Conventional surface fitting methods give twisted surfaces and complicates capping closures. This is a typical character of surfaces that lack rectangular topology. We suggest an algorithm which overcomes these limitations. The analysis of the algorithm is presented with experimental results. This algorithm assumes the mass center lying inside the object. Both capping closure and twisting are results of inadequate information on the geometric proximity of the points and surfaces which are proximal in the parametric space. Geometric proximity at the contour level is handled by mapping the points along the contour onto a hyper-spherical space. The resulting angular gradation with respect to the centroid is monotonic and hence avoids the twisting problem. Inter-contour geometric proximity is achieved by partitioning the point set based on the angle it makes with the respective centroids. Avoidance of capping complications is achieved by generating closed cross curves connecting curves which are reflections about the abscissa. The method is of immense use for the generation of the deep cerebral structures and is applied to the deep structures generated from the Schaltenbrand- Wahren brain atlas.

Recent advances in three-dimensional (3D) imaging techniques have expanded the scope of applications of volume visualization to many areas such as medical imaging, scientific visualization, robotic vision, and virtual reality. Advanced image filtering, enhancement, and analysis techniques are being developed in parallel in the field of digital image processing. Although the fields cited have many aspects in common, it appears that many of the latest developments in image processing are not being applied to the fullest extent possible in visualization. It is common to encounter the use of rather simple and elementary image pre- processing operations being used in visualization and 3D imaging applications. The purpose of this paper is to present an overview of selected topics from recent developments in adaptive image processing and demonstrate or suggest their applications in volume visualization. The techniques include adaptive noise removal; improvement of contrast and visibility of objects; space-variant deblurring and restoration; segmentation-based lossless coding for data compression; and perception-based measures for analysis, enhancement, and rendering. The techniques share the common base of identification of adaptive regions by region growing, which lends them a perceptual basis related to the human visual system. Preliminary results obtained with some of the techniques implemented so far are used to illustrate the concepts involved, and to indicate potential performance capabilities of the methods.

Volume visualization is the technique of displaying two dimensional projections of three dimensional data. The data is acquired from a medical scanner, like MRI, CT, SPECT, or US scanners. Visualizing a given three dimensional medical dataset can be done by surface rendering algorithms or by direct volume rendering algorithms. Surface rendering algorithms require an intermediate geometric representation and are therefore less attractive. In our approach volume rendering is used. To improve image quality of such projections of the volume data, special care should be taken to (a) the interpolation step, (b) the estimation of the local gradient and (c) the assignment of opacity values at sample positions. These aspects are addressed in this paper.

Visualization aids in the evaluation of a 3-D data set, if a -- perhaps imperfect -- separability of different objects by their function value is assumed. An interpretation without explicit analysis can be carried out by reflection/transmission (RT) volume rendering where voxels, depending on their surface probability, reflect and transmit light. Color coding voxels in RT rendering in order to attribute the surface to a specific object is not easy because the voxel value depends on the unknown objects meeting at the surface. This problem does not arise, if an emission/absorption (EA) method is applied. Emission and absorption of colored light are associated with the probability of a voxel belonging to a certain object. We combined both approaches to exploit the RT renderer's ability to reveal surface details and the EA renderer's ability to differentiate objects. A user-defined operator decides whether a voxel belongs to the background, to a surface, or to an object. Background voxels are skipped. Object voxels contribute to accumulated light emission. Surface voxels reflect light from an outside light source. Color is attributed only to object voxels, but it is possible to let colored object voxels shine through semi-transparent surface voxels. CT sequences were visualized in order to display structures such as the ventricular system or areas of pathological processes. Although the differentiation of these structures was not complete, the display showed enough detail to convey the relevant shape to the user.

The frequency domain volume rendering approach generates line integral projections of spatial data utilizing the Fourier Projection-Slice Theorem. Compared to spatial domain rendering, the computational cost is reduced from O(N³) to O(N²logN). However, this rendering technique has limited ability for interaction or manipulation of the rendered data. One of the common operations in the manipulation of volume data is planar-clipping or sectioning which reduces the visual complexity of the data and improves our ability to visualize information contained within the volume data. A method to implement spatial volume clipping in the frequency domain is described.

The identification of cortical sulci is of great importance. In neurosurgical procedures any target in the cranium can be accessed by following the corridors of the sulci and fissures. The fusion of functional and anatomical data also requires the identification of sulci. Several approaches have been proposed for segmentation of the cortical surface and identification of sulci and fissures. Most of them are bottom-up. They work satisfactorily provided that the sulci are well discernible on MRI images, limiting their use to some major sulci and fissures, such as the central sulcus, interhemispheric fissure, or Sylvian fissure. We propose a sulcal model based approach, overcoming some of the above limitations. The sulcal model is derived from two brain atlases: Co-Planar Stereotaxic Atlas of the Human Brain by Talairach- Tournoux (TT), and Atlas of Cerebral Sulci by Ono-Kubik-Abernathey (OKA). The OKA atlas contains 403 patterns for 55 sulci along with their incidence rates of interruptions, side branches, and connections. An electronic version of the OKA atlas was constructed, quantitatively enhanced by placing the sulcal patterns in a stereotactic space. The original patterns from the OKA atlas were digitized, converted into geometric representation, placed in the Talairach stereotactic space, preregistered with the TT atlas, and integrated with a multi- atlas, multi-dimensional neuroimaging system developed by our group. The registration of any atlas with the clinical data automatically registers all atlases with this data. This way the sulcal patterns can be superimposed on data, indicating approximate locations of sulci on images. The approach proposed here provides a simple and real-time registration of the sulcal patterns with clinical data, and an interactive identification and labeling of sulci. This approach assists rather the medical professional, instead of providing a complete automated extraction of a few, primary sulci with certain accuracy, where a higher accuracy usually demands a longer time of pattern extraction.

The capability of today's clinical scanners to create large quantities of high resolution and near isotropically sampled volume data, coupled with a rapidly improving performance/price ratio of computers, has created the challenge and feasibility of creating new ways to explore cross- sectional medical imagery. Perspective volume rendering (PVR) allows an observer to 'fly- through' image data and view its contents from within for diagnostic and treatment planning purposes. We simulated flights through 14 data sets and, where possible, these were compared to conventional endoscopy. We demonstrated colonic masses and polyps as small as 5 mm, tracheal obstructions and precise positioning of endoluminal stent-grafts. Simulated endoscopy was capable of generating views not possible with conventional endoscopy due to its restrictions on camera location and orientation. Interactive adjustment of tissue opacities permitted views beyond the interior of lumina to reveal other structures such as masses, thrombus, and calcifications. We conclude that PVR is an exciting new technique with the potential to supplement and/or replace some conventional diagnostic imaging procedures. It has further utility for treatment planning and communication with colleagues, and the potential to reduce the number of normal people who would otherwise undergo more invasive procedures without benefit.

