Refining dataset curation methods for deep learning-based automated tuberculosis screening.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning systems (DLS) require large labeled datasets for tuberculosis (TB) detection on chest radiographs (CXRs).
• Semi-supervised learning offers a potential solution to reduce reliance on extensive manual data curation.

Purpose of the Study:

• To evaluate the efficacy of using unlabeled chest radiographs to train and enhance a deep learning system for TB detection.
• To assess the performance of a two-stage semi-supervised approach for improving TB detection accuracy on CXRs.

Main Methods:

• A two-stage semi-supervised learning approach was employed using the National Institute of Health ChestX-ray14 database (111,622 CXRs).
• Phase I involved training a DCNN on 11,000 expert-labeled CXRs. Phase II utilized the Phase I model to label remaining CXRs, followed by retraining a second DCNN (TBNet).
• TBNet was validated on CXRs from independent US and Chinese sites with confirmed TB cases.

Main Results:

• The Phase I model achieved an AUC of 0.88. The Phase II model (TBNet) reached an AUC of 0.91 on a mixed dataset and 0.87 on a tertiary care hospital dataset.
• TBNet demonstrated high sensitivity (85%), specificity (76%), and negative predictive value (0.9).
• When used with radiologists, TBNet improved overall sensitivity to 94% and negative predictive value to 0.96, indicating a synergistic effect.

Conclusions:

• Semi-supervised learning enables the development of accurate deep learning models for TB detection using limited labeled data.
• The developed TBNet functions as a valuable computer-aided detection (CAD) tool, particularly in identifying findings missed by radiologists.
• This approach is crucial for building clinically relevant AI models in the era of precision medicine, especially when dealing with noisy or absent labels.