Promptly identification of multidrug-resistant tuberculosis (MDR-TB) patients at high risk of treatment failure is essential for improving cure rates. This study aimed to develop a habitat radiomics based transformer fusion model to assess treatment effectiveness of MDR-TB. Independent patient cohorts from two hospitals were included. Radiomics features were extracted from the habitat and peripheral regions of cavities to construct predictive models. Then, a transformer-based fusion model integrating features from all regions was established. The areas under the receiver operating characteristic curves (AUCs) were used to evaluate the performance. The transformer fusion model combining two subregions and peripheral area achieved remarkable performance, with AUC values of 1.000, 0.959, and 0.879 in the training, validation, and test cohort, respectively. The finding highlights the efficacy of our model in predicting treatment effectiveness of MDR-TB patients and its potential to guide individualized therapy.