subjective image quality and in some quantitative measurements. A liver phantom based on a MR liver image was created to simulate the effects of quantizing the MR signal. The MR signal which would be generated by this phantom was characterized and used to study quantization effects as well as to help in developing a more efficient quantization scheme. Uniform quantization of the signal was simulated to determine the effects of quantization noise on the liver phantom. Quantitative measurements using SNR and detectability were made and used as a basis of comparison for similar measurements utilizing other quantizers including uniform quantization with quantizer overload and logarithmic quantization. Quantitative measurements were again made and compared to the full range uniform quantizer. Simulations were performed without additive noise, and results are given in the form of graphs, tables, and actual images. Simulations were also performed with noisy data but are not reported in detail in this paper. Q uantitative measurements were consistent in most cases and agreed well with subjective evaluations. It was found that uniform quantization noise can significantly affect image quality. It was also determined, by using a logarithmic quantizer or by simple overloading of the uniform quantizer, that significant improvements in image quality can be achieved. The results are extensible to other image collection systems, particularly those with high dynamic range.

Digital film scanning/printing systems which do not alter the diagnostic content of radiographs are required for diagnostic teleradiography, electronic image archiving, and image enhancement. Described is a high fidelity system capable of digitizing and reprinting films with minimal alteration of detail or degradation of noise for either general radiographs or mammograms. The system transfers resolution up to 10 lp/mm and has a useful density range of 0-3.5. Examples of chest, bone and breast films are presented.

Performance figures are reported on absolute luminance, luminance uniformity, characteristic display function, internal scatter, dynamic range, distortion, modulation transfer and its spatial uniformity, and temporal and spatial noise of two commercial ORT display systems as well as films printed by laser image recorders and displayed on light-boxes. One of the ORT displays has a matrix of nominally 2000 x 2000 pixels, the other of 1000 x 1500 pixels. The laser image recorders cover a matrix of 3500 x 2200 pixels. When comparing equal pixel matrices of hard versus soft copy displays, typically hard copies facilitate greater information transfer than soft copies due to a presentation with higher absolute luminance, greater perceived dyiiainic range, and better spatial resolution. Perceived dynamic range and resolution are partially degraded in the ORT displays by internal scatter. Soft and hard copy displays are about equivalent in terms of luminance non-uniformity, noise, and geometrical distortion. All displays differ in their characteristic display functions and thus in perceived contrast resolution. A display function standard is proposed to the industry by which mismatches between hard and soft copy presentations can be nimized as well as means for maintaining standardized performance.

This paper describes the results of measurements which permit determination of the maximum information content of images displayed with a high resolution CRT of a nominal pixel matrix of 2000 x 2000 adressable pixels. These measurements are based on the precise determination of signal, noise and spatial resolution. Noise is measured both as temporal noise and as spatial noise, using both temporal and spatial sampling techniques. The dynamic range as given by the maximum and minimum values of luminance can be as high as 1900, however for practical settings of brightness and contrast it is about 424. The dynamic range is reduced from 424 to 1 38 as a result of veiling glare when the SMPTE pattern is displayed. The spatial noise was significantly larger than the temporal noise: For a command level of GL = 255, a single pixel and a single CRT frame, the temporal signal-to-noise Eatio was 59, while the spatial signal-to-noise ratio was only 7. The RMS noise (both spatial and temporal) was proportional to the squareroot of the respective bandwidths. Both temporal and spatial noise power spectra were independent of their respective frequencies (white noises). In the average, the size of a pixel was much larger than the size of an addressable pixel, indicating, that the total number of usable pixels was smaller than the number of addressable pixels: Instead of the addressable pixelmatrix of 2000 x 2000 pixels, we estimated only 532 x 548 "Noise-Equivalent" pixels. It was estimated, that for spatial integration over a single pixel and temporal integration over a single CRT frame, the maximum information content of images produced by this monitor, is 1.72 x 10⁶ bits, based on temporal noise; however due to spatial noise that number is reduced to 1.28 x 10⁶ bits. This contrasts sharply with the nominal information content of 3.2 x 10⁷ bits as determined from the nominal raster of 2000 x 2000 pixels and the 8 bits digitization at the display buffer.

