MITIGATING CLASS IMBALANCE IN TUBERCULOSIS DETECTION: COMBINING SMOTE AND TOMEK LINK WITH MODIFIED FOCAL LOSS AND CLASS WEIGHTING IN A TRANSFER LEARNING FRAMEWORK

Abstract

Tuberculosis (TB) remains a major global health concern, and early detection is crucial for effective treatment. This study addresses the challenge of class imbalance in existing machine learning models for TB prediction, which often leads to biased results. A novel approach combining the hybrid SMOTE-Tomek Links technique, modified focal loss, and class weighting was developed and applied to a dataset of X-ray images categorized into normal and TB classes. The hybrid SMOTE-Tomek Links method generates synthetic samples for the minority class while removing ambiguous samples, ensuring a balanced dataset. The modified focal loss and class weighting help focus on misclassified cases and address class disparities. The model was evaluated against benchmark models, including EfficientNetB3, Random Forest, and XGBoost, with and without SMOTE. The developed model achieved a remarkable accuracy of 99.7%, outperforming the benchmark models (92.72%–99.1%). These results demonstrate the effectiveness of the proposed approach in improving TB prediction accuracy and handling class imbalance. The study's findings provide valuable insights into medical image classification and offer a robust framework for enhancing diagnostic tools, with potential applications beyond TB detection. This research could significantly improve TB management and diagnosis in clinical settings.