Proposing a novel multi-instance learning model for tuberculosis recognition from chest X-ray images based on CNNs,...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Infectious Disease Diagnostics

Background:

• Tuberculosis (TB) is a significant global infectious disease, with 10 million diagnoses in 2018.
• Early TB diagnosis is crucial for effective treatment, survival rates, and transmission prevention.
• Chest radiographs (CXR) offer a low-cost, rapid screening method for TB, but interpretation requires expertise due to image complexity.

Purpose of the Study:

• To develop and evaluate a novel multi-instance classification model for automated tuberculosis recognition from CXR images.
• To reduce the disease burden of tuberculosis through improved diagnostic accuracy and efficiency.
• To create a computer-aided diagnosis (CADx) system for TB detection, minimizing reliance on specialist interpretation.

Main Methods:

• A novel CNNs, complex networks, and stacked ensemble (CCNSE) model was proposed for TB detection.
• The CCNSE model utilizes overlapping patches from CXR images, extracting features via CNNs and clustering.
• Feature engineering on local and global complex networks, followed by global clustering and purity score assignment, enables patch-level and image-level analysis for classification.

Main Results:

• The proposed CCNSE method achieved high performance on two datasets (Montgomery County CXR and Shenzhen).
• Achieved AUC scores of 99.00% ± 0.28% for MC and 98.00% ± 0.16% for SZ.
• Demonstrated accuracy of 99.26% ± 0.40% for MC and 99.22% ± 0.32% for SZ, outperforming comparative methods.

Conclusions:

• The CCNSE model shows superior performance in diagnosing tuberculosis from CXR images.
• This automated approach can significantly reduce manual time, effort, and dependency on expert radiologists.
• The developed system holds promise as an effective computer-aided diagnosis tool for widespread TB screening.