Two methods of image display are developed in which the best 5-7 images of the peripheral arteries are joined together to form a single, continuous image of the legs. First, a complete image of both legs, called WHOLE LEG, is reduced so that it fits on a single monitor or hardcopy image. Second, a full resolution image, called SCROLL, is assembled that can be scrolled on a video monitor. From the geometrical parameters of the acquisition (source-to_detector distance, table height, etc.), the portions of each image to remove beforejoining the remaining portions together are estimated. This estimate is refined using an automatic search for the best match. To correct for body taper, images are intensity equalized. Using a reconstruction method that assumes a planar geometry results in a volume that is displayed in neighboring image frames and a volume that is never displayed. Nevertheless, WHOLE-LEG is aesthetically pleasing. In the case of SCROLL, images are displayed such that every part of each input image is displayed at one time or another. A video tape of SCROLL shows how this works.

The work described here is the result of an attempt to implement a unified representation for the display, storage, model matching, location and analysis of cardiac images. The Symmetrical Axis Tranaform (SAT) has attributes which make it useful for object representation, induding 3-D objects. It is hierarchical, compact, exact and fast. Using these characteristics, object recognition has been done using SAT [6], and means of obtaining surface descriptions for 3-D object8 has been reported [4]. The SAT has been described using morphological operations [5]. This paper describes an extension of the SAT to the representation of arbitrary non-Eudidean surfaces using geodesic operators put forth by Lantejoul and Beucher [2].

Transparent Volume Imaging began with the stereo xray in 1895 and ended for most investigators when radiation safety concerns eliminated the second view. Today, similiar images can be generated by the computer without safety hazards providing improved perception and new means of image quantification. A volumetric workstation is under development based on an operational prototype. The workstation consists of multiple symbolic and numeric processors, binocular stereo color display generator with large image memory and liquid crystal shutter, voice input and output, a 3D pointer that uses projection lenses so that structures in 3 space can be touched directly, 3D hard copy using vectograph and lenticular printing, and presentation facilities using stereo 35mm slide and stereo video tape projection. Volumetric software includes a volume window manager, Mayo Clinic's Analyze program and our Digital Stereo Microscope (DSM) algorithms. The DSM uses stereo xray-like projections, rapidly oscillating motion and focal depth cues such that detail can be studied in the spatial context of the entire set of data. Focal depth cues are generated with a lens and apeture algorithm that generates a plane of sharp focus, and multiple stereo pairs each with a different plane of sharp focus are generated and stored in the large memory for interactive selection using a physical or symbolic depth selector. More recent work is studying non-linear focussing. Psychophysical studies are underway to understand how people perce ive images on a volumetric display and how accurately 3 dimensional structures can be quantitated from these displays.

We have developed the algorithm and the computer program that constructs three dimensional image (3D) of soft tissue from multiple 2D slices of computed tomograph(CT) and magnetic resonance(MR) images. In construction of MR-3D, we have used multi-spin-echo images, and performed segmentation by comparing the curve-shape(intensity vs number of echo) of each pixel with a standard curve-shape that belongs to the desired organ. In the segmentation for CT-3D, we have used two kinds of threshold. In addition to CT value, we have used differential intensity of image as the threshold. The 3D micro-program has been installed in the Image Workstation of PACS. The position of observer's eye can be changed in less than a few seconds.

SOFTVU is a collection of computer programs capable of performing several image manipulation tasks required to visualize and analyze multidimensional image data obtained from medical imaging scanners. In this paper we present some novel image segmentation, interpolation, and display techniques developed and implemented in the package. Our rule-based segmentation can automatically extract different structures of the brain from MR image data. The shape-based interpolation scheme presented here minimizes user involvement in interactive segmentation and enhances portrayal of dynamic as well as static objects. The rendering subsystem can render multiple objects with color and cut-away sections. The system as a whole provides efficient software solutions to several clinical image analysis tasks.

