External Validation of Deep Learning Algorithm for Tuberculosis Detection in Thai Population
Issued Date
2022-01-01
Resource Type
Scopus ID
2-s2.0-85151704542
Journal Title
6th International Conference on Information Technology, InCIT 2022
Start Page
314
End Page
319
Rights Holder(s)
SCOPUS
Bibliographic Citation
6th International Conference on Information Technology, InCIT 2022 (2022) , 314-319
Suggested Citation
Rajak A., Chaisangmongkon W., Chamveha I., Promwiset T., Rungsinaporn K., Saiviroonporn P., Tongdee T. External Validation of Deep Learning Algorithm for Tuberculosis Detection in Thai Population. 6th International Conference on Information Technology, InCIT 2022 (2022) , 314-319. 319. doi:10.1109/InCIT56086.2022.10067327 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/84298
Title
External Validation of Deep Learning Algorithm for Tuberculosis Detection in Thai Population
Other Contributor(s)
Abstract
Several studies have been conducted for the automatic detection of tuberculosis on chest X-ray (CXR) images using deep learning. Despite the excellent performance of deep learning algorithms, a major challenge faced by such models is its limited ability to generalize in unseen datasets. Previous works have highlighted the importance of local datasets for building a high-performance deep learning model tailored to a specific region or population, yet model's performance on heterogeneous datasets have not been addressed. In this paper, we present a state-of-the-art model for image-wise classification and lesionwise localization of tuberculosis (TB) in the Thai population. The model was trained on an extensive Thai CXR dataset, which was labeled with feature-specific keywords. Our model demonstrated outstanding performance with an average AUROC of 0.936 and a lesion-wise localization score of 88.18%. The model achieved high sensitivity (83.5%) and specificity (94.6%). When compared with the benchmark model based on EfficientNet, our model obtained excellent performance in terms of both classification and localization. Our model consistently outperformed the benchmark model when validated on multiple independent datasets.