Marcos Antonio Martins de Almeida*
Department of Electronics and Systems, Federal University of Pernambuco, Brazil
Received: 10 November 2018
Accepted: 13 December 2018
Version of Record Online: 4 January 2019
Martins de Almeida MA (2018) Use of Statistical Techniques to Analyze Textures in Medical Images for Tumor Detection and Evaluation. Adv Mol Imaging Interv Radiol 2018(1): 01-06.
Correspondence should be addressed to
Marcos Antonio Martins de Almeida, Brazil
Copyright © 2018 Marcos Antonio Martins de Almeida. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and work is properly cited.
Detection of malignant tumors in the diagnostic phase is a concern of clinicians, radiologists and oncologists. Medical images contribute greatly to early diagnosis. The use of data extraction techniques from 2D / 3D medical images of human tissues suspected of tumors and the use of statistical techniques are efficient for the detection of these tumors. This paper discusses recent advances in understanding the segmentation and representation of visual textures, in order to show predictive and probabilistic models through the analysis of human tissue textures. Texture feature is an important source of information for the process of image analysis and interpretation.
Keywords: Haralick’s Parameters; Region of Interest; Texture Analysis; Tumor
One of the most promising applications of texture analysis in medicine is the early detection of tumor signals and prophylaxis of cancer according to Gentillon .
The benefit of medical imaging depends on a precise diagnosis and this, in turn, depends on the quality of the image acquisition and its interpretation by specialists. The quality of the image acquired depends on the resolution, sensitivity, bandwidth and signal-to-noise ratio of the imaging system. Computer vision can, for example, be used to assist radiologists in focusing their attention on diagnostically relevant information and to provide quantitative measurements for suspicious regions . Studies on the influence of noise on ultrasound images have been presented by Singh . The experimental results presented show how these features are able to provide an accurate quality measurement that corresponds well with the subjective evaluations performed by clinical experts.
The tumor is an exceptional expansion generated by human cells reproducing in an unrestricted way. This can be identified by a variations in the texture of the human tissue under study, since the information contained in the images is of extreme value. Textures are visual patterns in the image pixels, which have brightness, color, slope, size, and other attributes, but which, once partitioned into sub-images in regions of interest, can be used for classification purposes.
Texture can be characterized by the spatial distribution of the intensity of pixels in a neighborhood. If two contiguous regions in an image have a different surface texture, this may lead to the detection of intermediate texture boundaries. Much of the work on perception is concerned with enabling observers to discriminate, without effort, certain pairs of textures.
In the medical field, radiologists are more interested in Region of Interest (ROI) rather than whole image. ROI is a subpart of the image that contains very important information related to the diagnosis. In addition, ROI size has been known to influence the sensitivity and specificity of the classification .
It turns out that human vision, though perfect, cannot accurately distinguish variations between low intensities from neighboring pixels, especially if the quality of the medical image acquired does not have a good resolution, which generally contributes to the noise present in the imaging process. Advancements in medical imaging systems have made it possible to obtain texture information which is not visible to the human eye.
The probabilistic methods initially used in texture analysis in images are the first order statistics, using the histogram of the image, or the probability of occurrence of pixels. The histogram is calculated from the intensity of the pixels, without taking into account any spatial relationships between the pixels within the image. The characteristics are simply statistical parameters of the histogram distribution, such as: mean brightness, variance, skewness, kurtosis and percentiles.
Similar to the first-order characteristics, the histogram of the image gradient is calculated and the statistical parameters of these histograms are determined. Haralick  proposed an approach for the development of new algorithms for microtextures analysis. Once the Gray Level Coocurrence Matrix (GLCM) is established, the statistical parameters can be calculated from this matrix. The fourteen textural features described by Haralick provide a wide range of parameters that can be used in image texture analysis, such as contrast, homogeneity, energy, entropy, mean, and variance. GLCM is one of the most studied and extensively used general approaches for texture analysis and has recently become the focus of study of several research groups whose aim is to increase the discriminability of GLCM descriptors . First-order texture analysis measurements use the image histogram, or pixel occurrence probability, to calculate texture.
In similar way to the image histogram features, the histogram of the image gradient is computed and statistical parameters of the histogram distribution are determined. Haralick provided the classic survey of texture measurements.
