Revisiting Computer-Aided Tuberculosis Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Tuberculosis (TB) poses a significant global health challenge, with diagnosis often difficult in resource-limited settings.
• Deep learning for computer-aided tuberculosis diagnosis (CTD) shows promise but is hampered by insufficient training data.
• Accurate and timely TB diagnosis is crucial for improving patient survival rates and controlling disease spread.

Purpose of the Study:

• To establish a large-scale chest X-ray (CXR) dataset (TBX11K) for training sophisticated CTD models.
• To propose SymFormer, a novel deep learning model for simultaneous CXR classification and TB infection detection.
• To create a benchmark for CTD research, including evaluation metrics and an online challenge.

Main Methods:

• Development of the TBX11K dataset with 11,200 CXR images and TB area annotations.
• Introduction of SymFormer, utilizing Symmetric Search Attention (SymAttention) and Symmetric Positional Encoding (SPE) for enhanced feature learning.
• Establishment of a CTD benchmark with evaluation metrics and baseline model assessments.

Main Results:

• The TBX11K dataset provides a valuable resource for training high-quality CTD detectors.
• SymFormer demonstrated state-of-the-art performance in classifying CXR images and detecting TB infection areas.
• The proposed methods effectively leverage the bilateral symmetry of CXR images for improved diagnostic accuracy.

Conclusions:

• The TBX11K dataset and SymFormer model represent significant advancements in computer-aided tuberculosis diagnosis.
• This work facilitates further research and development in deep learning for medical image analysis in TB detection.
• The established benchmark and online challenge will accelerate progress in CTD research globally.