We present two methods for acquiring and viewing integrated 3-D images of cerebral vasculature and cortical anatomy. The aim of each technique is to provide the neurosurgeon or radiologist with a 3-D image containing information which cannot ordinarily be obtained from a single imaging modality. The first approach employs recent developments in MR which is now capable of imaging flowing blood as well as static tissue. Here, true 3-D data are acquired and displayed using volume or surface rendering techniques. The second approach is based on the integration of x-ray projection angiograms and tomographic image data, allowing a composite image of anatomy and vasculature to be viewed in 3-D. This is accomplished by superimposing an angiographic stereo-pair onto volume rendered images of either CT or MR data created from matched viewing geometries. The two approaches are outlined and compared. Results are presented for each technique and potential clinical applications discussed.

Stereolithography, a new technique of prototype fabrication developed for the aerospace industry, offers a unique way to display patient anatomy. Like current computer-aided-design (CAD) systems, it uses digital image data from CT and MR to produce a physical model. Unlike conventional CAD it does not require a cutting tool and therefore CAD tool-path limitations do not exist. The stereolithography apparatus (SLA) uses an ultraviolet laser to selectively polymerize and solidify a polymeric liquid plastic solution under computer control. The device was used to produce a model of cranial bony anatomy from CT image data providing full internal detail in the constructed model including encased sinuses, foramen, and potentially complete internal anatomy within a closed skull. The advantages and disadvantages of this technology are reviewed with an emphasis on future development.

The aim of this application is to interactively transfer information between CT, MRI or DSA data and a 3D stereotactic atlas digitized on a C. Based on a 3D organization of data, this system is devoted to assist a neurosurgeon in surgical planning by numerically cross-assigning information between heterogeneous data (in-vivo or atlas). All these images can be retrieved in digital form from the PACS central archive (SIRENE PACS system). The basic feature of this confrontation is the Talairach's proportional squaring which consists in dividing the 3D cerebral space in independently deformable sub-parts. This 3D model is based on anatomical structures such as the AC-PC line and its two associated vertical lines VAC and VPC. Based on this proportional squaring, the atlas has been digitized in order to get atlas plates along the three orthogonal directions of this geometrical reference (axial, coronal, sagittal). The registration of in-vivo data to the proportional squaring is done by extracting either external framework landmarks or anatomical reference structures (i.e. AC and PC structures on the MRI sagittal mid-plane image). Geometrical transformations and scaling are then recorded for each modality or acquisition according to the proportional squaring. These transformations make for instance possible the transfer of a 3D point of a MRI examination to its 3D location within the proportional squaring and furthermore to its 3D location within another data set (in-vivo or atlas). From that stage, the application gives the choice to the neurosurgeon to select any confrontation between input data (in-vivo images or atlas) and output data (id).

With the integration of computer graphics to presently available computed tomography, magnetic resonance imaging, and positron emission tomography scan devices, processing and displaying three dimensional data arrays is rapidly becoming important. Present systems can model full color, complex two or three dimensional objects, and manipulate their perspective displayed appearance in real time. However, even with the many hardware and software tools available for providing enhanced versions of two or three dimensional information,the rapid developments within these areas continues to outpace the ability of users to learn and use them efficiently. At the University of Tennessee Space Institute, Department of Electrical Engineering, and Vanderbilt University Medical Center, Department of Radiological Sciences, research is being performed to develop a voice actuated computer graphics system that responds interactively with the human voice. The research uses a Texas Instruments voice recognition and speech synthesis board, a Texas Instruments 955 computer workstation, and a Sun Microsystems 4/280 color graphics workstation. Verbal commands are used to control the graphics system programs, set parameters for displaying two and three dimensional views, and instruct the system for additional cross-sections, viewing references, display formats, etc. Verbal responses from the computer inform the user when additional information is needed or error conditions have occurred. The system provides a means for users to conduct an interactive verbal dialogue with the computer for rendering and displaying complex medical imaging data.

