Self-Trained Convolutional Neural Network (CNN) for Tuberculosis Diagnosis in Medical Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Tuberculosis (TB) poses a significant global health challenge, necessitating early and accurate detection for effective treatment and transmission control.
• Chest X-rays are a primary diagnostic tool for TB, but interpretation can be complex and time-consuming.
• Advancements in deep learning offer potential for enhancing diagnostic accuracy and efficiency in medical imaging.

Purpose of the Study:

• To develop and evaluate a deep learning model, specifically a convolutional neural network (CNN), for the automated detection of tuberculosis from chest X-ray images.
• To investigate the effectiveness of various image preprocessing techniques in improving the performance of the CNN model for TB detection.
• To compare the performance of a self-trained CNN model against transfer learning approaches using pre-trained models.

Main Methods:

• A dataset of 7000 chest X-ray images (3500 TB positive, 3500 normal) was utilized.
• Image preprocessing included wavelet transformation, contrast-limited adaptive histogram equalization (CLAHE), and gamma correction.
• Data augmentation techniques such as random flipping, rotation, resizing, and rescaling were applied to enhance model robustness.
• A CNN architecture comprising convolutional, max-pooling, flatten, and dense layers was designed, with ReLU and sigmoid activations.

Main Results:

• The self-trained CNN model achieved a high accuracy of approximately 96.57% in detecting tuberculosis from chest X-ray images.
• The deep learning approach demonstrated superior performance compared to transfer learning methods using pre-trained models.
• Image preprocessing and data augmentation techniques contributed to improved model performance and robustness.

Conclusions:

• Deep learning, particularly CNNs, shows significant promise for accurate and efficient tuberculosis detection in chest X-ray analysis.
• Self-trained CNN models can be effectively optimized for medical image analysis tasks, potentially outperforming traditional transfer learning.
• Future research should focus on larger, diverse datasets from clinical settings to further validate and refine these AI-driven diagnostic tools for broader medical application.