Some studies on the use of Curvelet and Wavelet transforms for medical image processing

Abstract

Abstract Over the last few decades there is a rapid development in the field of medical image processing. Scientists and engineers all over the world have developed advanced diagnostic tools to detect the diseases at any level of complicacy. Keeping pace with fast and accurate diagnosis, this thesis aims at exploring the use of two multiresolution mathematical transforms called wavelet and curvelet in medical image processing. A survey on fifteen different advanced wavelets including curvelet is performed. The two transforms (wavelet and curvelet) are comprehensively studied and selected as the medical image analyzing tools. The different investigating areas of this thesis are the detection of human skin ringworm, Darier disease detection, diabetic retinopathy (DR) detection and denoising of mammogram. Three different approaches on denoising of mammogram are presented in this thesis. The human skin ringworm is investigated in the light of computer vision. Two independent methodologies are developed for its detection. The first methodology implements three level multi-wavelet decomposition of the skin images and subsequent evaluation of the approximation and detail subband energies which act as the texture characterizing features. The second methodology incorporates the curvelet to segment the circular protrusion of the skin images especially associated with ringworm images followed by statistical texture investigation. After feature extraction by both the methodologies, binary classifier called the support vector machine (SVM) recognizes the images as ringworm with detection accuracy of around 87% and 80% for the first and second methodologies respectively. In addition, the performance indexing parameters obtained from SVM classification like sensitivity, specificity, Positive Predictive Value (PPV) and Negative Predictive Value (NPV) are evaluated. Both the methodologies are comprehensively demonstrated and compared to select the better one. The selected method is then compared with the available techniques and commented upon. viii This thesis also deals with the detection of a genetic disorder called Darier disease manifested as dermal changes. It incorporates three methodologies. They are - 1) gray level co-occurrence matrix (GLCM), 2) local binary pattern (LBP) and 3) wavelet energy feature for skin texture recognition. The feedforward neural network (FNN) is implemented for Darier disease detection. All the methodologies are thoroughly compared to find the most suitable one as skin texture screening tool. The GLCM, LBP and wavelet based methodologies attain Darier disease detection accuracy of about 82%, 82% and 89% respectively. Though the GLCM based methodology provides inferior detection accuracy than wavelet based methodology, but it addresses the presence and location of typical skin textural abnormalities. The early detection of DR and its subsequent medication is of prime importance to the medical practitioners. This thesis focuses on the implementation of the curvelet transform to segment the blood vessels from the retina images. The blood vessel characterizing features of curvelet filtered images are quantified by the retinal image analyzer. The computed vessel features of DR and non-diabetic retinopathy (NDR) images are used to tabulate databases which are provided to a FNN for DR detection. The DR is detected with sensitivity of 83.9%, specificity of 77.4% and accuracy of 80.6% which is better than the results obtained by other researchers. Mammogram is an easy and affordable means of diagnosis of breast cancer. Like other medical data acquisitions, it is also commonly affected by noise. Therefore it is a challenge for the researchers to denoise the mammograms for clear data extraction. This thesis targets at denoising of mammogram by wavelet and curvelet transforms with a focus to investigate the role of an “embedded” thresholding algorithm. As the thresholding technique helps to compress the transform generated coefficients, an effort is made to find how the change of the thresholding algorithm coded within either the wavelet/curvelet transform produces different signal to noise ratio (SNR) of the denoised image. The scientific exploration also goals at finding the dependence of SNR of the output image on the type and depth of the noise. A standard mammogram is selected and different types of noise are added. Next, the noisy mammogram is denoised by the wavelet and the curvelet transforms. The commonly used thresholding techniques called hard, soft and block are implemented for noise reduction. Finally, a comparison table is drawn to find the applicability of the transforms with the incorporated thresholding algorithms. Generally a mammogram is denoised by curvelet transform based on conventional thresholding called hard thresholding (HT). Therefore, the motive of this investigation is to denoise mammogram with the same transform but with efficient thresholding technique. This thresholding technique is based on the information of the neighbouring curvelet coefficients which are generated after the application of fast discrete curvelet transform. It is also known as block thresholding. It is found that the curvelet transform applied with three different block thresholding techniques is visually and statistically better than the conventional approach. It has also been found that the denoising performance of wavelet, contourlet and curvelet implemented on mammogram with Poisson noise is unique in the sense that SNR of denoised mammogram by wavelet is better than that of by contourlet which in turn is better than that of by curvelet. The first part of this investigation deals with the confirmation of the above exceptional denoising performance (trend) with the result obtained by our approach. The later part of our investigation implements the recently developed denoising approach called the Poisson Unbiased Risk Estimation– Linear Expansion of Thresholds (PURE-LET) to the Poisson noise corrupted mammogram with an objective to improve the SNR. The PURE-LET successfully removes Poisson noise in a way better than the traditional mathematical transforms mentioned above.