Comparative Study of AI Models for Automated Tuberculosis Detection Using Image Processing Techniques

Tuberculosis is a critical global health issue, particularly in resource-limited regions where early and accurate diagnosis is important and is in need so that the treatment is effective and the control transmission is controlled. The known diagnostic methods, such as sputum smear microscopy and nucleic acid amplification tests are costly, time-consuming, and require trained professionals. Due this in some cases it is inaccessible in many regions. Deep learning-based automated TB detection offers a promising alternative by enhancing diagnostic efficiency through medical imaging analysis. This study presents a comparative evaluation of five deep learning models, InceptionResNetV2, DenseNet, VGG16, ANN, and a custom CNN, trained on a dataset of $\mathbf{3, 0 0 8}$ chest radiograph images, evenly distributed between TB-positive and normal cases. The dataset underwent advanced preprocessing techniques, pixel normalization, and data augmentation. The hyperparameter tuning process was applied, which optimized the learning rates, dropout rates, convolutional filter sizes, and batch sizes to enhance model performance. The models were assessed using accuracy, precision, recall, F1-score, sensitivity, specificity.. Experimental results indicated that the custom CNN achieved the highest classification accuracy (99.51). The superior performance of the custom CNN over other models is attributed to optimized feature extraction, effective preprocessing, and structured hyperparameter tuning. A comparative analysis with previous studies highlights how this approach mitigates dataset limitations and improves model interpretability, and the potential of AI-driven TB detection, enhancing future diagnostic efficiency by improving model generalizability and deployment in real-world healthcare settings.