Automatic Detection of Tuberculosis bacilli from Conventional Sputum Smear Microscopic Images Using Densely Connected Convolutional Networks

Early and correct diagnosis of Tuberculosis (TB) is important for proper treatment. According to WHO 2020 report, about 10 million people fell ill with TB and 1.4 million died in 2019. There are a number of well-established diagnosing methods for detecting TB, among them sputum smear microscopy examination remains the popular diagnostic test especially among the high burden countries due to its low cost and its ability to detect highly infectious pulmonary TB. However, the manual sputum smear microscopic examination is time-consuming and error-prone. Recently, many methods were proposed in the literature to automate the sputum smear microscopic examination. In this paper, we used the DenseNet CNN model for automating the detection of TB bacilli from non-bacilli objects. We used images from two public datasets for the development and evaluation of the model. We used 2220 samples from both datasets, which include 1110 positive cases (TB bacilli) and 1110 negative cases (non-bacilli). To avoid any bias towards the test set, we conducted a fivefold cross validation and our method achieved an average accuracy of 99.7%, which is superior to other recently proposed methods.