This PDF file contains the editorial “Guest Editorial: Special Section on Biomedical Image Representation” for JBO Vol. 8 Issue 01

This contribution discusses a selection of today’s techniques and future concepts for digital x-ray imaging in medicine. Advantages of digital imaging over conventional analog methods include the possibility to archive and transmit images in digital information systems as well as to digitally process pictures before display, for example, to enhance low contrast details. After reviewing two digital x-ray radiography systems for the capture of still x-ray images, we examine the real time acquisition of dynamic x-ray images (x-ray fluoroscopy). Here, particular attention is paid to the implications of introducing charge-coupled device cameras. We then present a new unified radiography/fluoroscopy solid-state detector concept. As digital image quality is predominantly determined by the relation of signal and noise, aspects of signal transfer, noise, and noise-related quality measures like detective quantum efficiency feature prominently in our discussions. Finally, we describe a digital image processing algorithm for the reduction of noise in images acquired with low x-ray dose.

An adaptive vector quantizer (VQ) using a clustering technique known as adaptive fuzzy leader clustering (AFLC) that is similar in concept to deterministic annealing (DA) for VQ codebook design has been developed. This vector quantizer, AFLC-VQ, has been designed to vector quantize wavelet decomposed sub images with optimal bit allocation. The high-resolution sub images at each level have been statistically analyzed to conform to generalized Gaussian probability distributions by selecting the optimal number of filter taps. The adaptive characteristics of AFLC-VQ result from AFLC, an algorithm that uses self-organizing neural networks with fuzzy membership values of the input samples for upgrading the cluster centroids based on well known optimization criteria. By generating codebooks containing codewords of varying bits, AFLC-VQ is capable of compressing large color/monochrome medical images at extremely low bit rates (0.1 bpp and less) and yet yielding high fidelity reconstructed images. The quality of the reconstructed images formed by AFLC-VQ has been compared with JPEG and EZW, the standard and the well known wavelet based compression technique (using scalar quantization), respectively, in terms of statistical performance criteria as well as visual perception. AFLC-VQ exhibits much better performance than the above techniques. JPEG and EZW were chosen as comparative benchmarks since these have been used in radiographic image compression. The superior performance of AFLC-VQ over LBG-VQ has been reported in earlier papers.

Accurate estimation of bronchial caliber in high resolution computerized tomography (HRCT) is essential for physicians in the management and following of patients with airway disease. Although there are at present different methods of bronchi analysis, none of them can provide an absolute diagnosis of bronchial caliber. The present paper addresses a new approach of bronchi segmentation in HRCT in order to estimate the bronchial caliber. The method developed is based on mathematical morphology theory, and relies on morphological filtering, marking techniques derived from the concept of connection cost, and conditional watershed-based segmentation. In order to evaluate the robustness of the segmentation and the accuracy of the caliber estimates, a realistic bronchi modeling based on physiological characteristics has been developed. It is shown that, according to the size of the bronchi, the estimation accuracy is up to 90%.

The purposes of this research are to investigate the effectiveness of our novel lung contour detection method in chest radiographs. The proposed method consists of five sections as follows. First, in order to reduce the amount of information, the images are smoothed and subsampled from 2 k by 2.5 k pixels to 256 by 310 pixels with rescaling from 12-bit to 8-bit based on the image maximum and minimum. Second, that the image is resolved into the profiles for two directions (i.e., horizontal x and vertical y axes). Then, for each direction, those profiles are tested by the neural network which has been trained using the profiles from 14 pairs of original and target images. Note that both horizontal and vertical neural networks are trained with the horizontal and vertical profiles, respectively. For each direction, the whole two-dimensional image is reconstructed from the output profiles of the neural network. Next, the binarization process followed by the labeling process is applied to each reconstructed image individually. Finally, two postprocessed images are combined through the OR operation, and the labeling process is performed for the combined image to get the final contour. A total of 85 screening chest radiographs from Johns Hopkins University Hospital were digitized to 2 k by 2.5 k pixels with 12-bit gray scale. Fourteen images were used for the training of the neural networks and the remaining 71 images for testing. The proposed method can detect the lung contour at 94% accuracy for test images following the same rules as for the training images.

This paper provides examples of several medical image analysis applications for which single-purpose border detection approaches were developed in the past. However, the utility of these and other existing automated and semiautomated medical image analysis systems is limited by their narrow, frequently singlepurpose orientation. After a general approach to graph-based optimal border detection is overviewed, a new method for design of image segmentation systems is reported, in which the criterion of optimality is automatically determined by learning from border tracing examples. Border features employed in the designed method are selected from a predefined global set using radial-basis neural networks. The method was validated in intracardiac, intravascular, and ovarian ultrasound images. The achieved performance was comparable to that of our previously reported single-purpose border detection methods. Our approach facilitates development of general multi-purpose image segmentation systems that can be trained for different types of image segmentation applications.

