classification of endometrial images for aiding the diagnosis of hyperplasia using logarithmic gabor wavelet

introduction: the process of discriminating among benign and malignant hyperplasia begun with subjective methods using light microscopy and is now being continued with computerized morphometrical analysis requiring some features. one of the main features called volume percentage of stroma (vps) is obtained by calculating the percentage of stroma texture. currently, this feature is calculated by pathologists on a manual basis. the proposed algorithm can automatically calculate vps with a good precision. this procedure could be the first and an essential step of diagnosing endometrial hyperplasia in the field of pathology. materials and methods: in this paper, a method based on logarithmic gabor wavelets and mathematical morphology is proposed for the segmentation of texture. attaining maximum joint space-frequency resolution is highly significant in the process of texture segmentation. the ability of logarithmic gabor filters in the discrimination of texture bands at different scales and orientations has been used to segment stroma texture which is mainly located in the middle frequency bands from glandular elements. in the proposed algorithm, the logarithmic gabor filter bank applied on the microscopic endometrial images and mathematical morphology is used for denoising. the segmentation method is averaging the weighted and denoised gabor filter outputs. results: the images used in this method segment with a good precision and can be used by pathologists to calculate volume percentage of stroma as an aided diagnosis feature. a sensitivity of 95.3, specificity of 95.6, accuracy of 95.4%, ppv of 98.4% and npv of 88% was achieved in distinguishing between benign and malignant hyperplasia based on vps. discussion and conclusion: the proposed procedure could be the first and an essential step of diagnosing endometrial hyperplasia in pathology.based on these experiments, the logarithmic gabor wavelet is one of the most effective methods in texture segmentation that can easily extract the texture information from the middle frequency bands and make it available to the segmentation algorithm.