Predicting Tuberculosis Drug Resistance Using Semi-Supervised and Multi-Task Learning

Tuberculosis remains a persistent global health threat and a leading cause of mortality. Drug therapy currently serves as the primary approach for tuberculosis treatment. However, drug-resistant cases have emerged among tuberculosis patients. Accurately identifying drug resistance in patients is paramount for effective treatment. A practical artificial intel-ligence algorithm relies heavily on a substantial volume of meticulously labeled data. Unfortunately, there is a scarcity of cases for drug resistance analysis in clinical settings. To address the challenge of small data, we propose a semi-supervised learning approach for predicting tuberculosis drug resistance and leverage tuberculosis morphological features for multi-task learning. Firstly, we employ a three-dimensional ResNet as the backbone. Subsequently, the practical training procedure is executed through the semi-supervised learning algorithm based on pseudo-labeling and the multi-tasking learning strategy. Our proposed method demonstrates promising results on the constructed semi-supervised CT image dataset.