Diagnosing tuberculosis from X-ray imaging based on contrast limited adaptive histogram equalization

Abstract

Tuberculosis is a serious threat, and one of the effective data types for diagnosing tuberculosis is chest X-ray data. In this paper, we hypothesize the effect of image enhancement on the effectiveness of deep learning models in the problem of diagnosing pulmonary tuberculosis from chest X-ray images. To clarify the hypothesis, we have designed a data processing process with an image enhancement step using the contrast limited adaptive histogram equalization (CLAHE) technique to enhance the quality of input chest X-ray data, and the experiments were conducted with a standard dataset that was published on the Kaggle system. The evaluation is performed comprehensively with popular convolutional neural network architectures, including DenseNet201, DenseNet121, EfficientNetB0, and MobileNetV2, compared in two scenarios with and without the image enhancement step. Experiments have shown that the image enhancement step effectively improves the classification performance of all models, clearly through important scores such as area under curve (AUC), accuracy, F1-score, precision, and recall. The best result tested is the EfficientNetB0 model with 0.925926 accuracy score, 0.970732 AUC score, 0.904762 precision score, 0.95 recall score, and 0.926829 F1-score. In addition, qualitative analysis using gradient-weighted class activation mapping (Grad-CAM) shows that the resulting models have shown a focus on the lung region, reflecting the interpretability and suitability for radiologist expertise.