Tuberculosis (TB) remains the most common cause of death from a single infectious agent. Early detection and treatment can limit the spread of the disease. One of the critical needs is to use existing diagnostic techniques effectively. Chest X-rays (CXR) examination is the primary diagnostic tool for tuberculosis. In this paper, we propose a deep learning framework for multiclass TB lesion semantic segmentation. Image augmentation and contrast limited adaptive histogram equalization (CLAHE) are used to improve the accuracy of segmentation results. We compare the performance of U-Net and U-Net++ networks. The experimental results show that we could achieve 100% image classification accuracy with U-Net++. On the other hand, the mean intersection over union (Mean IoU) of the detected multiclass lesions can achieve as high as 0.7. The proposed method can speed up TB diagnosis in low and middle-income countries where there is a lack of medical expertise and a severe TB epidemic.