Sputum smear images database: A resource for deep learning study based to detect Bacilli for TB diagnose

In Indonesia, sputum smear microscopy is one of method that widely adopted to detect Tuberculosis Bacilli, a bacteria causing Tuberculosis (TB). Due to sputum smear observation using conventional microscope rely on very little cost to start with, compared to its more modern solution. i.e. using fluorescence microscope or using GeneXpert. However, sputum smear observation using conventional microscope has drawbacks. According to the World Health Organization (WHO) standard, each sample is need to observed in 300 field of view. It leads to operator fatigueness, which can affect the output. Another factor is, it has very high inter operator variability. Therefore, an automatic process to detect the TB bacteria is highly needed. Deep learning based method to detect the bacteria is gaining attention in recent years. However, these methods rely on database, which is not designed for deep learning approach. The database contain limited number of annotated data. Moreover, the available database annotated images on a clean smear that completely different with Indonesian sputum smear images. This paper propose a sputum smear images database with data collected by Indonesian experts. Our database contain of 3 classes, which are: TBBacteria, NonTBbacteria, and Stainresidues. The database assists researcher for developing deep learning based TB bacteria detection, to produce deep learning's architecture with better accuracy.