A cost-effective personal workstation has been designed and developed using PC based commercial products, to assist physicians and scientists who are interested in creating teaching programs based on CT and MR images, collecting interesting cases, carrying out research projects requiring original digital image information, or generating reports. The development hardware is based on an IBM PS/2 Model 80 with a 85 14 graphics adaptor, an optical disk, a trackball, a digitizing pad, a 9 track tape drive, a mouse, a T800 transputer card, an Ethernet card, a voice recognition card, and a fax machine. The software development has been done using Microsoft C under Windows/286, such that software can be used with any 80286/80386 PCs capably of supporting Windows/286.

During a clinical trial, emergency physicians and radiologists at the Ottawa Civic Hospital used IRIS (Integrated Radiological Information System) to process patients' x-rays, requisitions, and reports, and to have consultations, for 319 active cases. This paper discusses IRIS user interface issues raised during the clinical trial. The IRIS workstation consists of three major system components: 1) an image screen for viewing and enhancing images; 2) a control screen for presenting patient information, selecting images, and executing commands; and 3) a hands-free telephone for reporting activities and consultations. The control screen and hands-free telephone user interface allow physicians to navigate through patient files, select images and access reports, enter new reports, and perform remote consultations. Physicians were observed using the system during the trial and responded to questions about the user interface on an extensive debriefing interview after the trial. Overall, radiologists and emergency physicians were satisfied with IRIS control screen functionality and user interface. In a number of areas radiologists and emergency physicians differed in their user interface needs. Some features were found to be acceptable to one group of physicians but required modification to meet the needs of the other physician group. The data from the interviews, along with the comments from radiologists and emergency physicians provided important information for the revision of some features, and for the evolution of new features.

In this document we describe the design of FilmPlane2, its user interface and implementation. The underlying mental model of the workstation design is that of a virtual electronic view box. Viewed "from a distance" the whole view box can be seen, but with only in rough detail. Viewed "from up close", individual images can be seen at full resolution. The system provides a natural way to determine which images show up in the "up close" view. Additional features allow for sequential and nonsequential access to images, measuring anatomical features, contrast adjustment, and limited zooming. FilmPlane2 is being implemented using X Windows and C++. The full system is designed to run on a Stellar graphics workstation with three 1280x1024 greyscale display screens, as well as Sun and DEC workstations. The system can be adapted to an arbitrary number of display screens.

A distributed CT/MR digital viewing station for the neuroradiology section has been developed and is being evaluated in our department. The major components of the station consist of a SUN host computer, a PIXAR II image processor, and four 1K x 1K progressive video monitors. The software modules operating in the station include an image acquisition process, a local database process, and an user image display and processing process. Functions provided by the station are described. Preliminary results obtained from clinical evaluation are reported. Future plans to refine the station are presented.

The Multi-modality Image Display Station (MIDS) is designed for the use of physicians outside of the radiology department. Connected to a local area network or a host computer, it provides speedy access to digitized radiology images and written diagnostics needed by attending and consulting physicians near the patient bedside. Emphasis has been placed on low cost, high performance and ease of use. The work is being done as a joint study with the University of Texas Southwestern Medical Center at Dallas, and as part of a joint development effort with the Mayo Clinic. MIDS is a prototype, and should not be assumed to be an IBM product.

