Mycobacterium tuberculosis images classification based on combining of convolutional neural network and support vector machine

Abstract

Mycobacterium Tuberculosis (TB bacteria) is a rod-shaped bacterium with a very small size. This bacterium can cause lung disease known as Tuberculosis. These TB bacteria can be seen at least by using a conventional microscope with magnification 1000 times. Images that have been seen in a microscope will be further processed by digital image processing. The data used in this study were 100 captions. Based on the color of the TB bacteria, a sputum image is detected and then cropping is done. Total data on TB bacteria and non-bacterial crops in automatic cropping were 1266 crops consisting of 633 TB bacteria and 633 non-TB bacteria. The size of the TB bacteria and open TB bacteria have different pixel sizes, so it needs to resize the image with a size of 50 x 50 pixels. There are several Convolutional Neural Networks (CNN) architectures that have been tried in solving classification problems among them LeNet, AlexNet, ZFNet, GoogleNet, VGGNet and ResNet. In other studies, the accuracy was 95.05% using the Inception V3 method. In the case of this classification of TB bacteria, researchers proposed the ResNet-101 architecture with 224x224x3 pixel input data specifications, 347layer and 1000 full connected layer (fc1000). As for the classification, researchers used the Support Vector Machine (SVM) to determine TB bacteria or not TB bacteria. The results of this study resulted in an accuracy of 97.6%, 97.9% precision, 97.4% recall and F1 score 97.6%.