CT Images Recognition of Pulmonary Tuberculosis Based on Improved Faster RCNN and U-Net

Abstract

Pulmonary tuberculosis has the characteristics of polymorphism, multi-site, multi-nodule and cavity, so it is difficult to be identified in lung CT images. At present, there are two main methods for tuberculosis recognition in CT images: manual recognition and computer-aided recognition. Manual recognition will expend much energy, and medical staff would miss diagnosis or even misdiagnosis under the high-intensity work pressure. Most of the computer-aided recognition methods use traditional methods, and the recognition effect is poor. In this paper, we focus on the recognition of pulmonary tuberculosis based on CT images, and improve the traditional CT analysis process, which has two steps of segmentation and classification. We construct an improved deep network model of combining U-Net and Faster-RCNN, which is optimized by Res block. The model can detect and identify the tuberculosis lesions in pre-processed CT images simultaneously. A large number of experimental results show that our model has better effects on tuberculosis recognition than other machine learning models.