Biomedical structures such as the beating heart are inherently multi-dimensional in nature. In addition to the three spatial directions which represent the object location and orientation, higher order dimensions can be assigned to represent various object parameters such as time and tissue density. In this paper, we propose a hierarchical data structure which can be mapped into a computer architecture that will efficiently store, manipulate, and display time varying images of multi-dimensional biomedical structures. This n-D object representation scheme which is called a linear hypertree is a generalization of the linear quadtree and octree from their respective 2-D and 3-D spaces to n-D environment. It is a hierarchical data structure which represents multi-dimensional volumetric information in a 2'-way branching tree. The basic properties of a linear hypertree are briefly presented along with the procedure for encoding the node rectangular coordinates into a hierarchical locational code. Two decoding techniques that transform the node locational code into its rectangular coordinate format are introduced. Some adjacency concepts in a multi-dimensional environment are defined. A neighbor finding algorithm which identifies the locational code of the adjacent hypertree node in a given direction is also presented. This algorithm does not convert the locational code to its rectangular coordinate form; instead, it operates directly on the node locational code in order to determine the neighbor's identification. Finally, Procedures for computing the locational codes of larger and smaller size neighbors are also included.

A new method for spatial-frequency analysis, the Frazier-Javierth transform (FJT), has recently been introduced which allows the simultaneous encoding of image properties in both the space and frequency domains. This technique appears to have applicability to the problem of image data compression, in a way similar to the Fourier transform or discrete cosine transform. A necessary step in the development of an FJT compression algorithm is to produce a method for appropriately quantizing the FiT coefficients. The current work undertakes the measurement of the thresholds for the visual detection of the FJT basis functions. These results will be used in the development of a psychovisually-based quantizer. The thresholds were determined by the use of psychovisual observer experiments. It is shown that the detectability of the basis functions is in accord with current knowledge of the spatial frequency sensitivity of the human visual system. However, these results are not predictable from simple psychovisual models, and because of the frequency domain behavior of the FJT, the measured values must be corrected before they can be used in a quantizer. The corrected values imply that the FJT quantizer may not match the properties of ordinary images as well as the psychovisual quantizers used for such methods as the discrete cosine transform.

The ACR-NEMA Standard was published five years ago. Implementations are just now becoming available in a form other than a prototype. Though this seems like a long interval between the initial work and results, the organization responsible for the Standard, the ACR-NEMA Digital Imaging and Communications Standards Committee, has not been idle. Much of the progress which has been made is the result of cooperative work involving industry, the Committee and its Working Groups (WG), and the medical imaging user community. This paper will briefly review the history of the development of what is now a family of ACR-NEMA standards, describe the current activity of the Working Groups, and indicate in what directions the WGs are headed for both new and updated standards.

A compression scheme is described that allows high-definition radiological images with greater than 8-bit intensity resolution to be represented by 8-bit pixels. Reconstruction of the images with their original intensity resolution can be carried out by means of a pipeline architecture suitable for compact, high-speed implementation. A reconstruction system is described that can be fabricated according to this approach and placed between an 8-bit display buffer and the display's video system thereby allowing contrast control of images at video rates. Results for 50 CR chest images are described showing that error-free reconstruction of the original 10-bit CR images can be achieved.

A new decomposition method using image splitting and gray-level remapping has been proposed for image compression, particularly for images with high contrast resolution. The effects of this method are especially evident in our radiological image compression study. In our experiments, we tested the impact of this decomposition method on image compression by employing it with two coding techniques on a set of clinically used CT images and several laser film digitized chest radiographs. One of the compression techniques used was full-frame bit-allocation in the discrete cosine transform domain, which has been proven to be an effective technique for radiological image compression. The other compression technique used was vector quantization with pruned tree-structured encoding, which through recent research has also been found to produce a low mean-square-error and a high compression ratio. The parameters we used in this study were mean-square-error and the bit rate required for the compressed file. In addition to these parameters, the difference between the original and reconstructed images will be presented so that the specific artifacts generated by both techniques can be discerned by visual perception.

