Abstract: Pulmonary tuberculosis typically presents as multifocal and pleomorphic lesions on X-ray computed tomography (CT) images, necessitating a comprehensive assessment of multiple lesion types and quantities in clinical diagnosis. To address this complexity, this paper designs a pulmonary tuberculosis detection algorithm based on feature context understanding and slice analysis, aiming to achieve binary classification of tuberculosis detection. The algorithm integrates graph neural networks with multiple instance learning, constructing a detection framework comprising the following core modules: a feature-label embedding module that fuses lesion labels with image features to enhance the model's representational capacity for lesions; a semantic-spatial graph module that employs graph structures to simultaneously model both the spatial positional relationships and deep semantic associations among slices, effectively capturing multifocal distribution characteristics; and a global information attention module that leverages a multiple instance network to aggregate all instance information and learn the embedding representation of the entire bag, ultimately yielding the final classification result. Experimental results demonstrate that the proposed algorithm achieves outstanding performance with an accuracy of 96.3% and an area under the receiver operating characteristic curve (AUC) of 99.1%, significantly outperforming conventional 3D classification networks.