ISSN: 0970-938X (Print) | 0976-1683 (Electronic)

Biomedical Research

An International Journal of Medical Sciences

Research Article - Biomedical Research (2017) Volume 28, Issue 4

A study on discrete wavelet transform compression algorithm for medical images

Thirumoorthi C1* and Karthikeyan T2

1Assistant Professor, Hindusthan College of Arts & Science, Bharathiar University, Coimbatore, India

2Associate Professor, PSG College of Arts and Science, Coimbatore, India

*Corresponding Author:
Thirumoorthi C
Hindustan College of Arts and Science
Bharathiar University, India

Accepted on August 19, 2016

Visit for more related articles at Biomedical Research

Abstract

Image compression is play vital role in medical area. In health field the bulk of medical image data produced every day is ever growing, mainly in grouping with the improved scanning resolutions and the importance of volumetric medical image data sets. In this work, six image compression methods are simulated. They are Karhunen-Loeve Transform (KLT), Walsh-Hadamard Transform (WHT), Fast Fourier Transform (FFT), Sparse Fast Fourier Transform (SFFT), Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT). The results of simulation are shown and compared different quality parameters of it are by applying on various lung cancer Computed Tomography (CT) scan medical images. The Discrete Wavelet Transform (DWT) method algorithm was given better results like Compression Ratio (CR), Structural Content (SC), Mean Square Error (MSE) and Peak Signal to Noise Ratio (PSNR) compare to other transform methods. The Discrete Wavelet Transform (DWT) technique given improved result compared with other methods in all evaluation measures.

Keywords

Medical image, Image compression, Karhunen-Loeve transform, Walsh-hadamard transform (WHT), Fast fourier transform (FFT), Sparse fast fourier transform (SFFT), Discrete cosine transform (DCT), Discrete wavelet transform (DWT), Lung cancer computer tomography images, Compression ratio (CR), Structural content (SC), Mean square error (MSE), Peak signal to noise ratio (PSNR).

Abbreviations

KLT: Karhunen-Loeve Transform; WHT: Walsh-Hadamard Transform; FFT: Fast Fourier Transform; FT: Fourier Transform; SFFT: Sparse Fast Fourier Transform; DCT: Discrete Cosine Transform; DWT: Discrete Wavelet Transform; CT: Computed Tomography; CR: Compression Ratio; SC: Structural Content; MSE: Mean Square Error; PSNR: Peak Signal to Noise Ratio

Introduction

Medical science field and medical zone have large amount of medical images. So they need reduction the large volume amount of medical images. The medical Image compression is a process that focused on decreasing the size without losing quality, decrease the storage space of medical images and information of medical image. Lung cancer is the most common cause of cancer death in the world. Tobacco smoking cause’s etiology breath and lung cancer. Patients at highly effected by the lung cancer [1]. It could present with a variety of indications and is regularly screened by more common sicknesses. Several patients report no indications what so always, so it is vital that exact questions be enquired to try to produce info about symptoms they may not reflect abnormal [2]. In the case of an abnormal chest x-ray, the patient should directly undergo a chest CT scan for further evaluation. Figure 1 shows that the unusual lung CT’s inpatients who originally obtainable with weight loss and cough. The positron emission tomography was clarifying that the extent of the cancer [3]. When lung cancer is assumed, the patient stated to an oncologist, who will decide the supreme suitable diagnostic process (e.g., Bronchoscopy, CT guided lung biopsy, sputum cytology or media stinoscopy) and utmost favourable whole course of therapy for the lung cancer patient's specific needs [4].

biomedres-lung-image

Figure 1. CT scan lung image.

Image Compression Methodology

In this work to take different lung cancer CT scan image and using different image compression to compress image without loss. Walaa et al. [5] said the compression analysis results have indicated that the performance of the suggested method is much better, where the constructed medical images are less distorted and compressed with higher factor. The Image compression is process to remove the redundant information from the image so that only essential information can be stored to reduce the storage size, the transmission bandwidth and transmission time. Figure 2 shows that the essential information is extracted by various transforms techniques such that it can be reconstructed without losing quality and information of the image [6]. In this work comparative analysis of image compression is done by seven transform methods [7], which are Karhunen-Loeve Transform (KLT), Walsh- Hadamard Transform (WHT), Sparse Fast Fourier Transform (SFFT), Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT).

biomedres-Compression-procedure

Figure 2. Compression procedure.