Transform based compression methods achieve their effect by taking advantage of the correlations between adjacent pLtels in an image. The increasing use of three-dimensional imaging studies in radiology requires new techniques for image compression. For time-sequenced studies such as digital subtraction angiography, pixels are correlated between images, as well as within an image. By using three-dimensional cosine transforms, correlations in time as well as space can be exploited for image compression. Sequences of up to eight 512 x 512 x 8-bit images were compressed using a single full volume three-dimensional cosine transform, followed by quantization and bit-allocation. The quantization process is a uniform thresholding type and an adaptive three-dimensional bit-allocation table is used. The resultant image fidelity vs. compression ratio was shown to be superior to that achieved by compressing each image individually.

In this paper, we report on an experiment in which various lossy image coding techniques are applied to a digitized radiographic image. The coding schemes are applied to radiograph, digitized using the Konica laser based scanner. Six different lossy coding techniques are employed as follows: full frame transforms, variable length block transform coding, three vector quantization schemes which differ in the degree of adaptivity, and a unified variable length transform coding and adaptive vector quantization. The first vector quantizer uses an universal code book which has been designed from a large set of radiographs; whereas, the second vector quantizer uses a reduced set of code words drawn from the universal codebook. The use of smaller set of code words effectively means that less bits are needed for specifying the codeword label and so decreasing the bit rate requirements. In the last vector quantizer, a new codebook is designed for each image, so decreasing the bit rate required for the label, but increasing the overhead required for transmitting the new code book. The unified variable length transform coding and adaptive vector quantization technique involves variable length transform coding on the code book of the last vector quantizer, thus reducing the overhead for code book transmission.

Color 35mm photographic slides are commonly used in dermatology for education, and patient records. An electronic storage and retrieval system for digitized slide images may offer some advantages such as preservation and random access. We have integrated a system based on a personal computer (PC) for digital imaging of 35mm slides that depict dermatologic conditions. Such systems require significant resources to accommodate the large image files involved. Methods to reduce storage requirements and access time through image compression are therefore of interest. This paper contains an evaluation of one such compression method that uses the Hadamard transform implemented on a PC-resident graphics processor. Image quality is assessed by determining the effect of compression on the performance of an image feature recognition task.

Both CT and MRI images present three-dimensional information as a series of two-dimensional slices. In conventional film sheet display, recovering the third dimension of information requires the physician to mentally reconstruct the volume. This complex process inhibits the understanding of three dimensional registration of both lesions and normal structures of surgical concern. We have developed a technique for the computerized generation of a wire-frame model from the raster image set. By using a 10242 pixel display, any three raster frames plus the wire-frame may be simultaneously displayed. Each raster frames's location is displayed on the wire-frame providing perspective clues as to the three-dimensional relationship between frames. Each raster image may be changed with real time tracking on the wire-frame model. The wire-frame may be formed from either CT or MRI images and takes less than two minutes for a standard 25 slice head series.

This paper addresses the issue of increasing productivity in experimenting with and using image analysis algorithms. Algorithm developers typically experiment with new algorithms in an environment consisting of an editor and compiler on general purpose computers. Such an environment has several disadvantages: (i) It is non-interactive in that the experimenter must go through several steps (e.g. editing, compiling and linking) to see the result of his changes; (ii) It forces the experimenter to concentrate on the details of the compiler instead of the details of the algorithm; (iii) The software so produced frequently demonstrates techniques of programming expediency instead of techniques which lend themselves to parallel implementation. These disadvantages stifle creativity and reduce productivity.

A teaching station for MR technologists based on a personal computer system has been designed and developed. The system is designed mainly to train technologists, who have experiences with other imaging modalities like CT, ultrasound, or nuclear medicine.

The application of digital technologies to chest radiography holds the promise of routine application of intage processing techniques to effect image enhancement. However, due to their inherent spatial resolution, digital chest images impose severe constraints on data storage devices. Compression of these images will relax such constraints and facilitate image transmission on a digital network. We have evaluated image processing algorithms aimed at compression of digital chest images while improving the diagnostic quality of the image. The image quality has been measured with respect to the task of tumor detection. Compression ratios of as high as 2:1 have been achieved. This compression can then be supplemented by irreversible methods.

