Automated Tuberculosis Detection from Chest X-Rays: A Deep Learning Approach with DenseNet121

In this paper presents the application of a modified DenseNet121 model for tuberculosis (TB) identification from chest X-ray images. Using 700 publicly available TB-positive images for the Tuberculosis Chest X-ray Database, 2800 more TB-positive images from the NIAID TB portal, and 3500 normal images, the model was tweaked for binary classification. Celebrated for its effective feature extraction, the DenseNet121 design was altered with sigmoid activation and a global average pooling layer. With an overall accuracy of 90%, the model was able to separate cases of normal from TB-positive. Performance criteria for normal images showed a precision of 0.90 and recall of 1.00, thereby indicating rather great accuracy in non-TB diagnosis. The recall for TB-positive cases was 0.43 even though the accuracy was exactly 1.00. Underlined the strong discriminative power of the model, the outstanding AUC of 0.976 indicated by the ROC curve analysis These results imply more work to increase TB-positive case identification and validate the promise of the DenseNet121 model in TB detection.