Screening TB Using Deep Transfer Learning

Tuberculosis is a major public health problem and has to be proactive screening for quarantine by means of developing machine learning model to screen suspected case. This can be mutually beneficial to healthcare providers and patients. The application of deep learning technique for medical image classification has been developed and grown exponentially over the past few years. We propose Convolution Neural Network (CNN) model because it is one of several well-known and high performance models for image classification. This research presents neural network to classify chest imaging into 2 classes: normal and tuberculosis. We collect 3 datasets of chest X-ray image: Montgomery, Shenzen and Bureau of tuberculosis. The researchers compared 4 CNN classification models to find out the best model that is suitable for chest X-ray. Performance was measured by using metrics: accuracy, precision, recall and AUC. The result of this study shows that DenseNet model is more accurate than others and we tune the model for the best threshold and train it with Thai Bureau of tuberculosis chest image for screening TB for Thai people. The accuracy for discrimination normal lung and TB-infected lung in the best model is 91% and AUC is 95%. This model would aided healthcare providers for TB screening large population in Thailand.