Identification of endometrial tuberculosis in infertility using Non-Subsampled Contourlet based convolution neural network

Abstract

Background and ObjectiveFemale Genital Tuberculosis (FGTB) has a major impact on female fertility but it can be effectively treated on timely diagnosis. Early, objective and reliable diagnostics of FGTB causing infertility is a need of the hour in a populous country like India. As per our literature study, there is no computational method available for the same till date. Since, Transvaginal Ultrasound (TVUS) imaging is a non-invasive, primary and first line investigative technique; aim of this work is to develop an effective method for diagnosis of abnormal endometrial TB from TVUS images. MethodTVUS images of female patients coming for infertility treatments to medical centers in India are collected under the supervision of medical expert. These images are preprocessed and used to train the proposed model. In this paper, a multi-scale and multi-direction Non-Subsampled Contourlet Transform (NSCT) based CNN model is proposed that integrates different levels of NSCT transformed features with different layers of a pre-trained CNN model. NSCT does spatial as well as spectral analysis and extracts significant features irrespective of orientation. The integration of these features makes the model more effective in discriminating normal and abnormal TB images. Further, the integrated model has less number of trainable parameters due to intrinsic multi-scale nature of NSCT. Results and ConclusionTo test the effectiveness of the proposed NSCT-CNN model, traditional CNN with different pooling methods, Wavelet-CNN and Contourlet-CNN models were also implemented. The experimental results show that the proposed model has an improved efficacy over other related models implemented with an average testing accuracy of 88%, a sensitivity of 0.832 and F1-score of 0.869. Two tailed t-test conducted on the model performances are statistically significant at 95% confidence level for the data in hand. The model shows an improved efficiency over traditional model with 16.01% reduction in number of trainable parameters and 41.08 % reduction in training time.