BME 590L Final Project: Classification of Microscopic Images With and Without Tuberculosis Bacilli

Zhen Lin      Huisi Cai     

zl187@duke.edu     hc239@duke.edu    

Paper PDF

About 1,700 million people, or one third of the world's population are, or have been, infected by Mycobacterium tuberculosis (TB). The process of examining the microscopic images could be time consuming and labor-intensive. Therefore, we propose to classify whether the tuberculosis is presented in the sample automatically using deep neural network training. The goal of this project is to train the network to correctly classify the microscopic specimen as "with TB cell" or "without TB cell". There are three main components in this project: 1) pre-process the image database by cropping them into suitable size for training and labeling them correctly, 2) perform image classification on the self-established dataset with different networks and identify a network with the best performance, and 3) Add a physical layer of weights on the three color channels to improve classification accuracy.

Relevant links: Artificial Intelligence and Data Science, Makerere University, Uganda


Paper:
Code and Data: