CNN-Transformer Combined with Skip Structure for Tuberculosis Cavity Segmentation

Tuberculosis is a global public health problem. Tuberculosis cavity is an important imaging sign of tuberculosis. Accurate segmentation of tuberculosis cavity from Computer Tomography (CT) images is the prerequisite and main task for computer-aided diagnosis (CAD) and clinic treatment of tuberculosis. However, irregular shape and fuzzy boundary of tuberculosis cavity bring challenges to the segmentation task. To address these problems, we propose a novel CNN-Transformer combined with skip structure (TransUNet++) method. Firstly, we employ the CNN-Transformer module as an encoder, in which 3 convolution layers are used as the local feature extractor and 12 Transformer layers are treated as the global feature extractor. Secondly, we utilize the skip structure of the decoder module to obtain the comprehensive features. TransUNet++ can not only overcome the limitations of convolutional neural network (CNN) in obtaining global features by the CNN-Transformer module, but also narrow the semantic information difference of features between the decoder and the encoder by the skip structure of the decoder module. Experiments on CT images of patients with tuberculosis cavity provided by our cooperative hospital show that the proposed TransUNet++ outperforms the state-of-the-art comparison methods in terms of segmentation accuracy, and it can provide a positive guiding role in the formulation of clinical treatment plan and evaluation of the recovery of tuberculosis cavity.