A Comparative Study On Performance Of Pre-Trained Convolutional Neural Networks In Tuberculosis Detection
European Journal of Molecular & Clinical Medicine,
2020, Volume 7, Issue 3, Pages 4852-4858
AbstractIndia accounts for 26% of the words Tuberculosis population. The WHO’s Global TB Program states that in India, the number of people newly diagnosed with TB increased by 74% when compared to other countries from 1.2 million to 2.2 million between 2013 and 2019. Tuberculosis was and still remains a disease that causes high death rates in the country. Many of these deaths can be easily prevented if diagnosed at an early stage. The easiest, cost-effective and non-invasive method of detecting tuberculosis is through a frontal chest x-ray (CXR). But this requires a radiologist to manually examine and analyse each of the X-ray, considering the heavy patient count this puts a great burden on the resources available. A computer aided diagnosis system can easily mitigate this problem and can greatly help in reducing the cost. In recent times deep learning has made great progress in the field of image classification and has produced remarkable outputs in terms of image classification in various domains. But there still remains a scope for improvements when it comes to Tuberculosis detection. The aim of this study is toapply three pre-trained convolutional neural networks that have proven record in image classification on to publically available CXR dataset and classify CXR’s that manifest tuberculosis and compare their performances. The CNN models that are used on our CXR images dataset as a part of this study are VGG-16 ,VGG-19,AlexNet ,Xception and ResNet-50. Also visualization techniques have been applied to help understand the features whose weights played a role in the classification process. With the help of this system, we can easily classify CXR’s that have active TB and even CXR’s that show mild abnormalities, thus ensuring that high risk patients get the help they require on time.
- Article View: 134
- PDF Download: 161