Uncertainty Assisted Robust Tuberculosis Identification With Bayesian Convolutional Neural Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Tuberculosis (TB) detection relies on identifying complex features in chest X-rays (CXRs).
• Standard Convolutional Neural Networks (CNNs) struggle with uncertainty in CXR classification, impacting diagnostic accuracy.
• Differentiating between TB and non-TB cases with subtle manifestations remains a challenge.

Purpose of the Study:

• To develop a Bayesian-based CNN (B-CNN) model for enhanced TB identification from CXRs.
• To address the limitations of traditional CNNs in handling uncertainty and ambiguous cases.
• To improve the accuracy and reliability of automated TB detection systems.

Main Methods:

• Implementation of a Bayesian-based convolutional neural network (B-CNN) architecture.
• Training and evaluation on two benchmark TB datasets: Montgomery and Shenzhen.
• Utilizing Google Colab with NVidia Tesla K80 for computational resources.

Main Results:

• B-CNN achieved high accuracy: 96.42% on the Montgomery dataset and 86.46% on the Shenzhen dataset.
• The model effectively identified and filtered ambiguous CXR cases based on output variance.
• Demonstrated superior performance compared to state-of-the-art machine learning and standard CNN approaches.

Conclusions:

• B-CNN offers a robust solution for TB detection, outperforming existing methods.
• The model's ability to quantify uncertainty enhances diagnostic reliability.
• B-CNN shows significant potential for improving automated analysis of medical images for infectious diseases.