Karhunen-loeve transform (KLT)

Thus for any image KLT is the optimal information compaction transformation as well as for any number of retained coefficients. KLT basic functioning is dependent on image but, this however makes pre-computing impossible and hence this does not qualify it as a practical option suitable for image compression [7]. The KLT kernel is a unitary matrix, [V], whose columns, {uk} (arranged in descending order of eigenvalue amplitude), are used to transform each zero-meaned vector:

image (1)

Walsh-hadamard transform (WHT)

Walsh system functions forms the very basis for Walsh transform. Walsh functions are orthogonal and have only +1 and -1 values [8]. In general, the Walsh transform can be generated by the hadamard matrix as follows:

image (2)

WHT here was employed to facilitate face as well as signature [9] recognition as WHT coefficients output array comprises only of integer values, resultantly hadamard transform thus becomes a quick transform, and implementation is feasible in O (N log N) additions as well as subtractions.

Fast fourier transform

Fourier transform is known to decompose image as two components: sinus and cosines. This implies that FT will transform any image from spatial to frequency domain respectively [10]. Fact is that any function can be precisely approximated using sum of infinite sinus and cosines functions [11]. The FFT is used to compute the Discrete Fourier Transform (DFT). The bigger size M=N × N matrix split into smaller size N’s of DFT in the time domain. The new radix-2 Decimation in Time (DIT) FFT is the effective and simplest method. It’s also used to compute 2n point DFT. In the FFT method the given weights are multiplied by indexes [5]. The output should be the bit reverse form of the given input signal. The compound conjugate regularity is used to signify the imaginary part. Fourier transform essentially is a method to execute this. Mathematically a two dimensional images Fourier transform in Equation 3.

image → (3)

Sparse fast fourier transform (SFFT)

SFFT computes k-sparse approximation, whose runtime depends upon k, the larger coefficient in the signal. This algorithm works in the process of identifying these k values by a filter G. ‘G’ concentrates in both time and frequency.

G is zero for all values except at a small number of time co-ordinates, and the Fourier transform of (�) is negligible except at a small fraction value which is about 1/k of the frequency co-ordinates. Gaussian or Dolph-Chebyshev function convolved with a box-car function is used to identify the large coefficients in sFFT [12]. Using this filter to find location and magnitude of the higher frequency component, parallel acceleration is achieved easily. Previous sub-linear algorithms have a runtime which is a polynomial in k and log n3. The fastest runtime of these algorithms will be calculated from Equation 4.

image (4)

For some constants C>2.

The key feature of sFFT is its simple structure. It has an efficient runtime with low big-0th Constant. For a typical case of n which is a power of 2, the run time is calculated using Equation 5.

image (5)

The implementation method and the working of sFFT with mathematical derivative.

Discrete cosine transform

The Discrete Cosine Transform (DCT) is an orthogonal transform, its attempts to de-correlate the medical image data. After the de-correlation each transform coefficient can be independently encoded without losing the compression efficiency [13]. The DCT transforms the signal from the spatial representation into the frequency representation. The DCT signify an image as a summation of sinusoids of the varying magnitudes and frequencies. DCT property is a typical image utmost of the visually substantial information about an image is focused in just few coefficients of DCT.

Afterwards the computation of the DCT coefficients, they are regularized according to a quantization table with the different measures provided by the JPEG image standard computed by the psycho visual evidence. The Selection of quantization table affects the compression ratio and entropy. The value of quantization is inversely proportional to quality of reconstructed image. Its gives better compression ratio and better mean square error [14]. In a quantization, the less significant frequencies are discarded, and then the maximum important frequencies that remain are used to recover the image in decomposition method. After quantization, the quantized coefficients are rearranged in the zigzag order for the further compressed by an effective lossy coding algorithm. DCT has several merits: (1). It has the ability to pack more information in least coefficients. (2). It decreases the block like appearance called the blocking artifact that outputs when margins between sub-images become visible [15].

An image is signified as the two dimensional matrix, the 2D DCT is used to compute a DCT coefficients of an image. The 2D DCT for an N × N input order can be defined with Equation 6.

image Where, P (x, y) is the input matrix image N × N, (x, y) are the coordinate of the matrix elements and (i, j) are the coordinate of the coefficients, and C (u) is calculated using Equation 7.

image (7)

The reconstructed image is the computed by using the inverse DCT (IDCT) according to the Equation 8.

image

The black and white image pixels are ranged from 0 to 255, where 0 matches to pure black and 255 matches to a pure white. As DCT is constructed to work on pixel the values ranging from -128 to 127, the original block is the level off by 128 from every entry [16]. The step by step procedure of getting compressed image is using DCT and getting reconstructed image from the compressed image.

