Comparison of ResNet101V2 and ResNet152V2 Architectures in Microscopy-Based Tuberculosis Bacteria Identification
Aeri Rachmad *, Mohammad Syarief, Suci Hernawati, Eka Mala Sari Rochman, Husni, Kurniawan Eka Permana

University of Trunojoyo Madura

Abstract

Tuberculosis (TB) is a preventable and treatable infectious disease, but it remains a serious problem in countries at risk, such as those with poverty and limited access to healthcare services. Caused by the bacterium Mycobacterium tuberculosis, TB can be fatal without proper treatment. Accurate early identification is challenging, despite prevention efforts being made. The primary method for detecting TB is by identifying bacteria in sputum samples using a microscope, but there are weaknesses such as varying interpretations and inconsistent image quality. Convolutional Neural Networks (CNN) have shown potential in improving the accuracy of identifying TB bacteria in microscopic images. This study compares the performance of two CNN architectures, ResNet101V2 and ResNet152V2, in identifying TB bacteria in microscopic images. ResNet152V2 shows better results with an accuracy of 83.86%, precision of 100.00%, recall of 66.39%, and an F1-score of 80.00%. Despite requiring longer computational time, efficiency remains high, demonstrating strong potential for medical applications. Future research can explore variations in architecture and parameters for even more optimal results.

Keywords: Tuberculosis, Convolutional Neural Networks, ResNet101V2 and ResNet152V2

Topic: Machine Learning and Deep Learning