Detection of Mycobacterium Tuberculosis Using Residual Neural Network

In 2020, even though the number of people with TB has decreased by 1.3 million. However, TB deaths have increased by 0.1 million. There are two main problems in Mycobacterium Tuberculosis (MT) detection to diagnose tuberculosis, i.e. the overload and the detection consistency. To overcome these problems, this study explores the use of a residual neural network (ResNet) in detecting MT. In total, 20,240 patch images are applied to build the model that consists of 6,583 MTs and 13,657 non-MTs. Image dimension of $320 \times 320$ pixels and train with the ResNet 152 v1 gives the best model’s performance. This proposed scheme has an accuracy of 0.8592 and a loss of 0.3320. However, it requires about 22 hours to build the model and the data storage of about 238 MB. Even though this model is not the optimal one, however, in the medical application the best detection performance is the most critical parameter than the other ones. The results of this study can assist MT detection by reducing the workload. Furthermore, the proposed system gives the detection result more consistent.