Texture Classification of Diffused Liver Diseases Using Wavelet Transforms

Introduction: A major problem facing the patients with chronic liver diseases is the diagnostic procedure.  The conventional diagnostic method depends mainly on needle biopsy which is an invasive method. There  are  some  approaches  to  develop  a  reliable  noninvasive  method  of  evaluating  histological  changes  in  sonograms. The main characteristic used to distinguish between the normal, hepatitis and cirrhosis liver is the  texture of liver surface. The problem of defining a set of meaningful features that explores the characteristics  of the texture, leads to several methods of determining tissue texture. Some of these methods, which have  been developed so far, are based on wavelet transform. The selection of wavelet transform type affects the  accuracy of determining the texture. In this study, an optimal wavelet transform called Gabor wavelet was  introduced and three different methods of determining tissue texture were evaluated. These include statistical,  dyadic wavelet transform and Gabor wavelet transform methods.    Materials and Methods: The proposed algorithm was applied to differentiate ultrasonic liver images into  two disease states (hepatitis and cirrhosis) and normal liver. In this experiment, 50 liver sample images for  each three states which already been proven by needle biopsy were used. These images are taken from a  Toshiba Sonolayer SSA250A device using a 3.75 MHz transducer. For each image, a region of interest (ROI)  with 75×35 pixels is selected. The ROI is chosen to include only liver tissue without major blood vessel or  hepatic duct.  The  classification  method  used  for  this  work  is  "Minimum  Distance",  where  the  distance  is  calculated  between feature vectors of test image and reference images. In order to evaluate the diagnostic results, two  quantities named “Sensitivity” and “Specificity” were calculated for each method.  Results: The obtained results show that Gabor wavelet has 85% and dyadic wavelet has 77% sensitivity in  the  hepatitis  liver  images.  On  the  other  hand,  Gabor  wavelet  shows  86%  sensitivity  in the cirrhosis liver  images, while dyadic wavelet has 78%. The specificity of Gabor wavelet in the hepatitis and cirrhosis liver  images is 77% and 79% respectively, while the specificity of dyadic wavelet is 65% and 72%, respectively.  Discussion  and  Conclusion:    Based  on  this  experiment,  the  Gabor  wavelet  is  more  appropriate  than  the  dyadic wavelet and statistical based method for the texture classification as it leads to higher classification  accuracy,  because  the  dyadic  wavelet  loses  some  middle-band  information,  while  the  Gabor  wavelet  preserves it. Based on what was observed, the most significant information regarding the texture is mainly  located in the middle-frequency bands of wavelet decomposition. Therefore, using Gabor wavelet, a more  flexible decomposition of the entire frequency band can be achieved leading to a superior differentiation of  the texture information.