The conventional diagnosis of breast cancer by a combination of mammography and physical examination has limited efficacy. Only 20 - 30% of the suspicious breast lesions biopsied are actually malignant. Breast MRI (BMRI), using intravenous contrast injection to detect and characterize breast lesions, has shown promise in several studies. We are developing a 3D image visualization and analysis system to assist radiologists to detect breast cancer in BMRI. Dynamic contrast-enhanced MRI has emerged to become an effective procedure in BMRI for the diagnosis and management of breast carcinoma. A 3D image visualization and analysis system that allows a radiologist to rapidly search through BMRI images and visualize three- dimensional volume of a whole breast has been developed. Three dimensional breast images were constructed from 2D slices. The system can register and subtract a pre-contrast image from each of the time sequenced post-contrast images. The dynamic time sequences of the breast before and after the contrast administration can be visualized. Suspicious lesions can be detected based on the dynamic time-sequence changes in images. The system also allows interactive manipulation of images for viewing from different angles and examination of 2D projections at specific locations. The developed 3D image visualization and analysis system provided radiologists an efficient way to analyze MR breast images. Our automated detection scheme has the potential to accurately detect suspicious lesions and can be an effective tool to facilitate the clinical application of MRI for breast imaging.

Specialized high speed volume rendering tools and image preprocessing methods required for the automated interactive microscope (AIM) are described. AIM will allow the biologist to perform 'closed loop' experiments, meaning tea the biologist will be able to view preliminary results to allow critical observations which facilitate the changing of the course of image data collection dynamically during the experiment. To facilitate this process, we are developing high-speed volume rendering tools utilizing 3-D texture mapping hardware. In addition, specific preprocessing methods are described which enable the seeing of 3-D structures in volume renderings of differential interference contrast (DIC) volume images. Without these methods, structures of interest, which are critically related to the experiment, would be obscured as though in a dense fog.

Shape-based interpolation as applied to binary images causes the interpolation process to be influenced by the shape of the object. It accomplishes this by first applying a distance transform to the data. This results in the creation of a gray-level data set in which the value at each point represents the minimum distance from that point to the surface of the object. (By convention, points inside the object are assigned positive values; points outside are assigned negative values.) This distance transformed data set is then interpolated using linear or higher order interpolation and is then thresholded at a distance value of 0 to produce the interpolated binary data set. In this paper, we describe a new method that extends shape-based interpolation to gray-level input data sets. This generalization consist of first lifting the n-dimensional image data to represent it as a surface, or equivalently as a binary image, in an (n plus 1)- dimensional space. The binary shape-based method is then applied to this image to create an (n plus 1)-dimensional binary interpolated image. Finally, this image is collapsed (inverse of lifting) to create the n-dimensional interpolated gray-level data set. We have conducted several evaluation studies involving patient CT and MR data as well mathematical phantoms. They all indicate that the new method produces more accurate results than conventional gray-level interpolation methods, although at the cost of increased computation.

We have developed a 3D reach-in tool to manually reconstruct 3D cortical surface patches from 2D brain atlas images. The first application of our cortex editor is building 3D functional maps, specifically Brodmann's areas. This tool may also be useful in clinical practice to adjust incorrectly mapped atlas regions due to the deforming effect of lesions. The cortex editor allows a domain expert to control the correlation of control points across slices. Correct correlation has been difficult for 3D reconstruction algorithms because the atlas slices are far apart and because of the complex topology of the cortex which differs so much from slice to slice. Also, higher precision of the resulting surfaces is demanded since these define 3D brain atlas features upon which future stereotactic surgery may be based. The cortex editor described in this paper provides a tool suitable for a domain expert to use in defining the 3D surface of a Brodmann's area.

In pediatric cardiac angiography, there are several peculiarities such as limitation of both x-ray dose and the amount of contrast medium in comparison with conventional angiography. Due to these peculiarities, the catheter examinations are accomplished in a short time with biplane x- ray apparatus. Thus, it is often difficult to determine 3D structures of blood vessels, especially those of pediatric anomalies. Then a new 3D reconstruction method based on selective biplane angiography was developed in order to support diagnosis and surgical planning. The method was composed of particular reconstruction and composition. Individual 3D image is reconstructed with the particular reconstruction, and all 3D images are composed into standard coordinate system in the composition. This method was applied to phantom images and clinical images for evaluation of the method. The 3D image of the clinical data was reconstructed accurately as its structures were compared with the real structures described in the operative findings. The 3D visualization based on the method is helpful for diagnosis and surgical planning of complicated anomalies in pediatric cardiology.

The goal of our work is to help radiologists remove obscuring structures from a large volume of computed tomography angiography (CTA) images by editing a small number of sections prior to three-dimensional (3D) reconstruction. We combine automated segmentation of the entire volume with manual editing of a small number of sections. The segmentation process uses a neural network to learn thresholds for multilevel thresholding and a constraint- satisfaction neural network to smooth the boundaries of labeled segments. Following segmentation, the user edits a small number of images by pointing and clicking, and then a connectivity procedure automatically selects corresponding segments from other sections by comparing adjacent voxels within and across sections for label identity. Our results suggest that automated segmentation followed by minimal manual editing is a promising approach to editing of CTA sequences. However, prerequisites to clinical utility are evaluation of segmentation accuracy and development of methods for resolution of label ambiguity.

Three dimensional (3-D) rendering is a useful technique for visualizing anatomical structures. However, the rendering process is computationally expensive. In this paper, we present a 3D renderer which can render 3D medical images in real time but requires little data pre- processing, a small amount of additional memory and no special hardware. The system has the capacity of rendering multiple objects so that the complex relationship between different segmented anatomical structured can be illustrated in the rendered view.

Image-guided neurosurgery (IGNS), in which anatomical images generated from patient MRI or CT scans provide surgical guidance, is now routinely employed in numerous institutions. However, IGNS systems generally lack the ability to display functional data, a significant shortcoming for many types of procedures. We have enhanced the IGNS system used at our institution (the ISG viewing wand) allowing the surgeon to display and interact with patient electroencephalography (EEG) data in the operating room. The surgeon can: determine 3D electrode locations; display electrode locations with respect to the underlying 3D patient anatomy obtained from MRI; visualize the EEG potential field map interpolated onto the scalp; graphically analyze the time evolution of these maps; and view the location of equivalent sources within the patient cerebral structures. Display of EEG information is clinically significant in cases involving the surgical treatment of epilepsy, where EEG data plays an important role in characterizing and localizing epileptic foci, both preoperatively and during the operation.

In interactive, image-guided surgery (IIGS) we use stacked slice tomographic image sets as three dimensional maps of the patient's anatomy. Such sets can provide exquisite information on bony anatomy (computed tomography), soft tissue structure and lesion definition (magnetic resonance) or function (positron emission tomography). However, none of these tomographic sets clearly show the location and extent of vascular structures, which may be of critical importance to the surgical process. Vascular information can be obtained from conventional x- ray angiography (XRA), a form of projection imaging, or from tomographic scans sensitive to flow such as magnetic resonance angiography (MRA) or computed tomography angiography (CTA). Projection images show the extent and intersection of vessels at high resolution but lose the three-dimensional relationship of the vessels during image formation. The high resolution available from conventional angiograms makes them attractive for the localization of small aneurysms and other vascular anomalies, but the projection nature of the image makes displaying surgical position difficult. We have developed techniques for interactively displaying surgical position on both topographically derived vascular sets such s CTA and MRA and on projection sets such as XRA. In the XRA images,homologous points are determined on the film and on the patient. Since we use extrinsic, bone-implanted fiducial markers for our surgical guidance system, these markers can serve as most, if not all, of our homologous points and can be localized very precisely on the film and in 3-space. A homogeneous transform matrix (HTM) is constructed to provide a best 'first-guess' of position. The HTM is decomposed into nine spatial parameters which represent the XRA process and those parameters are optimized using Powell's Conjugate Direction Method. In the tomographic vascular images, the issue is not one of registration but one of display. We presently track probe position on tomographic slices but the information inherent in vascular tomograms is masked by a raster-slice display; the vessel's path and extent are difficult to discern. Vascular path and extent are best displayed as a rotoscope of projection images created from the raster-slice set. By 'rotating' a series of projection images the three- dimensional geometry is discerned by motion parallax.

