A unified framework for image compression and segmentation by using an incremental neural network

Abstract

This paper presents a novel unified framework for compression and decision making by using artificial neural networks. The proposed framework is applied to medical images like magnetic resonance (MR), computer tomography (CT) head images and ultrasound image. Two artificial neural networks, Kohonen map and incremental self-organizing map (ISOM), are comparatively examined.

Compression and decision making processes are simultaneously realized by using artificial neural networks. In the proposed method, the image is first decomposed into blocks of 8 × 8 pixels. Two-dimensional discrete cosine transform (2D-DCT) coefficients are computed for each block. The dimension of the DCT coefficients vectors (codewords) is reduced by low-pass filtering. This way of dimension reduction is known as vector quantization in the compression scheme. Codewords are the feature vectors for the decision making process.

It is observed that the proposed method gives higher compression rates with high signal to noise ratio compared to the JPEG standard, and also provides support in decision-making by performing segmentation.

Introduction

Medical images like magnetic resonance (MR) and computer tomography (CT) acquired from different modalities comprise huge amounts of data, rendering them impracticable for storage and transmission. Archiving this large amount of image data in the computer memory is very difficult without any compression. An important issue in lossy compression of medical images is the risk of destroying diagnostically relevant information. Current lossy compression standards, such as JPEG (Wallace, 1991) and MPEG, are designed for conventional still-image and video display.

Transform-based techniques have been proposed for the efficient reduction of the high redundancy usually encountered in real life images (Strintzis, 1998). Unsupervised neural networks can perform nonlinear principal component analysis as a transform-based method in image compression (Tzovaras & Strintzis, 1998). They outperform linear principal component analysis, and are relatively easy to implement.

Another common method to compress images is to code them through vector quantization (VQ) techniques. Self-organizing Kohonen maps have been used to achieve the VQ process of image compression (Amerijckx, Verleysen, Thissen, & Legat, 1998). They represent an efficient compression scheme based on the fact that consecutive blocks in an image are often similar, and thus coded by similar codewords with a VQ algorithm. Early studies of lossy compressed medical images performed compression using variations on the standard discrete cosine transform coding algorithm combined with scalar quantization and lossless coding. More recent studies of efficient lossy image compression combined with scalar/vector quantization (Tsekouras, 2005, Pal et al., 2000), or neural networks (Jiang, 1999).

The constitution of the right feature space plays a critical role in decision making. In order to construct realistic classifiers, the features that are sufficiently representative of the physical process must be searched. In the literature, it is observed that different transforms are used to extract desired information from medical images. Image intensities at one or two neighborhood of the pixel (Dokur et al., 2000) and image intensities in multiple images (T1, T2 and proton density) (Qian, Cheng, & Liang, 1989) are utilized to represent the tissues in magnetic resonance and computer tomography images. Wavelet transform (Qian, Li, & Clarke, 1999), co-occurrence matrix (Arrowsmith et al., 1999, Haering and Lobo, 1999), Fourier transform (Feleppa et al., 1996) and spatial gray-level dependence matrices (Wu, Chen, & Hsieh, 1992) are used to extract textures in ultrasound images. Zhang, Yoshida, Nishikawa, and Doi (1998) used wavelet transform for the detection of the microcalcifications in digital mammograms.

Recent developments in spatial-scale (or frequency) analysis such as Gabor transform, Wigner distribution (Reed & Wechsler, 1990), and wavelet transform (Zhang et al., 1998) have provided a new set of multi-resolution analytical tools. The scales that represent the textures in different image types must be determined before the classification process. Different scales are used to extract information from ultrasound images and digital mammograms. The determination of the scales and also the mother wavelet is time consuming.

The main difficulty of the above methods is due to the lack of an adequate tool that characterizes different scales of textures effectively. In this study, the elements of the feature vectors are formed by the coefficients calculated by applying 2D-discrete cosine transform (2D-DCT) (Ahmed, Natarajan, & Rao, 1974) to image blocks of 8 × 8 pixels. Only two parameters are adjusted to represent the textures in the image: The amount of ignored coefficients, and the size of the image blocks. The dimension of the feature vectors can be decreased by discarding some coefficients which represent high frequencies. This leads to a decrease in the computational times of learning and classification processes.

In this paper, a unified framework for image compression and segmentation schemes based on discrete cosine transform and vector quantization by artificial neural networks is presented.