TBNet: learning from scratch and limited training data, a CNN based tuberculosis bacilli detection

Abstract

Tuberculosis (TB) is an infectious disease caused by the micro-bacteria. Several studies that have been conducted previously aimed to reduce the burden of observing tuberculosis bacilli using the digital image processing method. In this study, we proposed a newly developed convolutional neural network (CNN) based deep learning model to detect tuberculosis bacilli in sputum smear images. Recent advances in deep learning apply large scale image dataset to achieve convergent weight model. However, medical image dataset commonly available in relatively small quantity. In contrary with common deep learning approach, our model is capable to learn from our small dataset which consist of highly diverse hue and contrast of sputum smear images. Furthermore, its performance is proven to be reliable to detect sputum smear image content, which are TB bacillus and debris.