Roger S. Gaborski

Multiresolution unsharp masking technique for mammogram image enhancement

A multi-resolution unsharp masking (USM) technique is developed for image feature enhancement in digital mammogram images. This technique includes four processing phases: (1) determination of parameters of multi-resolution analysis (MRA) based on the properties of images; (2) multi-resolution decomposition of original images into sub-band images via wavelet transformation with perfect reconstruction filters; (3) modification of sub-band images with adaptive unsharp masking technique; and (4) reconstruction of image from modified sub- band images via inverse wavelet transformation. An adaptive unsharp masking technique is applied to the sub-band images in order to modify the pixel values based on the edge components at various frequency scales. Smoothing and gain factor parameters, employed in the unsharp masking, are determined according to the resolution, frequency, and energy content of the sub-band images. Experimental results show that this technique is able to enhance the contrast of region of interest (microcalcification clusters) in mammogram image.

Comparative study of wavelet and discrete cosine transform (DCT) decompositions with equivalent quantization and encoding strategies for medical images

Wavelet-based image compression is receiving significant attention, largely because of its potential for good image quality at low bit rates. In medical applications, low bit rate coding may not be the primary concern, and it is not obvious that wavelet techniques are significantly superior to more established techniques at higher quality levels. In this work we present a straightforward comparison between a wavelet decomposition and the well-known discrete cosine transform decomposition (as used in the JPEG compression standard), using comparable quantization and encoding strategies to isolate fundamental differences between the two methods. Our focus is on the compression of single-frame, monochrome images taken from several common modalities (chest and bone x-rays and mammograms).

Bayesian and neural network pattern recognition: a theoretical connection and empirical results with handwritten characters

Abstract Statistical Pattern Recognition and Artificial Neural Networks provide alternative methodologies to the classification of patterns represented as feature vectors. This paper provides a theoretical relationship and an empirical comparison between the Bayes decision rule and the backpropagation model. It is shown that backpropagation performs least mean square approximation to the Bayes discriminant function. While a three-layer backpropagation network (one hidden layer) with a sufficient number of hidden units is known to possess universal mapping ability, gradient-descent based backpropagation learning does not guarantee finding the minimum probability of error solution. Experimental results with handwritten character recognition (digits and letters extracted from handwritten addresses) are presented. The experiments are with two different representations of characters: binary pixel arrays and structural features represented as binary vectors. With pixel arrays, the backpropagation model performs better than the first-order Bayes discriminant that assumes statistical independence between pixels. With structural features, the first-order Bayes and backpropagation have similar performance. However, training of a backpropagation network is much more involved. Inherent difficulties with both classifiers are discussed.

The Polynomial Method Augmented by Supervised Training for Hand-Printed Character Recognition

We present a pattern recognition algorithm for handprinted and machine-printed characters, based on a combination of the classical least squares method and a neural-network-type supervised training algorithm. Characters are mapped, nonlinearly, to feature vectors using selected quadratic polynomials of the given pixels. We use a method for extracting an equidistributed subsample of all possible quadratic features.

Genetic Algorithm Selection of Features for Hand-printed Character Identification

We have constructed a linear discriminator for handprinted character recognition that uses a (binary) vector of 1, 500 features based on an equidistributed collection of products of pixel pairs. This classifier is competitive with other techniques, but faster to train and to run for classification.

Application of neural-network-based multistage system for detection of microcalcification clusters in mammogram images

A multi-stage system with image processing and artificial neural techniques is developed for detection of microcalcification in digital mammogram images. The system consists of (1) preprocessing stage employing box-rim filtering and global thresholding to enhance object-to- background contrast; (2) preliminary selection stage involving body-part identification, morphological erosion, connected component analysis, and suspect region segmentation to select potential microcalcification candidates; and (3) neural network-based pattern classification stage including feature map extraction, pattern recognition neural network processing, and decision-making neural network architecture for accurate determination of true and false positive microcalcification clusters. Microcalcification suspects are captured and stored in 32 by 32 image blocks, after the first two processing stages. A set of radially sampled pixel values is utilized as the feature map to train the neural nets in order to avoid lengthy training time as well as insufficient representation. The first pattern recognition network is trained to recognize true microcalcification and four categories of false positive regions whereas the second decision network is developed to reduce the detection of false positives, hence to increase the detection accuracy. Experimental results show that this system is able to identify true cluster at an accuracy of 93% with 2.9 false positive microcalcifications per image.

Contrast-detail analysis of the effect of image compression on computed tomographic images

Three compression algorithms were compared by using contrast-detail (CD) analysis. Two phantoms were designed to simulate computed tomography (CT) scans of the head. The first was based on CT scans of a plastic cylinder containing water. The second was formed by combining a CT scan of a head with a scan of the water phantom. The soft tissue of the brainwas replaced by a subimage containing only water. The compression algorithms studied were the full-frame discrete cosine (FDCT) algorithm, the Joint Photographic Experts Group (JPEG) algorithm, and a wavelet algorithm. Both the wavelet and JPEG algorithms affected regions of the image near the boundary of the skull. The FDCT algorithm propagated false edges throughout the region interior to the skull. The wavelet algorithm affected the images less than the other compression algorithms. The presence of the skull especially affected observer performance on the FDCT compressed images. All of the findings demonstrated a flattening of the CD curve for large lesions. The results of a compression study using lossy compression algorithms is dependent on the characteristics ofthe image and the nature of the diagnostic task. Because of the high density bone of the skull, head CT images present a much more difficult compression problem than chest x-rays. We found no significant differences among the CD curves for the tested compression algorithms. Key Words: Image compression, contrast-detail analysis.

Evaluation of the diagnostic quality of chest images compressed with JPEG and wavelet techniques: a preliminary study

Image compression reduces the amount of space necessary to store digital images and allows quick transmission of images to other hospitals, departments, or clinics. However, the degradation of image quality due to compression may not be acceptable to radiologists or it may affect diagnostic results. A preliminary study with small-scale test procedures was conducted using several chest images with common lung diseases and compressed with JPEG and wavelet techniques at various ratios. Twelve board-certified radiologists were recruited to perform two types of experiments. In the first part of the experiment, presence of lung disease on six images was rated by radiologists. Images presented were either uncompressed or compressed at 32:1 or 48:1 compression ratios. In the second part of the experiment, radiologists were asked to make subjective ratings by comparing the image quality of the uncompressed version of an image with the compressed version of the same image, and then judging the acceptability of the compressed image for diagnosis. The second part examined a finer range of compression ratios (8:1, 16:1, 24:1, 32:1, 44:1, and 48:1). In all cases, radiologists were able to make an accurate diagnosis on the given images with little difficulty, but image degradation perceptibility increased as the compression ratio increased. At higher compression ratios, JPEG images were judged to be less acceptable than wavelet-based images, however, radiologists believed that all the images were still acceptable for diagnosis. Results of this study will be used for later comparison with large-scale studies.