Enhanced textural analysis for endometrial tuberculosis identification from ultrasound images

Abstract

Endometrial Tuberculosis (ETB) is a major observation in Infertility due to Female Genital Tuberculosis (FGTB). Transvaginal ultrasound (TVUS) images are used to initiate the diagnostic procedure. An effective computational method to identify and classify ETB from TVUS images is presented in this paper. The images from female patients have been collected from medical centers in India. A composite co-occurrence model is proposed; where a Non-Subsampled Contourlet Transformation (NSCT) enhances the textural appearances of the images before feature extraction. Since original TVUS are inherently ill defined; the transformation highlights the directional, multi- scale spectral texture features for discriminatory analysis. Experimental results show that the textural features extracted from these transformed images are more discriminative. The proposed approach achieved an F-score of 0.821and a sensitivity of 0.801 for the dataset in hand. A dimensionality reduction of 66.6% is achieved by Entropy based Normalized Mutual Information ranking and subsequent feature subset selection for classification.