Segmentation via morphological granulometric features is based on fitting structuring elements into image topography from below and above. Each structuring element captures a specific texture content. This paper applies granulometric segmentation to digitized mammograms in an unsupervised framework. Granulometries based on a number of flat and nonflat structuring elements are computed, local size distributions are tabulated at each pixel, granulometric-moment features are derived from these size distributions to produce a feature vector at each pixel, the Karhunen–Loeve transform is applied for feature reduction, and Voronoi-based clustering is performed on the reduced Karhunen–Loeve feature set. Various algorithmic choices are considered, including window size and shape, number of clusters, and type of structuring elements. The algorithm is applied using only granulometric texture features, using gray-scale intensity along with the texture features, and on a compressed mammogram. Segmentation results are clinically evaluated to determine the algorithm structure that best accords to an expert radiologist’s view of a set of mammograms.

The existence of clustered microcalcifications is one of the important early signs of breast cancer. This paper presents an image processing procedure for the automatic detection of clustered microcalcifications in digitized mammograms. In particular, a sensitivity range of around one false positive per image is targeted. The proposed method consists of two main steps. First, possible microcalcification pixels in the mammograms are segmented out using wavelet features or both wavelet features and gray level statistical features, and labeled into potential individual microcalcification objects by their spatial connectivity. Second, individual microcalcifications are detected by using the structure features extracted from the potential microcalcification objects. The classifiers used in these two steps are feedforward neutral networks. The method is applied to a database of 40 mammograms (Nijmegen database) containing 105 clusters of microcalcifications. A free response operating characteristics curve is used to evaluate the performance. Results show that the proposed procedure gives quite satisfactory detection performance. In particular, a 93% mean true positive detection rate is achieved at the price of one false positive per image when both wavelet features and gray level statistical features are used in the first step.

Modeling of diagnostic examinations reduces the sample size needed to acquire statistical power. This paper compares the models of receiver operating characteristic analysis, channel models of diagnostic examinations, and uncertainty measures in modeling mammography examination. The cited models are applied to screening mammography. The results of the use of these models are less data required for achieving chosen levels of statistical power. These models are able to manage the large number of variables involved in comparing diagnostic examinations.

This paper describes a system for surface recovery and visualization of the three-dimensional topography of the optic nerve head, as support of early diagnosis and follow up of glaucoma. In stereo vision, depth information is obtained from triangulation of corresponding points in a pair of stereo images. In this paper, the use of the cepstrum transformation as a disparity measurement technique between corresponding windows of different block sizes is described. This measurement process is embedded within a coarseto- fine depth-from-stereo algorithm, providing an initial range map with the depth information encoded as gray levels. These sparse depth data are processed through a cubic B-spline interpolation technique in order to obtain a smoother representation. This methodology is being especially refined to be used with medical images for clinical evaluation of some eye diseases such as open angle glaucoma, and is currently under testing for clinical evaluation and analysis of reproducibility and accuracy.

In fractal image compression an image is partitioned into ranges for each of which a similar subimage, called domain, is selected from a pool of subimages. However, only a fraction of this large pool is actually used in the fractal code. This subset can be characterized in two related ways: (1) It contains domains with relatively large intensity variation. (2) The collection of used domains is localized in those image regions that show a high degree of structure. Both observations lead to improvements of fractal image compression. First, we accelerate the encoding process by a priori discarding the low variance domains from the pool that are unlikely to be chosen for the fractal code. Second, the localization of the domains may be exploited for an improved encoding of the domain indices, in effect raising the compression ratio. When considering the performance of a variable rate fractal quadtree encoder we found that a speedup by a factor of 2–3 does not degrade the ratedistortion curve ranging from compression ratio 5 up to 30.

Dithering quality of the void and cluster algorithm suffers due to fixed filter width and absence of a well-defined criterion for selecting among equally likely candidates during the computation of the locations of the tightest clusters and largest voids. Various researchers have addressed the issue of fixed filter width by adaptively changing the width with experimentally determined values. This paper addresses both aforementioned issues by using a Voronoi tessellation and three criteria to select among equally likely candidates. The algorithm uses vertices of the Voronoi tessellation, and the areas of the Voronoi regions to determine the locations of the largest voids and the tightest clusters. During void and cluster operations there may be multiple equally likely candidates for the locations of the largest voids and the tightest clusters. The selection among equally likely candidates is important when the number of candidates is larger than the number of dots for a given quantization level, or if there are candidates within the local neighborhood of one of the candidate points, or if a candidate’s Voronoi region shares one or more vertices with another candidate’s Voronoi region. Use of these methods leads to more uniform dot patterns for light and dark tones. The improved algorithm is compared with other dithering methods based on power spectrum characteristics and visual evaluation.

This paper presents predicted performance for twodimensional cross correlation where two images taken from a planar object model differ by a general perspective geometric transformation. The study shows there exists a window size that will maximize or minimize certain performance parameters for a given perspective distortion. The analysis also indicates many performance criteria have an optimum window size if the geometric distortion includes rotation or scale, but for a given perspective distortion where pitch angle is the only parameter, these measures are not appreciably optimized by any given correlation window size. The performance measures examined are expected peak value, peak-to-sidelobe ratio, probability of correct acquisition (PCA) and false acquisition (PFA), registration error covariance, and average signal-to-noise ratio. The results use statistically consistent image models with arbitrary autocorrelation functions. Monte Carlo simulation verification of theoretical predictions is performed and results are extended to a variety of common area-based image matching techniques.

Many people naively believe that the CIE standard observer defined in 1931, and the CIELAB color space defined in 1976, comprise ‘‘all I ever really needed to know about color science.’’ However, it can easily be shown that the world of human color perception is much more complex than this in reality.