Hybrid Spatial Temporal Deep Learning for Early Tuberculosis Detection and Risk Stratification

Abstract

Tuberculosis (TB) continues to be a major global health problem, and is responsible for more than 10.6 million new cases each year. Early and accurate detection and automation of disease progression is critical in controlling the disease process, especially in resource-limited environments. In this study, a novel Attention-Fused Hybrid Deep Learning Framework is proposed that combines the EfficientNetB0 and DenseNet121 models to complement each other in spatial feature extraction, a CNN-GRU spatio-temporal representation learning module, and a channel-wise attention fusion module for adaptive multi-source feature weighting. The framework is able to detect TB simultaneously and stratify the risk of the patient at three levels (Low, Medium, High) based on chest X-ray images. The proposed model shows a classification accuracy of 98.87%, a precision of 98.72%, recall of 98.65% and F1-Score of 98.68% on a subset of 7,000 images (3,500 TB+ve and 3,500 TB-ve) which is significantly higher than the baseline models, such as standalone CNN, GRU, CNN-LSTM, CNN-GRU, Logistic Regression, SVM, and Random Forest. In ablation studies, all the architectural components are confirmed and it has been established that each one of them contributes to the increment. An explainability analysis via Grad-CAM shows that the focus areas are clinically interpretable regions, corresponding to pulmonary zones of clinical importance. This proposed framework shows excellent generalizability, efficiency of computation, and clinical interpretability, which has become a new benchmark for intelligent tuberculosis screening and clinically interpretable clinical decision support.