<p>Tuberculosis remains a highly fatal infectious disease worldwide. Chest X-ray imaging is a low-cost tool for the preliminary diagnosis of tuberculosis. However, most existing detection models have a large number of parameters and are computationally intensive, limiting their applicability in low-resource environments. This paper aims to develop computationally inexpensive models for efficient and rapid diagnosis of tuberculosis. We propose Dual-RepViT, a novel lightweight and efficient tuberculosis detection model. First, to reduce parameters, we streamline the RepViT block by removing its terminal convolutional layer and using it as the backbone of the proposed network. Subsequently, we introduce Haar wavelet transformer decomposition in parallel with the simplified RepViT block to retain essential frequency-domain information during feature extraction. Finally, a feature fusion block is utilized to fuse information from the wavelet branch and the backbone. The dual-branch architecture synergistically combines spatial features from the simplified RepViT branch with frequency-domain features from the wavelet decomposition branch. Compared to state-of-the-art approaches, our method achieves competitive performance with fewer parameters and lower latency. The experimental results show that this model achieves a 99.57% F1 score in the Tuberculosis Chest X-ray Database with approximately 11.3K parameters. Code publicly available at: <a href="https://github.com/mingai-007/Dual-RepViT">https://github.com/mingai-007/Dual-RepViT</a></p>