In gamma unit radiosurgery treatment planning, dose delivery is based on the unit 'shot,' a distribution of dose approximately spherical in shape. Multiple shots are used to cover different parts of a given target region. Effective 3D optimization for gamma unit treatment has not been previously reported. In this article, a novel optimization method is introduced based on medial axis transformation techniques. Given a defined target volume, the target's medial axis, which uniquely characterizes the target, is used to determine the optimal shot positions and sizes. In using the medial axis, the 3D optimization problem is reduced to a 1D optimization, with corresponding savings in computational time and mathematical complexity. In addition, optimization based on target shape replicates and automates manual treatment planning, which makes the process easily understandable. Results of optimal plans and the corresponding dose distributions are presented. The relationship between skeleton disks and the dose distributions they predict are also discussed.

Wavelet-based image compression is proving to be a very effective technique for medical images, giving significantly better results than the JPEG algorithm. A novel scheme for encoding wavelet coefficients, termed set partitioning in hierarchical trees, has recently been proposed and yields significantly better compression than more standard methods. We report the results of experiments comparing such coding to more conventional wavelet compression and to JPEG compression on several types of medical images.

A neural network based framework has been developed to search for an optimal wavelet kernel that is most suitable for a specific image processing task. In this paper, we demonstrate that only the low-pass filter, h_u, is needed for orthonormal wavelet decomposition. A convolution neural network can be trained to obtain a wavelet that minimizes errors and maximizes compression efficiency for an image or a defined image pattern such as microcalcifications on mammograms. We have used this method to evaluate the performance of tap-4 orthonormal wavelets on mammograms, CTs, MRIs, and Lena image. We found that Daubechies' wavelet (or those wavelets possessing similar filtering characteristics) produces satisfactory compression efficiency with the smallest error using a global measure (e.g., mean- square-error). However, we found that Harr's wavelet produces the best results on sharp edges and low-noise smooth areas. We also found that a special wavelet, whose low-pass filter coefficients are (0.32252136, 0.85258927, 0.38458542, -0.14548269), can greatly preserve the microcalcification features such as signal-to-noise ratio during a course of compression. Several interesting wavelet filters (i.e., the g filters) were reviewed and explanations of the results are provided. We believe that this newly developed optimization method can be generalized to other image analysis applications where a wavelet decomposition is employed.

Wavelet-based image compression is receiving significant attention because of its potential for good image quality at low bit rates. In this paper, we describe and analyze a lossy wavelet compression scheme that uses direct extensions of the JPEG quantization and Huffman encoding strategies to provide high compression efficiency with reasonable complexity. The focus is on the compression of 12-bit medical images obtained from computed radiography and mammography, but the general methods and conclusions presented in this paper are applicable to a wide range of image types.

In this paper different methods for the quantization of wavelet transform coefficients are compared in view of medical imaging applications. The goal is to provide users with a comprehensive and application-oriented review of these techniques. The performance of four quantization methods (namely standard scalar quantization, embedded zerotree, variable dimension vector quantization and pyramid vector quantization) are compared with regard to their application in the field of medical imaging. In addition to the standard rate-distortion criterion, we took into account the possibility of bitrate control, the feasibility of real-time implementation, the genericity (for use in non-dedicated multimedia environments) of each approach. In addition, the diagnostical reliability of the decompressed images has been assessed during a viewing session and with the help of a specialist. Classical scalar quantization methods are briefly reviewed. As a result, it is shown that despite the relatively simple design of the optimum quantizers, their performance in terms of rate-distortion tradeoff are quite poor. For high quality subband coding, it is of major importance to exploit the existing zero-correlation across subbands as proposed with the embedded zerotree wavelet (EZW) algorithm. In this paper an improved EZW-algorithm is used which is termed embedded zerotree lossless (EZL) algorithm -- due to the importance of lossless compression in medical imaging applications -- having the additional possibility of producing an embedded lossless bitstream. VQ based methods take advantage of statistical properties of a block or a vector of data values, yielding good quality results of reconstructed images at the same bitrates. In this paper, we take in account two classes of VQ methods, random quantizers (VQ) and geometric quantizers (PVQ). Algorithms belonging to the first group (the most widely known being that developed by Linde-Buzo-Gray) suffer from the common drawback of requiring a computationally demanding training procedure in order to produce a codebook. The second group represents an interesting alternative, based on the multidimensional properties of the distribution of the source to code. In particular a pyramid vector quantization has been taken into account. Despite being based on the implicit geometry of independent and identically distributed (i.i.d.) Laplacian sources, this method proved to achieve good results with other distributions. Tests show that zerotree yields the most promising results in the rate- distortion sense. Moreover, this approach allows an exact rate control and has the possibility of a progressive bitstream which can be used either for data browsing or up to a lossless representation of the input image.

A new task-oriented image quality metric is used to quantify the effects of distortion introduced into magnetic resonance images by lossy compression. This metric measures the similarity between a radiologist's manual segmentation of pathological features in the original images and the automated segmentations performed on the original and compressed images. The images are compressed using a general wavelet-based lossy image compression technique, embedded zerotree coding, and segmented using a three-dimensional stochastic model-based tissue segmentation algorithm. The performance of the compression system is then enhanced by compressing different regions of the image volume at different bit rates, guided by prior knowledge about the location of important anatomical regions in the image. Application of the new system to magnetic resonance images is shown to produce compression results superior to the conventional methods, both subjectively and with respect to the segmentation similarity metric.

Compression of medical images to reduce their storage and transmission bandwidth requirements is of great interest in the implementation of systems such as the picture archiving and communication system (PACS). Direct application of discrete cosine transform (DCT) coding to medical images such as CT or MRI images is not effective as the characteristics of such medical images are not exploited. Firstly, the noisy background in medical images exhibits largely uncorrelated data which is difficult to compress using transform coding. Secondly, the overhead in representing the background information using fixed block-size transform coding is inefficient. A novel adaptive coding algorithm is proposed to yield high compression rate for medical images. The proposed algorithm is a two-stage process: the first stage (pre-processing stage) attempts to remove the background noise and identifies the border of the medical data by using a visual mask; the second stage (encoding stage) uses an adaptive block-size DCT coding algorithm to compress the image data. The proposed coding algorithm is evaluated and compared with the JPEG baseline algorithm where results on the compression ratio and peak signal-to-noise ratio (PSNR) are presented. The results show that the proposed coding algorithm achieves a higher compression rate than the JPEG baseline algorithm. In addition, the PSNR values of the new coder is marginally higher than the results obtained with the JPEG coder.

