CViTM-Net: A Hybrid Deep Learning Method for Tuberculosis Detection in CT Images

Tuberculosis (TB) persists as a substantial global public health concern, with a considerable number of new cases documented annually. Conventional diagnostic methodologies encounter constraints, including diminished sensitivity and elevated expenses. This study introduces a Convolutional neural network (CNN)-ViT-Meta fusion model (CViTM-Net) that combines CNN's local feature extraction, ViT's global contextual modeling, and patient metadata (gender and age) for multi-level feature integration. The model utilizes U-Net for precise lung segmentation and designs a CNN model to infer missing metadata, markedly enhancing dataset quality. Experiments on three publicly available datasets (4,999 samples) demonstrate an overall accuracy of 92.2% and specificity of 95.3%, with robust performance across age groups and gender demographics. Ablation studies substantiate the indispensability of each module, with the elimination of CNN resulting in a 20.6% performance reduction.