Currently encoding is completed and transmitted to receiver side in the form of the one dimensional array. This transmitted order saves in the text format. The array of compressed blocks that establish the image is stored in an extremely reduced amount of the space. Further the compression can be reached by applying suitable scaling factor [17]. In order to rebuild the output data, the rescaling and the de-quantization should be achieved as given in Equation 9.

image (9)

The de-quantized matrix is then transmuted back using the 2D inverse DCT. The equation for the 2D inversed DCT change is given in the above said Equation 10.

image

DCT separates images into the parts of different frequencies where less important frequencies are the discarded through quantization and important frequencies is used to retrieve to the image through decompression [8]. The following Equation 11 is given by the forward 2D-DCT transformation:

image

Where u, v=0, 1, 2, 3, …, N-1

The following Equation 12 is given by the inverse 2D-DCT transformation:

image

Where

D (u) = (1/N) ^ 1/2 for u=0

D (u) =2(/N) ^ 1/2 for u=1, 2, 3, ……., (N-1)

Compared to the other input dependent transforms, DCT have several advantages: (1) It have been implemented in single combined circuit; (2) It has the capability to pack most info in fewest coefficients; (3) It reduces the block like appearance so-called blocking artifact that results when limits between sub-images convert visible [18,19].

Discrete wavelet transform

The wavelet can be made from a scaling function which describes its scaling properties [20]. The limit that the scaling functions important be orthogonal to its discrete conversions implies some mathematical conditions on them which are mentioned ubiquitously, e.g. the dilation Equation 13.

image (13)

Here ‘S’ is a scaling factor. Furthermore, the area between the function essential be normalized and scaling function should be orthogonal to its integer translations, i.e. in Equation 14.

image (14)

After presenting some more conditions (as the limitations above don’t produce a discrete solution) we can obtain results of all these calculations, the finite set of coefficients ak that describe the scaling function and similarly the wavelet [21]. The wavelet is got from the scaling function as N. Here ‘N’ is an even integer. The given set of wavelets then constructs an orthonormal basis which it uses to decompose the signal. The research is made with 128 × 128 medical image to compression and decompression with SFFT, DCT, DWT and calculating PSNR, MSE, CR, SC.

Performance Evaluation

Peak signal to noise ratio (PSNR in dB)

PSNR (Peak Signal-to-Noise Ratio) approximates image quality index, however on its own it does not have the capability of drawing a comparison between the features of two distinct images [22]. There is a possibility however that lower PSNR image may be considered as an image whose tone quality is comparatively better than one with a higher signal to noise ratio. PSNR can be calculated using Equation 15.

image (15)

It refers to the relation between the power of the humiliating noise that affects the loyalty of its representation and the supreme imaginable power of the signal. As several signals inherently have an extensive dynamic range, PSNR generally is expressed as logarithmic decibel scale.

Mean square error (MSE)

It is a measure of image quality index. Large value of mean square basically implies that a poor quality image [23]. Mean square error called as MSE. Generally, it is a criterion that has been widely used and is representative of the classical error estimate as denoted by Equation 16.

image (16)

Where M and N are the image dimensions.

Compression ratio (CR)

Compression Ratio is the ratio of number bits required to represent the data before compression to the number of bits required to represent data after compression [24]. Increase of compression ratio causes more effective compression technique employed and vice versa. CR is calculated using Equation 17.

image (17)

Structural content (SC)

Structural content measure that is employed to draw a comparison between two images inherent in several small image patches and to determine the images have in common [25]. A comparison is drawn between patches by deploying 2D continuous wavelet that acts as a low level corner detector. Large value of structural content SC basically implies a poor quality image. SC is calculated using Equation 18.

image (18)

Experimental Results and Discussions

The experiment is carried out in MATLAB environment. In this research, the images deployed here for compression are all 128 × 128, 256 × 256 and 512 × 512 medical images. Basically meaning that images comprise of 128 × 128=16,384 pixels, 256 × 256=65536 pixels and 512 × 512=262144 pixels, wherein each pixel is represented by 8 bits or 1 byte. KLT, WHT, FFT, SFFT, DCT and DWT algorithm implementation has been carried out on MATLAB software and several different parameters have been ascertained with respect to the three medical images; comparison of reconstructed and original images PSNR, MSE, CR, SC values have been estimated.

