A Deep Multi-Modal Approach for Automatic Analysis of Tuberculosis-related Lung Diseases from Computed Tomography (CT) Images
Authors: Alebiosu David Olayemi, Anuja Dharmaratne, Lim Chern Hong
Affliation: Monash University, Malaysia
Abstract
Tuberculosis is a deadly disease and it is caused by a germ called mycobacterium. Tuberculosis is a bacterial infection which usually affects the lungs but can be treated if early diagnosis is carried out on the patient. Though it was discovered about 130 years ago, it has continued to remain a constant threat and has become a major cause of death across the globe. Presently, tuberculosis is the leading infectious diseases which have caused millions of deaths, giving rise to an estimated 2.5 million new cases every year. The need to pay urgent attention to effective diagnosis of tuberculosis is becoming increasingly clear with the recent outbreak of the novel COVID-19 pandemic across the globe. Tuberculosis has been found to be one of the most serious underlying health conditions that can slow down the recovery process of any COVID-19 patient. In medical imaging, the use of handcrafted techniques previously employed by researchers for CT image analysis has been ineffective due to their limitations in extracting discriminative features from the images. Over the last 10 years, deep learning applications have been employed to eradicate the challenges faced by the handcrafted techniques and has produced outstanding results. Our research aims to employ the use of convolutional neural networks (CNN) and recurrent neural networks (RNN) for the automatic analysis of chest CT scan images for tuberculosis severity assessment and for the automatic CT report generation. We first present a comprehensive pre-processing technique with the employment of an improved adaptive thresholding + CNN for segmentation and the use of 2 dimensional (2D) CNN + Bi-directional Long Short Time Memory (Bi-LSTM) for the classification task. Our proposed technique is expected to produce a more efficient segmentation approach and improve the classification accuracy of chest CT images for tuberculosis-related diagnosis.