AUTHOR=Sun Xin , Xing Zhiheng , Wan Zhen , Ding Wenlong , Wang Li , Zhong Lingshan , Zhou Xinran , Gong Xiu-Jun , Li Yonghui , Zhang Xiao-Dong 

TITLE=A robust ensemble deep learning framework for accurate diagnoses of tuberculosis from chest radiographs

JOURNAL=Frontiers in Medicine

VOLUME=Volume 11 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2024.1391184

DOI=10.3389/fmed.2024.1391184

ISSN=2296-858X

ABSTRACT=Tuberculosis (TB) is one of the leading causes of death globally, and early diagnosis with cost-effective medical resources is key to building a sustainable global TB response. For more precise resource allocation, chest X-ray radiography (CXR) data with automated screening can be at the forefront of the TB response. In this study, a robust ensemble deep learning framework with 43 implementations (models) is investigated to label active TB and its morphological and pathological subtypes. Based on our dataset with subtype details, the capability for detailed diagnoses is enabled. The deep learning framework contains a feature extraction stage, a fusion stage, and a classification stage. The feature extraction phase consists of 13 pre-trained convolutional neural network (CNN) models, leaving only the convolutional layers. The feature set extracted from the convolutional layer is then processed by the fusion phase using one of three strategies: voting, attention, and concatenation. The classification phase contains five binary classifiers that process fusion outcomes and predict robust clinical outcomes. The model is trained on a manually compiled deidentified dataset containing 915 active TB patients and 1276 healthy individuals. In addition, the efficiency of the model is validated by the visualization representations for further evaluating the performance of the ensemble model under different strategies, thereby helping clinicians screen and classify TB.