Explainable deep-neural-network supported scheme for tuberculosis detection from chest radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Tuberculosis diagnosis relies on chest radiographs, which are time-consuming and subjective.
• Machine learning offers potential for improving medical diagnostics, including tuberculosis screening.

Purpose of the Study:

• To develop a shallow convolutional neural network (CNN) for efficient and accurate tuberculosis screening from chest X-rays.
• To enhance diagnostic interpretation and reduce subjectivity in tuberculosis detection.

Main Methods:

• A shallow CNN with four convolution-maxpooling layers was designed.
• Hyperparameters were optimized using Bayesian optimization.
• Model performance was evaluated using accuracy, F1-score, sensitivity, specificity, and ROC AUC.
• Explainability was assessed using Class Activation Maps (CAM) and Local Interpretable Model-agnostic Explanations (LIME).

Main Results:

• The shallow CNN achieved a peak classification accuracy, F1-score, sensitivity, and specificity of 0.95.
• The receiver operating characteristic (ROC) curve demonstrated a peak area under the curve (AUC) of 0.976.
• The model's transparency and explainability were assessed against a state-of-the-art DenseNet.

Conclusions:

• The developed shallow CNN provides a highly accurate and potentially more objective method for tuberculosis screening from chest X-rays.
• The model's explainability features contribute to its clinical utility and trustworthiness.
• This approach offers a promising alternative to conventional diagnostic methods, addressing limitations of time and subjectivity.