According to the Ministry of Health in Kenya, tuberculosis (TB) is now the largest infectious disease killer in Kenya (USAID) and the fifth greatest cause of death (Ministry of Health, 2020). TB continues to harm numerous vulnerable groups, homes, and communities in Kenya and across Africa, despite the fact that it is preventable and treatable.
The conventional TB tests, which include a skin test and a blood test, are insufficient for determining the kind of TB. The World Health Organization (WHO) advises that chest X-rays (chest radiography) be used for more screening (WHO,2021). However, in many TB-affected areas of Kenya, there is a scarcity of radiological interpretation skills, which may impede screening and follow-up activities. In Kenya, a successful automated and cost-effective technology could enhance screening evaluation efforts and allow for earlier disease detection. As a result, there has been a lot of interest in utilizing computer-aided diagnosis to diagnose tuberculosis using chest X-rays, and there have been a lot of different ways.
Computer assisted diagnostics solutions have mostly relied on traditional machine learning approaches. Deep learning has grown rapidly over the previous decade, attracting a lot of attention in the field of image processing and classification. A neural network with a number of hidden layers is known as deep learning. Deep learning approaches can extract characteristics automatically without the need for human interaction, and they have shown excellent outcomes for real-world issues. Deep learning has recently gotten a lot of attention as a way to create a system that is autonomous, fast, and accurate.
However, training of these deep networks is challenging since the computation cost associated with it is exceptionally expensive. It is still computationally expensive to train, in that, it takes a long time to train the network even on a small dataset. Usually, the training time is measured in terms of days and at times weeks. Therefore, this limits the realization of deep learning. Further, the demand to utilize lots of data slows down the advancement and the viability of exploiting deep learning. Additionally, high inference time of the trained model limits its realization. To reduce the time that it takes to develop deep learning models with eminent precision there is a bid to lessen the time linked to the training of deep learning networks. Furthermore, the problem of overfitting due the depth of the network is still challenging the implementation of deep learning in real world problems. Moreover, practical applications in real-world problems are still lacking since more work is focused on improving the networks using some benchmarking datasets.
This study aims to optimize the efficiency of deep learning networks for the task of detecting TB using X-rays. In this work, we propose methods for improving the computation cost of deep learning methods for the task of detecting TB.
- Dr Edna Too, Chuka University
- Dr David Mwathi, Chuka University
- Prof Lucy Gitonga, Chuka University
- Saif Kanyori, Chuka University
- Pauline Mwaka, Chuka University