In Table 1, to take 128 × 128 size three CT medical images and simulated images using KLT, WHT, FFT, SFFT, DCT and proposed DWT compression techniques. In CT image 1 the proposed DWT compression produce better result in PSNR value is 36.9746 less than KLT, WHT, FFT, SFFT and DCT value is 43.3436, 41.0181, 40.7637, 39.7656 and 38.9746 respectively. In CT image 2 the proposed DWT compression produce better result in PSNR value is 36.874 less than KLT, WHT, FFT, SFFT and DCT value is 51.9335, 48.5279, 44.6743, 42.3932 and 40.6542 respectively. In CT image 3 the proposed DWT compression produce better result in PSNR value is 35.584 less than KLT, WHT, FFT, SFFT and DCT value is 52.8757, 50.6854, 47.9874, 45.0043 and 38.8023 respectively.

In CT image 1 the proposed DWT compression produce better result in MSE value is 15.342 higher than KLT, WHT, FFT, SFFT and DCT value is 2.342, 3.165, 8.437, 10.373 and 14.424 respectively. In CT image 2 the proposed DWT compression produce better result in MSE value is 19.653 higher than KLT, WHT, FFT, SFFT and DCT value is 7.0676, 9.9067, 10.0932, 12.3254 and 16.232 respectively. In CT image 3 the proposed DWT compression produce better result in MSE value is 19.313 higher than KLT, WHT, FFT, SFFT and DCT value is 6.874, 7.429, 10.453, 12.842 and 18.453 respectively.

In CT image 1 the proposed DWT compression produce better result in CR value is 7.7642 higher than KLT, WHT, FFT, SFFT and DCT value is 2.6822, 3.7365, 4.4844, 5.3433 and 5.9784 respectively. In CT image 2 the proposed DWT compression produce better result in CR value is 6.654 higher than KLT, WHT, FFT, SFFT and DCT value is 1.369, 2.9231, 3.242, 4.043 and 4.985 respectively. In CT image 3 the proposed DWT compression produce better result in CR value is 7.784 higher than KLT, WHT, FFT, SFFT and DCT value is 2.836, 3.1302, 4.043, 5.132 and 6.382 respectively.

In CT image 1 the proposed DWT compression produce better result in SC value is 1.987 less than KLT, WHT, FFT, SFFT and DCT value is 3.984, 3.098, 2.965, 2.615 and 2.105 respectively. In CT image 2 the proposed DWT compression produce better result in SC value is 2.004 less than KLT, WHT, FFT, SFFT and DCT value is 4.534, 4.053, 3.785, 3.166 and 2.816 respectively. In CT image 3 the proposed DWT compression produce better result in SC value is 1.786 less than KLT, WHT, FFT, SFFT and DCT value is 3.723, 3.108, 2.768, 2.564 and 2.008 respectively. In order to 256 × 256 and 512 × 512 size of CT medical scan image also had the better performance for given proposed DWT method.

Figures 3-6 and Tables 1-3 explained the PSNR, MSE, CR and SC parameters performance evaluation of 128 × 128, 256 × 256 and 512 × 512 size respectively. The DWT method gave less SC value are given three CT image. Also it produced high MSE and CR value compare to the other exiting methods. It also shows in Figures 3-6.

biomedres-Zigzag-ordering

Figure 3. Zigzag ordering for DCT coefficients.

biomedres-medical-images

Figure 4. Graphical results of performance analysis for 128 128 medical images.

biomedres-performance-analysis

Figure 5. Graphical results of performance analysis for 256 256 medical images.

biomedres-Graphical-results

Figure 6. Graphical results of performance analysis for 512 512 medical images.

Medical image Image compression technique used PSNR MSE CR SC
Image 1 KLT 43.3436 2.342 2.6822 3.984
WHT 41.0181 3.165 3.7365 3.098
FFT 40.7637 8.437 4.4844 2.965
sFFT 39.7656 10.373 5.3433 2.615
DCT 38.9746 14.424 5.9784 2.105
DWT 36.5643 15.342 7.7642 1.987
Image 2 KLT 51.9335 7.0676 1.369 4.534
WHT 48.5279 9.9067 2.9231 4.053
FFT 44.6743 10.0932 3.242 3.785
sFFT 42.3932 12.3254 4.043 3.166
DCT 40.6542 16.232 4.985 2.816
DWT 36.874 19.653 6.654 2.004
Image 3 KLT 52.8757 6.874 2.836 3.723
WHT 50.6854 7.429 3.1302 3.108
FFT 47.9874 10.453 4.043 2.768
sFFT 45.0043 12.842 5.132 2.564
DCT 38.8023 18.453 6.382 2.008
DWT 35.584 19.313 7.784 1.786
KLT: Karhunen-Loeve Transform; WHT: Walsh-Hadamard Transform; FFT: Fast Fourier Transform; FT: Fourier Transform; SFFT: Sparse Fast Fourier Transform; DCT: Discrete Cosine Transform; DWT: Discrete Wavelet Transform; CT: Computed Tomography; CR: Compression Ratio; SC: Structural Content; MSE: Mean Square Error; PSNR: Peak Signal to Noise Ratio

Table 1. Performance evaluation of various image compression techniques for 128 128 size medical image.