Progressive image transmission is receiving attention for application in interactive image communication over low-bandwidth channels. In progressive image transmission, a low-resolution image is first transmitted. Upon the user's request, the low-resolution image can be then refined progressively with further transmission until the original image is losslessly reproduced. One important class of encoding for progressive image transmission is based upon pyramid data structures where the intermediate levels correspond to reduced-resolution approximations. Progressive image transmission is achieved by sending the data stored in the pyramid starting from the top level. In this paper, we present a comparison of various pyramid data structures for progressive transmission of medical imagery. These pyramid data structures include the mean, reduced-sum, difference, reduced-difference, S-transform and Gaussian-Laplacian pyramids. The simulations are carried out on a set of digitized radiographic images. The simulation results demonstrate that the reduced-difference pyramid achieves the best performance in terms of three criteria: the equivalent entropy, rate distortion performance and total lossless transmission hit rate. Furthermore, the quality of the intermediate level images is seen to be improved by using an appropriate interpolation function.

Vector quantization has been widely used in image encoding systems and speech recognition systems science 1980's. In this paper, three kinds of vector quantization approaches are introduced, which are based on neural networks. The principles of using neural networks to improve the performance of vector quantization are described. Because of high parallel computation, learning function, high fault tolerance and selforganizing capability of neural networks, the performance of vector quantizers is improved.

The use of coherent fiber-optic bundles as image conduits has increased significantly in recent years. Fiber bundles permit optical inspection of small, previously inaccessible places. Applications in industrial vision, medicine, aerospace instrumentation and other areas are developing rapidly. The resolution of the image transmitted by fiber bundles is primarily limited by the diameter of the individual fibers. The image is further distorted by a number of factors such as the wide and occasionally skewed point spread function of each fiber and the non-uniformity of fiber characteristics throughout the bundle. Additionally, the appearance of the characteristic circular or hexagonal fiber walls is distracting and objectionable to viewers and increases the complexity of machine vision algorithms. We report a sequence of image processing steps on digitized coherent bundle images which results in complete removal of the fiber wall patterns from the transmitted images; along with corrections for factors creating distortion. The processing steps essentially involve 1) calibrating the fiber bundle under uniform illumination, 2) locating the center of each fiber and extracting the centroid value, 3) disseminating the centroid values onto a uniform grid, and 4) reconstructing the complete image by interpolating between nearest neighbor points on the uniform grid. The process has been applied to both monochrome and color images by applying the above steps individually to each of the red, green and blue components. Future work needs to be performed to correct for a visible moire pattern and the inherent smoothing caused by the interpolation process.

The medical imaging field relies increasingly on imaging and graphics techniques in diverse applications with needs similar to (or more stringent than) those of the military, industrial and scientific communities. However, most image processing and graphics systems available for use in medical imaging today are either expensive, specialized, or in most cases both. High performance imaging and graphics workstations which can provide real-time results for a number of applications, while maintaining affordability and flexibility, can facilitate the application of digital image computing techniques in many different areas. This paper describes the hardware and software architecture of a medium-cost floating-point image processing and display subsystem for the NeXT computer, and its applications as a medical imaging workstation. Medical imaging applications of the workstation include use in a Picture Archiving and Communications System (PACS), in multimodal image processing and 3-D graphics workstation for a broad range of imaging modalities, and as an electronic alternator utilizing its multiple monitor display capability and large and fast frame buffer. The subsystem provides a 2048 x 2048 x 32-bit frame buffer (16 Mbytes of image storage) and supports both 8-bit gray scale and 32-bit true color images. When used to display 8-bit gray scale images, up to four different 256-color palettes may be used for each of four 2K x 2K x 8-bit image frames. Three of these image frames can be used simultaneously to provide pixel selectable region of interest display. A 1280 x 1024 pixel screen with 1: 1 aspect ratio can be windowed into the frame buffer for display of any portion of the processed image or images. In addition, the system provides hardware support for integer zoom and an 82-color cursor. This subsystem is implemented on an add-in board occupying a single slot in the NeXT computer. Up to three boards may be added to the NeXT for multiple display capability (e.g., three 1280 x 1024 monitors, each with a 16-Mbyte frame buffer). Each add-in board provides an expansion connector to which an optional image computing coprocessor board may be added. Each coprocessor board supports up to four processors for a peak performance of 160 MFLOPS. The coprocessors can execute programs from external high-speed microcode memory as well as built-in internal microcode routines. The internal microcode routines provide support for 2-D and 3-D graphics operations, matrix and vector arithmetic, and image processing in integer, IEEE single-precision floating point, or IEEE double-precision floating point. In addition to providing a library of C functions which links the NeXT computer to the add-in board and supports its various operational modes, algorithms and medical imaging application programs are being developed and implemented for image display and enhancement. As an extension to the built-in algorithms of the coprocessors, 2-D Fast Fourier Transform (FF1), 2-D Inverse FFF, convolution, warping and other algorithms (e.g., Discrete Cosine Transform) which exploit the parallel architecture of the coprocessor board are being implemented.