Almost all present-day lossy image-compression methods are so-called waveform coders, i.e. they attempt to approximate the original image waveform as closely as possible with the available number of bits. At medium-to-low compression ratios this can usually be achieved quite well, but at high compression ratios clearly visible artefacts may be introduced into the coded images. During the compression of medical x-ray images we noticed that the data that suffer most from a high degree of compression are the noise-like components present in these images. The loss of these components is found to depreciate the perceived image quality. This paper proposes a method for modeling the noise components removed due to lossy compression and for regenerating these components during the decompression of the image.

Radiology uses large images and series of images that can consume large amounts of storage or of communication bandwidth in the utilization of those images. An update to the standard for compressing radiology images is being considered by the medical imaging compression standards committee. A standard for compression of radiology images is proposed for consideration. The proposed standard uses four basic techniques to achieve very high quality reconstructed images: (a) image decomposition into high frequency and low frequency elements, (b) lapped orthogonal discrete cosine transforms, (c) local quantization, and (d) Huffman encoding. Degenerate forms of the standard include the JPEG standard, already included in the DICOM medical image interchange standard. The proposed standard is a departure from the JPEG standard because of the low quality of the baseline JPEG lossy compression. At the same time, much of the hardware and software that have been used for JPEG compression are applicable to the proposed standard technique. A preprocessing step changes the format of the image to a form that can be processed using JPEG compression. A post-processing step after the JPEG restoration will restore the image. The proposed standard does not permit many techniques that have been used in the past. In particular, decomposition by the level of the significant bits is not permitted, the only transform permitted is the lapped orthogonal discrete cosine transform, the block size of the transform is limited to 8 by 8, and only Huffman coding is allowed. There are many variations that can be used in compression. This proposal allows some variations, but restricts many other variations in the interest of simplicity for the standard. The quality of the compression is very good. The extra complexity in the standard to allow more variations is not warranted.

To achieve higher compression ratios in medical images while preserving quality, a new fractal method is developed, examined, and tested in this paper. The basic idea of fractal image compression is to reduce the similarity redundancy by identifying a given image with the fixed-point of an appropriate partitioned iterated function system (PIFS), that consists of a set of contractive affine transformations. Since conventional PIFS models use only affine transformations to represent similarity for the whole image, the quality of the re-created images is quite limited in some cases. Because the grayscale in most images is dependent on location, it is not sufficient to describe the relationship using linear transforms when there are complex textures present. Effective interpolation polynomials should adapt to the nature of the underlying texture; that is the basis of the new method. The new fractal image-compression algorithm uses adaptive PIFS (APIFS), that is based on variants of affine transformations and lossless compression methods. Polynomials of various orders are used to represent adaptively the similarity of grayscale based on the local details of the image, and the contractive condition for the generalized transformation is shown to hold. In this approach, quadratic and linear models are applied adaptively to the contractive transformations. The variants of the affine transform are used where similarities can be identified, while lossless compression techniques are used for those local areas in which similar domains do not exist or cannot be found. Preliminary experiments indicate that the APIFS model has the potential to increase the useful compression ratio. Experiments with medical images indicate that this new algorithm can be extended to yield a compression ratio of about 30:1 without perceptible degradation.

As medical image data are acquired at higher spatial, gray-scale, and temporal resolution, with increasing image dimensionality and vectorial components per image element, the need to be able to compress them becomes increasingly important, especially for the practical utilization of PACS. Further, a compression method that allows an efficient in-memory representation would be very useful for those processing situations in which the non-compressed data are beyond RAM resources. Consider an n-dimensional (nD) scene intensity function (m- valued) as a digital surface in an (n plus m)D space as described below. The approach leads to a spectrum of lossless and lossy compression methods. For a 2D, scalar valued (m equals 1) scene S, for example, the 3D digital surface representing S is simply the graph of the scene intensity function. The first two coordinates of points on the surface represent the coordinates of a pixel in S and the third coordinate represents the pixel's intensity. In this example, if m equals 2, then the third and the fourth coordinates of the points in the 4D surface represent the two pixel values. The digital surfaces obtained from scenes via this process (referred to as lifting as described below) have certain distinct topological properties not shared by general digital surfaces (for example, they do not enclose holes). These properties allow us to encode them elegantly. For example, the entire surface (for any finite n greater than 0, m greater than 0) can be represented by a Hamiltonian path of surface elements in the graph representing the surface. This path is chain-encoded facilitated by the small number of possible configurations and orientations of adjacent surface elements. Our (very) preliminary results indicate lossless compression ratios of 4.6:1 for CT slices and 1.7:1 for CR scalar-valued scenes (n equals 2, m equals 1) representing 2D scenes. Our preliminary implementation uses simple chain codes that do not exploit the repetitive patterns in the Hamiltonian path. When this is done, and the implementation is extended to scenes with n plus m greater than 3, we expect a higher degree of compression due to the full exploitation of the spatio-temporal and vectorial component coherence. This paper opens a new direction for data compression not taken in previous research. It seems to capture regional information about the uniformity of the intensity distribution better than other methods. It also extends the current shape-based methods used in image interpolation to include compression.

The current trend of using ROC style observer performance studies to evaluate image compression schemes is inefficient and has greatly limited our ability to apply image compression to radiographic images. Because of this, we are developing more efficient automated techniques for assessing the loss of image quality due to compression. The figure- of-merit (FOM) presented here is based on plausible hypotheses about what factors are important in observer performance and evaluates not only the magnitude of difference images but also their structure. This FOM should avoid the limitations of simpler measures such as mean-square-error as well as the limitations and complexity of measures based on psychovisual theory. We applied our FOM to portions of compressed mammograms that had been previously evaluated in a just-noticeable-difference study, and to CT images that had been degraded by compression at various compression ratios with the JPEG algorithm or through other processes. In general, we found that this FOM offers many advantages over other common FOMs and demonstrates the feasibility of developing efficient measures. Although no one-dimensional measure can be expected to be highly correlated to observer performance, for small levels of image degradation we expect this measure to be useful in placing limits on the loss of observer performance.

To date, the lack of adequate electronic display devices has been the primary factor limiting the use of digital radiographic images. Flat panel display technology using field emissive cathodes in microvacuum cells has emerged as a potential solution. However, the performance characteristics of low-voltage cathodoluminescence in the range below 1 kV is not well understood. In this paper, we discuss issues concerning the prediction of phosphor efficiency, and describe the processes of light generation and transport which influence the performance of phosphor screens.

We investigate a simulation tool for the optimization of soft copy displays in radiology. The digital image is traced through the individual components of a black and white cathode ray tube (CRT) monitor and the luminance image observed at the glass faceplate is simulated. The simulated images can be evaluated numerically or rendered on film by a high-resolution printer for viewing. We validated the program simulating a real existing monitor by comparing the results with measured values, as well as by visually comparing the actual image with the simulated one. The gross properties like luminance, dynamic range, and spatial resolution are sufficiently well described. The visual impression of the simulated image is very similar to that of the real soft copy. We investigate the influence of individual parameters on image quality. We find that the bandwidth of the video amplifier has to be larger than half the pixel rate. We demonstrate the influence of the electron beam spot size on spatial resolution. It is shown how the spatial resolution depends on phosphor luminous efficiency and on glass transmission. Furthermore, for a given target display curve, it is found that the current-to- voltage relationship of the electron gun influences the number of perceived gray values. Finally we discuss phosphor noise in context with dynamic range.

