Implementasi Arsitektur ResNet18 dalam Deteksi Tuberkulosis melalui Citra X-Ray dengan Evaluasi K-Fold Cross Validation

Tuberculosis (TB) remains a serious global health threat, particularly in developing countries. Early detection is crucial to curb its spread. however, conventional methods like culture are time-consuming and rely on expert personnel. This study develops a deep learning-based TB detection model using the ResNet18 architecture, employing 5-fold cross-validation for rigorous evaluation and Grad-CAM for visual interpretation. The dataset comprises 7,000 balanced X-ray images, with 3,500 images each for normal and tuberculosis classes. Evaluation results demonstrate that the model achieved an accuracy of 99.57% on the test data, with an average validation accuracy of 99.47% across the five folds. Furthermore, the model's precision reached 99.71% and recall was 99.43%. Grad-CAM successfully visualizes how the model recognizes tuberculosis (TB) patterns and makes predictions based on heatmaps, serving as the basis for its decisions. These findings indicate that the ResNet18 architecture with 5-fold cross-validation is effective and consistent in detecting TB, as well as being computationally efficient without requiring lung segmentation. This research contributes to the classification of tuberculosis (TB) through X-ray images by combining an efficient architecture and visual interpretability, applied to a balanced dataset. The study's limitations include a restricted dataset variation, thus suggesting that future research should encompass more diverse demographic data.