Modality-specific deep learning model ensembles toward improving TB detection in chest radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Tuberculosis (TB) detection from chest X-rays (CXRs) remains a critical diagnostic challenge.
• Current deep learning models often lack robustness and generalization capabilities for TB detection.

Purpose of the Study:

• To evaluate the efficacy of knowledge transfer from modality-specific deep learning models for improving TB detection.
• To develop an ensemble model that surpasses the state-of-the-art in classifying TB-infected and normal CXRs.

Main Methods:

• Training custom and pre-trained Convolutional Neural Networks (CNNs) on large-scale CXR datasets (RSNA, Pediatric pneumonia, Indiana).
• Transferring learned knowledge to fine-tune models for TB detection on the Shenzhen CXR collection.
• Combining predictions of top-performing models using ensemble methods and evaluating through 5-fold patient-level cross-validation.

Main Results:

• A stacked ensemble of the top-3 retrained models achieved high performance (accuracy: 0.941, AUC: 0.995).
• No statistically significant differences were observed among ensemble methods (ANOVA, P > 0.05).
• Ensemble model demonstrated reduced prediction variance and improved robustness compared to individual models.

Conclusions:

• Knowledge transfer through modality-specific deep learning significantly improves TB classification accuracy.
• Ensemble methods provide superior and more robust performance for TB detection in CXRs.
• The proposed approach advances the state-of-the-art in automated TB diagnosis.