The issue of the spatial resolution required in order to present diagnostic quality digital images, especially for softcopy reporting, has received much attention over recent years. The aim of this study was to compare the diagnostic performance reporting from hardcopy and optimized softcopy image presentations. One-hundred-fifteen radiographs of the hand acquired on a photostimulable phosphor computed radiography (CR) system were chosen as the image material. The study group was taken from patients who demonstrated subtle erosions of the bone in the digits. The control group consisted of radiologically normal bands. The images were presented in three modes, the CR system's hardcopy output, and softcopy presentations at full and half spatial resolutions. Four consultant radiologists participated as observers. Results were analyzed using the receiver operating characteristic (ROC) technique, and showed a statistically significant improvement in observer performance for both softcopy formats, when compared to the hardcopy presentation. However, no significant difference in observer performance was found between the two softcopy presentations. We therefore conclude that, with appropriate attention to the processing and presentation of digital image data, softcopy reporting can, for most examinations, provide superior diagnostic performance, even for images viewed at modest (1 k²) resolutions.

The proposed standard defines a gray-scale display function for monochrome image presentation devices, such as cathode-ray-tube (CRT) monitor/display-controller systems and digital laser image printers. The display function is based on perceptual linearization. It is defined for the luminance range of 0.05 to 4000 cd/m². The standard provides a mathematical formula for the standard display function as well as a table of the luminance levels of the just-noticeable differences (JNDs) of a standard target. The standard facilitates similarity in gray-scale between different image display devices independent of their luminance. The standard does not eliminate the use of application-specific display functions, but rather assures their effectiveness through the standard display function. To realize conformance of an image presentation device with the display function standard, standard test patterns are defined. The patterns are used to measure the luminance of the display or optical density of a print as a function of digital input. By proper interpolation of the inverted measured characteristic function of the display system and inserting the desired standard luminance values for every digital input, a transformation can be computed so that the display system may conform with the display function standard. The standard also defines the dynamic range of an image presentation device as the number of JNDs that theoretically can be realized for a given digitization resolution and the standard target. Interpolation processes and potential conformance measures for the standard are discussed. The proposed display function standard is essential for the proper realization of image presentation services according to the DICOM Standard. The relation of the standard display function to DICOM-defined look-up tables is outlined.

This study compared diagnostic accuracy and image quality for laser imaging film from two systems: traditional 'wet' using chemical processing and the new 'DryView' system from 3M which is wet-chemistry-free. Three separate ROC studies (for CT, MRI, and US) were conducted. The 'wet' and 'dry' laser film imaging systems were connected in parallel and identical images for each case were printed on both systems. For each of the 3 studies, 10 radiologists reviewed each case, reporting diagnostic decision confidence. Evaluations of image quality (overall quality, visibility, sharpness, color, contrast) were also obtained. In all three studies, there were no major differences in diagnostic accuracy (ROC Az) for 'wet' vs 'dry' films, although performance was on average higher for the 'dry' film in all three modalities. Judgments of image quality were comparable for 'wet' and 'dry' films for all three modalities. There were no significant differences in viewing time in any of the studies. Thus, 'dry' processing represents a very useful, cost-effective and pollution-free alternative to currently used 'wet' laser imaging systems.

The data types of graphics, images, audio and video or collectively multimedia are becoming standard components of most computer interfaces and applications. Medical imaging in particular will be able to exploit these capabilities in concert with the database engines or 'information furnaces' that will exist as part of the information superhighway. The ability to connect experts with patients electronically enables care delivery from remote diagnostics to remote surgery. Traditional visual computing tasks such as MRI, volume rendering, computer vision or image processing may also be available to more clinics and researchers as they become 'electronically local.' Video is the component of multimedia that provides the greatest sense of presence or visual realism yet has been the most difficult to offer digitally due to its high transmission, storage and computation requirements. Advanced 3D graphics have also been a scarce or at least expensive resource. This paper addresses some of the recent innovations in media processing and client/server technology that will facilitate PCs, workstations or even set-top/TV boxes to process both video and graphics in real-time.

The suitability of computers to the task of medical image visualization for the purposes of primary diagnosis and treatment planning depends on three factors: speed, image quality, and price. To be widely accepted the technology must increase the efficiency of the diagnostic and planning processes. This requires processing and displaying medical images of various modalities in real-time, with accuracy and clarity, on an affordable system. Our approach to meeting this challenge began with market research to understand customer image processing needs. These needs were translated into system-level requirements, which in turn were used to determine which image processing functions should be implemented in hardware. The result is a computer architecture for 2D image processing that is both high-speed and cost-effective. The architectural solution is based on the high-performance PA-RISC workstation with an HCRX graphics accelerator. The image processing enhancements are incorporated into the image visualization accelerator (IVX) which attaches to the HCRX graphics subsystem. The IVX includes a custom VLSI chip which has a programmable convolver, a window/level mapper, and an interpolator supporting nearest-neighbor, bi-linear, and bi-cubic modes. This combination of features can be used to enable simultaneous convolution, pan, zoom, rotate, and window/level control into 1 k by 1 k by 16-bit medical images at 40 frames/second.

Tremendous computational capabilities are required in modern ultrasound machines. Unique and very high data throughput rates combined with the demanding processing requirement have restricted the designs of such ultrasound machines to algorithm-specific hardware with limited programmability. Specialized electronic boards are dedicated to each of the several subsystems such as gray scale, color flow, and Doppler processing. In many cases, improving the functionality of an ultrasound machine requires hardware redesigns and replacements of boards or of the entire machine. These redesigns and replacements generally require significant expenditures in manpower, time, and cost, thus imposing stringent limits on the types of image processing which can be supported. In an effort to address these problems, we have architected and designed a high-performance programmable ultrasound processing subsystem, the UWGSP8, to fit within an existing ultrasound machine and support native ultrasound image processing. Its flexible design means that it can support a much wider range of ultrasound image processing algorithms than can be realized in any conventional fixed hardware design.

The World Wide Web (WWW) is becoming the predominant force for global information dissemination. The Web browsers are originally designed for the client computers to navigate and display hypermedia documents and image bitmaps stored at the server machines. Their capabilities must be extended with database query and interactive visualization before the Web could be useful for the medical imaging community. This paper presents the system design and tools for interactive query and visualization of medical images on the Web. Examples from breast and brain imaging applications are used to illustrate the operations and capabilities of such tools.

Current training for regional nerve block procedures by anesthesiology residents requires expert supervision and the use of cadavers; both of which are relatively expensive commodities in today's cost-conscious medical environment. We are developing methods to augment and eventually replace these training procedures with real-time and realistic computer visualizations and manipulations of the anatomical structures involved in anesthesiology procedures, such as nerve plexus injections (e.g., celiac blocks). The initial work is focused on visualizations: both static images and rotational renderings. From the initial results, a coherent paradigm for virtual patient and scene representation will be developed.

