Diagnostic assistance method for RR-TB/MDR-TB patients under treatment based on CNN-LSTM.
Authors
Affiliations (8)
Affiliations (8)
- Department of Tuberculosis, Jiangxi Chest Hospital, The Third Affiliated Hospital of Nanchang Medical College, Nanchang, 330006, People's Republic of China.
- Medical Affairs, Johnson & Johnson, Somerset, USA.
- Department of Infectious Disease and Public Health, Affiliated Hospital of Hunan University, Central Hospital of Xiangtan, Xiangtan, People's Republic of China.
- Department of Infection, Shangrao Second People's Hospital, Shangrao, People's Republic of China.
- Department of Tuberculosis, Jiujiang Third People's Hospital, Jiujiang, People's Republic of China.
- Department of Radiology, Jiangxi Chest Hospital, The Third Affiliated Hospital of Nanchang Medical College, Nanchang, 330006, People's Republic of China.
- Department of Thoracic Surgery, Jiangxi Chest Hospital, The Third Affiliated Hospital of Nanchang Medical College, Nanchang, 330006, People's Republic of China. [email protected].
- Department of Neurology, Jiangxi Chest Hospital, The Third Affiliated Hospital of Nanchang Medical College, Nanchang, 330006, People's Republic of China. [email protected].
Abstract
The rapid development of deep learning has promoted its application in disease diagnosis, treatment, and prognosis prediction. Medical imaging plays a crucial role in the management of rifampicin-resistant tuberculosis/multidrug-resistant tuberculosis (RR-TB/MDR-TB). In particular, chest computed tomography (CT) scans offer detailed lung images that can reveal subtle features. In this study, we propose a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model to predict treatment outcomes in RR-TB/MDR-TB patients, aiming to support clinicians in timely adjustment of therapeutic strategies and improving treatment success. The model integrates CNN for image feature extraction with LSTM for sequential analysis of patient monitoring data, including two types of immune detection indicators (T-cell subsets and peripheral blood tuberculosis-related CD161-positive cells). Transfer learning with weight initialization was applied to enhance model performance, and three backbone architectures (DenseNet201 + ABC, ResNet-50 + ABC, CheXNet + ABC) were compared to assess their impact on predictive accuracy. Experimental results demonstrated that the CNN-LSTM model with DenseNet201 + ABC as the backbone achieved the highest accuracy in predicting subsequent treatment indicators. These findings demonstrated the feasibility and effectiveness of using CNN-LSTM for treatment outcome prediction in RR-TB/MDR-TB, and highlighted its potential to assist clinicians in precision-tailoring treatment plans, thereby improving therapeutic efficacy and offering both theoretical and practical value in healthcare.