Deep Convolutional Neural Networks for Lungs Segmentation and Diagnosis of Tuberculosis from Chest X-Rays

Abstract

Technology nowadays is revolving around intelligent systems that mimic the learning capability of human brain. These systems learn, adapt, act and make decisions autonomously instead of just executing predefined programmed instructions. This research presents such intelligent computer-aided diagnostic (CAD) system that tries to mimic the brain of a radiologist by learning how to distinguish between normal and Tuberculosis (TB) infected chest x-ray. Such systems can facilitate reducing TB epidemic as it is a curable disease and early diagnosis is a critical parameter of its cure. Proposed research consists of two main modules i.e. segmentation of lungs region and classification of x-ray as normal or TB infected. Segmentation of lungs from chest x-rays is a crucial step in designing of such CAD system and a fully convolutional neural network was used for this purpose. Post processing was done to fill holes, separate left and right lung from each other and remove unwanted objects that appeared in few cases. The process was repeated 10 times, with random split of data into a 60:40 ratio as training and testing sets respectively, to calculate the average accuracy. The segmentation methodology was tested on three datasets: Japanese Society of Radiological Technology (JSRT), Montgomery County (MC), and a local dataset that achieved average accuracy of 97.1%, 97.7% and 94.2% respectively. For TB classification purpose, an ensemble was created by feature-level fusion of three deep neural network models: ResNet, Inception-ResNet and DenseNet. The models were used as feature extractors and support vector machine (SVM) was used as a classifier. The methodology was tested on publically available Shenzhen dataset, which was randomly split into a 90:10 ratio as training and testing set respectively. The process was repeated 10 times with random split data, which achieved 90.5% average accuracy that is among top accuracies achieved till date on Shenzhen dataset. The results proved that the proposed methodologies are efficient enough and can be generalized for other such segmentation and classification problems in medical imaging domain.