The classic radiologic teaching file consists of hundreds, if not thousands, of films of various ages, housed in paper jackets with brief descriptions written on the jackets. The development of a good teaching file has been both time consuming and voluminous. Also, any radiograph to be copied was unavailable during the reproduction interval, inconveniencing other medical professionals needing to view the images at that time. These factors hinder motivation to copy films of interest. If a busy radiologist already has an adequate example of a radiological manifestation, it is unlikely that he or she will exert the effort to make a copy of another similar image even if a better example comes along. Digitized radiographs stored on CD-ROM offer marked improvement over the copied film teaching files. Our institution has several laser digitizers which are used to rapidly scan radiographs and produce high quality digital images which can then be converted into standard microcomputer (IBM, Mac, etc.) image format. These images can be stored on floppy disks, hard drives, rewritable optical disks, recordable CD-ROM disks, or removable cartridge media. Most hospital computer information systems include radiology reports in their database. We demonstrate that the reports for the images included in the users teaching file can be copied and stored on the same storage media as the images. The radiographic or sonographic image and the corresponding dictated report can then be 'linked' together. The description of the finding or findings of interest on the digitized image is thus electronically tethered to the image. This obviates the need to write much additional detail concerning the radiograph, saving time. In addition, the text on this disk can be indexed such that all files with user specified features can be instantly retrieve and combined in a single report, if desired. With the use of newer image compression techniques, hundreds of cases may be stored on a single CD-ROM depending on the quality of image required for the finding in question. This reduces the weight of a teaching file from that of a baby elephant to that of a single CD-ROM disc. Thus, with this method of teaching file preparation and storage the following advantages are realized: (1) Technically easier and less time consuming image reproduction. (2) Considerably less unwieldy and substantially more portable teaching files. (3) Novel ability to index files and then retrieve specific cases of choice based on descriptive text.

The well-known hospital information system (HIS) and the picture archiving and communication system (PACS) are typical applications of multimedia to medical area. This paper proposes a personal medical information save-and-carry system using a laser card. This laser card is very useful, especially in emergency situations, because the medical information in the laser card can be read at anytime and anywhere if there exists a laser card reader/writer. The contents of the laser card include the clinical histories of a patient such as clinical chart, exam result, diagnostic reports, images, and so on. The purpose of this system is not a primary diagnosis, but emergency reference of clinical history of the patient. This personal medical information system consists of a personal computer integrated with laser card reader/writer, color frame grabber, color CCD camera and a high resolution image scanner optionally. Window-based graphical user interface was designed for easy use. The laser card has relatively sufficient capacity to store the personal medical information, and has fast access speed to restore and load the data with a portable size as compact as a credit card. Database items of laser card provide the doctors with medical data such as laser card information, patient information, clinical information, and diagnostic result information.

With the goal of eventually increasing the quality of medical care, especially in remote areas, we have developed a system for telemedicine research based on a combination of ATM networking and a high-speed DSP board based on the Texas Instruments TMS320C80. The purpose of the system is to give health care providers at remote locations the ability to consult with specialists using a combination of video, audio, and externally-acquired images. The system can also be used for education purposes to support bi-directional video/audio communications for grand round lectures, classes, and case conferences. In order to maximize the utilization of the available transmission medium (ranging from land-based copper and fiber optic cable to satellite link) while providing the best possible video and audio quality, the compression performed by the system is adaptable to a wide variety of bandwidths. After about two years of experience with telemedicine in a research environment, we have some preliminary findings to report regarding the performance of a telemedicine application combining ATM and programmable multimedia processors in PC environments.

New and modern image modalities generate digital images and increase the amount of data in a hospital. Without data compression techniques the practical use of many applications like teleradiology is not usable. A typical CT-series is about 25-50 MB. Many non-standardized lossy data compression techniques have been developed for the clinical environment. This paper describes qualitative and quantitative evaluation of three standardized compression methods, JPEG, MPEG, and H.261 for clinical cases. The evaluation shows that JPEG and MPEG can actually still compete with special non-standardized medical image compression methods and the compressed images can be decompressed with very little effort.

We have designed an image compression scheme based on multiresolution decomposition using the quadrature mirror filters (QMFs), and applied it to chest radiographs. In order to assess the preservation of diagnostic image quality, we performed image review study using a diagnostic workstation. From the result of review by diagnostic radiologists, our method may be available in clinical practice.

This paper proposes a novel on-line structure lossless compression method for digital mammograms during the film digitization process. The structure-lossless compression segments the breast and the background, compresses the former with a predictive lossless coding method and discards the latter. This compression scheme is carried out during the film digitization process and no additional time is required for the compression. Digital mammograms are compressed on-the-fly while they are created. During digitization, lines of scanned data are first acquired into a small temporary buffer in the scanner, then they are transferred to a large image buffer in an acquisition computer which is connected to the scanner. The compression process, running concurrently with the digitization process in the acquisition computer, constantly checks the image buffer and compresses any newly arrived data. Since compression is faster than digitization, data compression is completed as soon as digitization is finished. On-line compression during digitization does not increase overall digitizing time. Additionally, it reduces the mammogram image size by a factor of 3 to 9 with no loss of information. This algorithm has been implemented in a film digitizer. Statistics were obtained based on digitizing 46 mammograms at four sampling distances from 50 to 200 microns.

Functional magnetic resonance imaging (fMRI) experiments are becoming recognized in a number of areas of neuroscience. Presenting useful information to the clinician in a reasonable time and in an understandable way is of paramount importance for the use of fMRI protocols in the clinical setting. We have developed a series of tools for fMRI analysis and presentation encapsulated by a commercially-available graphical-user interface which allows the user to immediately make use of fMRI data for multiple analyses. The application visualization system (AVS) was chosen to provide a graphical environment for the tools. A series of AVS modules were created to allow the user to perform several processing and analysis tasks using the serial fMRI image data as a starting point. Modules were developed to provide t-test analysis and cross-correlation analysis, in which the user is able to select a suitable idealized driving function which can be interactively modified to suit the given fMRI protocol. Both time and frequency analyses are possible on each pixel in the images. In addition, several co- registration methods have been developed to resolve problems arising from patient motion.

Although OpenGL is not usually thought of as a library for imaging, it was designed to expose the capabilities of modern frame buffer hardware. The emphasis in OpenGL is on 3D graphics (i.e., geometry), but OpenGL also includes a fairly rich set of capabilities for 2D imaging. This paper describes the capabilities of OpenGL for imaging applications, including pixel transfer operations (draw, read, copy); color lookup tables; linear transformation of color values; pixel conversion capabilities; and pixel operations such as blending, masking, and clipping. Several recently proposed extensions to OpenGL add significant capabilities to the core imaging model, including convolution, window level mapping, and image transformation and resampling. These capabilities are discussed in the context of the pixel processing pipeline defined by OpenGL.

Meaningful performance evaluation of video equipment can be complex, requiring specialized equipment in which results must be interpreted by technically trained operators. The alternative to this has been to attempt evaluation by visual inspection of patterns such as the SMPTE RP-133 Medical Imaging Standard. However this involves subjective interpretation and does not indicate the point in a system at which degradation has occurred. The video waveform of such a pattern is complex and not suitable for quantitative analysis. The principal factors which influence quality of a video image on a day-to-day basis are resolution, gray scale and color, if employed. If these qualities are transmitted and displayed without degradation beyond acceptable limits, suitable performance is assured. Performance evaluation by inspection of the image produced on a video display monitor is subject to interpretation; this is resolved by inserting, at the display, the original 'perfect' electronically generated waveform to serve as a reference. Thus the viewer has a specific visual comparison as the basis for performance evaluation. Another valuable feature of the test pattern insert is that a test segment can be placed on recorded images. Thus each image recalled by tape playback or from digital storage will carry an integral means for quality assurance.