The statistical approaches are better suited to micro textures, for which Haralick identified analytical techniques based upon auto correlation functions, frequency domain analysis, edge operators, grey-level co-occurrence matrices, grey-level run lengths, and autoregressive models. But it has been contested that the discrimination between two image textures depends in great part of the difference in the second-order statistics of the textures. That is, for two textures with identical second-order statistics, a deliberate quantitative effort is needed to discriminate between them. In contrast, little effort is required when the second-order texture statistics are different. This observation, however, does not extend to textures that differ in a third or higher order, and cannot readily be discerned by the naked eye. Segmentation divides an image into discrete fields, so that the pixels in each region are similar and there is a visible distinction between regions, according to Agrawal .
Another statistical method is the Gray Level Run-Length Matrix (GLRLM) that calculates for each ROI features using Spatial Gray Level Dependence Matrices (SGLDM). This is a statistical method which consists in constructing co-occurrence matrices to reflect the spatial distribution of gray levels in the ROI. SGLDM is based on an estimate of the second order conditional probability density g (i, j, d, θ).
This means that a pixel element at location (i, j) of the SGLD matrix signifies the probability that two different resolution cells which are in a specified orientation θ from the horizontal and specified distance from each other, will have gray level values i and j respectively.
Different textural parameters of texture reflect different properties of the image. These parameters, derived from the Haralick matrix, reflect the characteristics such as: contrast, homogeneity, energy, entropy, mean, and variance. The study of developing these perceptual textural spaces was conducted by Rao . Twelve perceptual features aiming to capture different aspects of texture were elaborated selected for psychophysical experiments.
Juelesz  verified that discriminating between two image textures depends largely upon the difference in the second-order statistics of the textures. That is, for two textures with identical second-order statistics, a deliberate amount of effort is required to discriminate between them. In contrast little effort is required when the second-order statistics of the textures are different. However, this observation does not extend to textures that differ in third or higher-order statistics, which are not readily discriminated by eye. Image segmentation is one of the widely used methods to organize the pixels of an image accurately in a decision-oriented application. It divides an image into a number of discrete fields so that the pixels have high resemblance in each region and high distinction between regions. The significance of image analysis for cancer was treated by Sela .
Although there are other methods of texture analysis for tumor detection, statistical models accurately reveal, in the order of 85% to 95%, the clinical diagnoses.
Some research work was related to texture analysis for the detection of tumors in different tissue types. From these articles, their common conclusions, strengths, weaknesses, gaps and a comparisons of suggested techniques were extracted.
The statistical methods analyzed were used to calculate a variety of texture characteristics, through the parameters based on histogram matrices and co-occurrence. In some papers, texture analysis problems such as segmentation and classification were discussed.
Texture analysis uses the changes in the grey value of image pixels and their distribution pattern, to identify microscopic pathological changes that are not visible and that can be used in the analysis of various images. Texture analysis in medical imaging can be a substantial support for the clinical decision-making process in the diagnosis and classification of tumors. This methodology is expected to become more accurate than the human eye and mind in detecting minute deviations in cell and tissue structures. Although the results using statistical techniques were satisfactory, when second order GLCM parameters are used, it is necessary to take some precautions regarding the size of the region of interest. In some cases, the size of the ROI can change values for some parameters. For example, parameters describing the image homogeneity and complexity (angular second moment, entropy, sum entropy, and difference entropy) are examples of parameters that depend on the ROI size, especially with small ROI sizes, and approach a limit value, . The optimal ROI size also reduces the computational cost in extracting the texture features from the smaller ROI. The experimental results encourage the use of 200 × 200 pixels ROI for classifying a mammogram in one of the two (Fatty, Dense) breast density classes .
The results demonstrate that histogram parameters are highly dependent on variations in image contrast and brightness, and provide little additional information to that obtained by visual inspection. Features based on the GLCM and GLRLM contain information that cannot be evaluated visually. The size-dependence of specific features should be noted by standardizing the size and shape of the ROI .
Statistical methods that use GLCM to perform microtextural analyzes of human tissues and image classification for tumor detection have shown great efficiency. The accuracy of these analyzes reaches the level between 90% and 95%. This shows the effectiveness of second-order statistics.
It is hoped in the future to design and implement these web-based applications in medical-hospital equipment to assist clinicians, oncologists and radiologists in decision-making.
Copyright © 2019 - All Rights Reserved - ocimumpublishers.com