A receiver operating characteristic (ROC) experiment is used to compare the diagnostic performance of analog film to digital hardcopy and compressed hardcopy. The compression is done using a full-frame bit-allocation algorithm carried through on a custom designed board residing on a SUN computer bus. Image set included thirthy-one radiographs with septal lines (n=18) and/or parenchymal nodules (n=14), with six radiographs that had neither abnormality. Six radiologists viewed the image set under each modality. There is no statistically significant difference between the three modalities.

Signal processing concepts are used to study the problems of obtaining continuous image representations in terms of B-spline basis functions, and performing reconstructions at various magnifications. Filters that compute direct or indirect B-spline transformations are derived and characterized in terms of their z-transforms for polynomial splines of any order. Efficient recursive implementations of these operators are proposed. Recursive filters that efficiently solve the problem of approximating a signal through the use of least squares splines are also derived. This last procedure is analogous to the application of an anti-aliasing filter prior to decimation for the representation of a signal with fewer samples.

Image compression at rates of 10:1 or greater could make PACS much more responsive and economically attractive. This paper describes a protocol for subjective and objective evaluation of the fidelity of compressed/decompressed images to the originals and presents the results ofits application to four representative and promising compression methods. The methods examined are predictive pruned tree-structured vector quantization, fractal compression, the discrete cosine transform with equal weighting of block bit allocation, and the discrete cosine transform with human visual system weighting of block bit allocation. Vector quantization is theoretically capable of producing the best compressed images, but has proven to be difficult to effectively implement. It has the advantage that it can reconstruct images quickly through a simple lookup table. Disadvantages are that codebook training is required, the method is computationally intensive, and achieving the optimum performance would require prohibitively long vector dimensions. Fractal compression is a relatively new compression technique, but has produced satisfactory results while being computationally simple. It is fast at both image compression and image reconstruction. Discrete cosine iransform techniques reproduce images well, but have traditionally been hampered by the need for intensive computing to compress and decompress images. A protocol was developed for side-by-side observer comparison of reconstructed images with originals. Three 1024 X 1024 CR (Computed Radiography) images and two 512 X 512 X-ray CT images were viewed at six bit rates (0.2, 0.4, 0.6, 0.9, 1.2, and 1.5 bpp for CR, and 1.0, 1.3, 1.6, 1.9, 2.2, 2.5 bpp for X-ray CT) by nine radiologists at the University of Washington Medical Center. The CR images were viewed on a Pixar II Megascan (2560 X 2048) monitor and the CT images on a Sony (1280 X 1024) monitor. The radiologists' subjective evaluations of image fidelity were compared to calculations of mean square error (MSE), normalized mean square error (NMSE), percentage mean square error (PMSE), and fractal normalized mean square error (FMSE) for each compression method and bit rate.