Medical image Image compression technique used PSNR MSE CR SC
Image 1 KLT 46.2234 3.659 21.6598 3.354
WHT 44.5483 4.243 26.7483 2.865
FFT 42.6754 7.549 28.3232 2.245
sFFT 41.7685 9.085 30.5442 2.098
DCT 39.9867 11.875 32.9322 1.856
DWT 38.0933 18.386 34.0985 1.4325
Image 2 KLT 49.6754 8.6087 15.8769 2.764
WHT 46.2376 9.9435 18.6741 2.456
FFT 44.5643 11.0874 26.7834 2.006
sFFT 40.4524 14.3453 29.9454 1.893
DCT 39.6654 17.4342 32.8948 1.534
DWT 35.784 19.6504 35.656 1.324
Image 3 KLT 48.054 7.3453 13.543 3.654
WHT 47.283 9.0273 18.954 3.104
FFT 45.132 11.3776 24.543 2.875
sFFT 41.242 14.242 27.23 2.505
DCT 37.0398 17.3453 32.532 1.813
DWT 35.765 21.6435 36.532 1.165
KLT: Karhunen-Loeve Transform; WHT: Walsh-Hadamard Transform; FFT: Fast Fourier Transform; FT: Fourier Transform; SFFT: Sparse Fast Fourier Transform; DCT: Discrete Cosine Transform; DWT: Discrete Wavelet Transform; CT: Computed Tomography; CR: Compression Ratio; SC: Structural Content; MSE: Mean Square Error; PSNR: Peak Signal to Noise Ratio

Table 2. Performance evaluation of various image compression techniques for 256 256 size medical image.

Medical image Image compression technique used PSNR MSE CR SC
Image 1 KLT 49.3488 4.4732 25.546 2.985
WHT 45.4341 5.1613 29.337 2.687
FFT 43.8734 9.8453 31.1216 2.045
sFFT 41.6484 12.1232 33.0082 1.767
DCT 39.7563 18.4633 34.6578 1.546
DWT 37.763 21.3412 36.425 1.365
Image 2 KLT 43.6357 2.8152 15.5156 2.964
WHT 42.6754 4.8243 20.9442 2.404
FFT 39.7542 6.2434 23.5433 1.897
sFFT 36.9232 10.3433 26.2353 1.761
DCT 33.4599 19.315 31.8172 1.286
DWT 32.764 30.123 38.124 1.004
Image 3 KLT 42.874 3.986 19.4232 3.267
WHT 41.453 5.343 22.342 2.682
FFT 39.354 7.342 24.345 2.103
sFFT 35.343 9.0345 26.432 1.973
DCT 34.547 19.453 30.5422 1.675
DWT 32.843 28.4126 36.8724 1.008
KLT: Karhunen-Loeve Transform; WHT: Walsh-Hadamard Transform; FFT: Fast Fourier Transform; FT: Fourier Transform; SFFT: Sparse Fast Fourier Transform; DCT: Discrete Cosine Transform; DWT: Discrete Wavelet Transform; CT: Computed Tomography; CR: Compression Ratio; SC: Structural Content; MSE: Mean Square Error; PSNR: Peak Signal to Noise Ratio

Table 3. Performance evaluation of various image compression techniques for 512 512 size medical image.

Conclusion

This work is concentrated on the comparison among six main KLT, WHT, FFT, SFFT, DCT and DWT methods of image compression. In this paper analysis of various Image compression approaches for different images is done based on parameters, Mean Square Error (MSE), Compression Ratio (CR) and Peak Signal to Noise Ratio (PSNR). The comparison of KLT, WHT, FFT, SFFT, DCT and DWT after compression the parameters like PSNR values, MSE, CR and SC values were good outcome in proposed DWT work. From the above study it can be seen that proposed DWT take less PSNR value and SC value than KLT, WHT, FFT, SFFT and DCT. The proposed method has more compression ratio and MSE value than KLT, WHT, FFT, SFFT and DCT. Our simulation results from result and discussion chapter shows that we can achieve higher compression ratio using DWT compression technique. DWT transform gives higher compression ratio but for getting that clarity of the image is partially trade off. It is more suitable for regular applications as it is having a good compression ratio along with preserving most of the information. Further research work can be performed to relax high compression ratio constraint and transform. This work has been concentrate transform image using for the elimination of the spatial redundancy by different compression of medical images.

References