A voting-based ensemble deep learning method focusing on image augmentation and preprocessing variations for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Tuberculosis (TB) is a significant global infectious disease primarily affecting the lungs.
• Early detection and treatment are crucial to reduce mortality and transmission.
• Chest radiography (CXR) is a common imaging technique, and Computer-Aided Detection (CADe) systems enhance diagnostic reliability.

Purpose of the Study:

• To develop and evaluate an ensemble Convolutional Neural Network (CNN) model for improved tuberculosis detection using CXR images.
• To investigate the impact of preprocessing variations and data augmentation on model performance.

Main Methods:

• Utilized fine-tuned InceptionV3 and Xception CNN models with Contrast-Limited Adaptive Histogram Equalization (CLAHE) preprocessing.
• Applied 10 different image transformations for data augmentation.
• Implemented ensemble methods, including Bayesian optimization-based weighted voting and soft voting with averaging probabilities, using 3-5 base classifiers.

Main Results:

• Achieved high accuracy rates of 97.500% on the Montgomery dataset and 97.699% on the Shenzhen dataset.
• The proposed ensemble CNN model demonstrated superior performance compared to state-of-the-art methods on both TB CXR datasets.

Conclusions:

• The developed ensemble CNN model offers a highly accurate and efficient approach for automated tuberculosis detection from chest X-rays.
• This method has the potential to significantly aid clinicians in early TB diagnosis, thereby improving patient management and public health outcomes.