The production of hardcopy output, of suitable quality for modern digital radiography systems, is generally performed using a laser imager exposing a silver-halide film. Until recently the use of technologies that do not require the 'wet' photographic processing of the film have been hampered by poor image quality. However the Polaroid Helios imager offers high imaging specification, without the environmental overheads, by utilizing a totally digital printing paradigm. The Helios system was evaluated in terms of its physical imaging properties, and a clinical trial was undertaken. The results, although preliminary in nature, indicate that the Helios imager does indeed offer a viable alternative to conventional laser imagers. This encourages us to carry out a full trial in clinical routine.

This investigation is a preliminary attempt to derive a figure-of-merit (FOM) characterizing the image degradation that can occur as a result of image compression. The usefulness of a visible difference predictor (VDP), which is based on previously published psychophysical models, is assessed. Specifically, this is a preliminary attempt to relate physical differences between an original image and a compressed version of the original to visible differences and then to calculate a FOM indicating the degree to which images have been degraded based on this visible difference. The FOM was applied to phantom images and to CT images that had been compressed at various ratios using either the JPEG algorithm or a wavelet compression algorithm. Comparisons between the VDP based measures and more traditional measures of image degradation suggest that, while this FOM may overcome some of the shortcomings of simpler measures such as root-mean-square (RMSE), the use of a VDP has limitations of its own.

Many medical applications require the full perception of human organs or tissues for advanced interpretation and reliable decision. It is useful to generate a three-dimensional (3-D) view from its serial cross sections for surgical planning and diagnosis. The cross-sectional images are usually represented by contours after segmentation. Interpolation has to be carried out to fill the space between the successive contours. A new approach to 3-D image interpolation using the co-matching corresponding finding (CMCF) is proposed. The start and goal contours are mapped onto a unit square respectively and then divided into four regions with each side of the square in order to determine the four bounding points. Four segments are formed between the bounding points. Hence, the matching process becomes the matching of a segment to another (segment of another contour) and it is repeated four times. An objective mapping is applied to the correspondence points of each segment and additional points which follow a precise decision rule may be inserted for determining the best correspondence pair.

The development of microscopes is nearly fixed since the introduction of confocal microscopy. The credit-point of confocal microscopy is the low depth of focus. But a small depth of focus is very often a problem in conventional light microscopy, because the objects are higher than the depth of focus. There is a physical limitation based on the numerical aperture (NA) of the objective lens of the microscope. The connection of the conventional light microscope, video technique and computers open new fields of applications for the microscopy. It is possible to generate an image containing focused areas from series of images of different focus- or z- positions. In this way it is possible to extend the depth of focus without a physical limitation by the NA of the objective lens. The connection of microscope, video and PC also opens the field of high resolution low light color applications, like fluorescence in situ hybridization (FISH). Image processing technology enables the enhancement of contrast for objects with a very low contrast. And such a system also opens the way for a view to objects with small motion, like processes of growing of biological objects.

Medical imaging companies have traditionally supplied the industry with image visualization solutions based on their own custom hardware designs. Today, more and more systems are being deployed using only off-the-shelf workstations. Two major factors are driving this change. First, workstations are delivering the functionality and performance required to replace custom hardware for an ever increasing subset of visualization techniques, while continuing to come down in cost. Second, cost pressures are forcing medical imaging companies to OEM the hardware platform and focus on what they do best -- delivering solutions to health care providers. This industry shift is challenging the workstation vendors to deliver the maximum inherent performance in their computer systems to medical imaging applications without locking the application into a specific vendor's hardware. Since extracting the maximum performance from a workstation is not always intuitively obvious and often requires vendor-specific tricks, the best way to deliver performance to an application is through an application programmer's interface (API). The Hewlett-Packard Image Visualization Library (HP-IVL) is such an API. It transparently delivers the maximum possible imaging performance on Hewlett-Packard workstations, while allowing significant portability between platforms. This paper describes the performance tricks and trade-offs made in the software implementation of HP's Image Visualization Library and how the HP Image Visualization Accelerator (HP-IVX) fits into the overall architecture.

In this paper, lossy compression of 3D MR images of the human brain is associated with a segmentation algorithm, in the context of an interactive brain sulci delineation application. Influence of compression losses is analyzed according to the segmentation results. Lossy compression is performed by subband coding leading to a multiresolution representation of the image. Wavelets are adapted for medical images statistics. The decompressed images are segmented by directional watershed transform (DWST), providing an accurate 3D segmentation of the brain. Impact of losses on the quality of the segmentation is estimated either by a 3D Chamfer distance function and by visual appreciation. In this article, we show that lossy compression can be combined with some applications, providing high compression ratio without significantly altering the results of the application.

Perceptual linearization has been advocated for medical image presentation, both for the faithful reproduction of images, and for standardizing the appearance across different display devices. It is currently being proposed as the standard display function for medical image presentation by ACR/NEMA working group 11 (display function standard). At this time, studies have not been made to evaluate how close existing display systems are to being perceptually linearized. This paper presents a methodology for quantitatively calculating the perceptual linearity of a display device based on a statistical measure, the linearization uniformity measure (LUM), of standard deviation of the ratio of contrast thresholds of the display system versus the contrast thresholds of the human observer. Currently available medical image display systems are analyzed using LUM metric, and their pre-linearization and post-linearization results are compared with that of the desired human observer response curve. We also provide a better description of the achievable dynamic range of a display device, based on the three quantitative measures: the standard deviation of the contrast threshold ratios, the mean of the contrast threshold ratios, and the number of DDL steps used.

In the field of medical imaging there is a need for high-resolution high-brilliance monochromatic CRT displays. However, at higher brightness levels the resolution of these displays decreases, due to the increasing spot size. In order to improve the performance of the CRT display a relatively simple method, called the multi-beam concept, is introduced. Using this technique a higher brightness can be realized without an increase of the spot size and therefore a better display quality can be achieved. However, for successful exploitation of the multi-beam concept it is necessary to minimize the convergence error of the CRT display. For this purpose two circuits have been realized, where the convergence error is reduced by an analogue and a digital method respectively. The analogue implementation means an improvement for the applicability of the multi-beam concept, however, in order to achieve major image quality improvement the use of the digital system is necessary.

In this paper, we present a new approach to the generation and manipulation of oblique slices of MR/CT or any voxel space images. We consider the result of the voxel space as if it would be cut by a scalpel, simulating the action of a surgeon. With the intervention of a new definition of discrete 3D voxel plane, we show that this can be done in a highly efficient way and that there are very few computations to do. It is shown that the supercover of the continuous oblique plane is a 6-connected discrete plane that has many very interesting properties that can be usefully exploited. For instance, geometrical considerations avoid much of the intersection computations and once one oblique slice is generated, all the other parallel oblique slices can be generated with few further computations. These results are applied to improve the existing algorithms and provide some ideas for new ones. It should also improve the handling possibilities of oblique slices, indeed, almost all that can be done on a sagittal or axial slice can then also be done on oblique slices. An extension to 4D oblique slices is possible.

The diagnosis of breast disease for screening or symptomatic women is largely arrived at by a multi-disciplinary team. We report work on the development and assessment of an inter- disciplinary computer based learning system to support the diagnosis of this disease. The diagnostic process is first modelled from different viewpoints and then appropriate knowledge structures pertinent to the domains of radiologist, pathologist and surgeon are depicted. Initially the underlying inter-relationships of the mammographic diagnostic approach were detailed which is largely considered here. Ultimately a system is envisaged which will link these specialties and act as a diagnostic aid as well as a multi-media educational system.

In this paper, we investigate the application of fractal concept to the coding of medical images, taking into account the self-similarity at different scales. The approach proposed by Jacquin is very flexible, enabling us to optimize different steps of the associated algorithm. We remark that the choice of the distance used to measure the self-similarity between the range block R_i and the estimated range block R_i is essential to the algorithm. In fact, the choice of the metric determines the optimal parameters of the affine transformation. In our study, we propose two metrics, L₂ and L_(infinity ) and compare their performances. We also develop a simple fast decoding scheme, necessary for a clinical use. This paper addresses the adaptation of the fractal compression algorithm to medical image modalities. We present the results obtained with two image data bases (numerized mammograms and x-ray angiograms). A comparison with JPEG results shows the improvement with our technique, particularly for low bit rates.

A design-based approach to ethical analysis examines how computer scientists, physicians and patients make and justify choices in designing, using and reacting to computer-aided diagnosis (CADx) systems. The basic hypothesis of this research is that values are embedded in CADx systems during all phases of their development, not just retrospectively imposed on them. This paper concentrates on the work of computer scientists and physicians as they attempt to resolve central technical questions in designing clinically functional CADx systems for lung cancer and breast cancer diagnosis. The work of Lo, Chan, Freedman, Lin, Wu and their colleagues provides the initial data on which this study is based. As these researchers seek to increase the rate of true positive classifications of detected abnormalities in chest radiographs and mammograms, they explore dimensions of the fundamental ethical principal of beneficence. The training of CADx systems demonstrates the key ethical dilemmas inherent in their current design.

In this paper, we compute quantization matrices which are tuned to individual images for JPEG compression of 12 bit radiographs. The quantization matrices were derived using a perceptual model, and tested in a set of psychophysical experiments to find the just-noticeable- difference (jnd) quality level between the original and lossy compressed images. Each of the images used in the study was compressed to 17 different quality levels, where 'quality level' is a perceptually meaningful metric. The results show that the technique can be used to provide selectable quality image compression of radiographs. We provide a list of recommended quantization matrices at various quality levels for images of this class (peripheral bone radiographs).

Despite the proven superiority of vector quantization (VQ) over scalar quantization (SQ) in terms of rate distortion theory, currently existing vector quantization algorithms, still, suffer from several practical drawbacks, such as codebook initialization, long search-process, and optimization of the distortion measure. We present a new adaptive vector quantization algorithm that uses a fuzzy distortion measure to find a globally optimum codebook. The generation of codebooks is facilitated by a self-organizing neural network-based clustering that eliminates adhoc assignment of the codebook size as required by standard statistical clustering. In addition, a multiresolution wavelet decomposition of the original image enhances the process of codebook generation. Preliminary results using standard monochrome images demonstrate excellent convergence of the algorithm, significant bit rate reduction, and yield reconstructed images with high visual quality and good PSNR and MSE. Extension of this adaptive VQ to color image compression is currently under investigation.

Teleradiology is being implemented in the U.S. military. Soft-copy reading of computed radiology (CR) images and film-digitizer (FD) images are becoming a common practice. The Medical Diagnostic Imaging Support (MDIS) Office at the Medical Advanced Technology Management Office (MATMO), Fort Detrick, Maryland, installed an 'off-the-shelf' DICOM teleradiology system by which CR images and FD images acquired at Hickam Air Force Base (AFB), Hawaii, are transmitted electronically over a T-1 telecommunications line to Tripler Army Medical Center (TAMC), Hawaii. The goal was to provide a diagnostic quality teleradiology system to the military services to extend the expertise and training of physicians to remote sites. In order to guarantee a diagnostic quality image throughout the system, a rigid set of quality control standards had to be designed and implemented. This poster presents the results of a successful teleradiology implementation where quality control is maintained throughout the imaging chain.

The implementation of teleradiology is spreading throughout the U.S. military. In an effort to contain cost, the Medical Diagnostic Imaging Support (MDIS) Office at the Medical Advanced Technology Management Office (MATMO), Fort Detrick, Maryland, implemented an 'off- the-shelf' DICOM teleradiology system by which computed radiography (CR) images acquired at Hickam Air Force Base (AFB), Hawaii, can be transmitted electronically over a T-1 telecommunications line to Tripler Army Medical Center (TAMC), Hawaii. The goal was to provide a teleradiology system to the military services which extends the expertise and training of physicians to remote sites, while realizing cost savings through off-the-shelf DICOM technology. The TAMC teleradiology hub equipment was provided to support soft copy diagnostic reading. The Army X-Ray ISO-Shelter along with the teleradiology equipment allows Hickam AFB to maintain radiology services during an on-going construction project. This poster presents the results of a successful off-the-shelf DICOM teleradiology implementation.

The Department of Defense Telemedicine Test Bed produced a CD-ROM including information on telemedicine, teleradiology and military medical advanced technology projects. The CD-ROM was produced using media from the Telemedicine Test Bed World Wide Web site and academic papers and presentations. Apple Media Tools software was used to produce the interactive program and the authoring was done on a high speed Apple Macintosh Power PC computer. The process took roughly 100 hours to author 50 Mb of data into 200 frames of interactive material. Future versions of the Telemedicine CD-ROM are in progress which will include much more material to take advantage of the 650 Mb available on a compact disk. This paper graphically depicts and explains the authoring process.

In September 1995 the United States military conducted a demonstration project to provide live ultrasound video and diagnostic DICOM still images using GTE's asynchronous transfer mode (ATM) technologies over an Orion T-1 satellite link. Still images were frame-grabbed from a Diasonics ultrasound and sent to the ALI Wide Area Network system. A group of diagnostic images was then sent in DICOM 3.0 format over a virtual ethernet satellite link from Chantilly, Virginia to Dayton, Ohio. These images came across a DICOM gateway into the Medical Diagnostic Imaging Support (MDIS) System. Live video from the ultrasound was also routed through a CLI Radiance VTC over the satellite to a VTC in Ohio. The video bandwidth was progressively narrowed with two radiologists determining the minimal acceptable bandwidth for detecting test objects in a phantom. The radiologists accepted live video ultrasound at bandwidths as low as 384 kbps from the hands of an experienced ultrasonographer located hundreds of miles away. DICOM still images were sent uncompressed and were of acceptable image quality when viewed on the MDIS system. The technology demonstrated holds great promise for both deployed U.S. Military Forces and civil uses of remote radiology. Detailed network drawings and videotapes of the ultrasound examinations at the remote site are provided.