Generalized Lesion Detector Based on Convolutional Neural Network

In current computer-aided diagnostic systems, existing detection functions are usually only for a specific type of lesion, such as skin lesions, pulmonary nodule lesions and liver lesions. However, in actual clinical diagnosis, many lesions are actually related. For example, pulmonary nodular lesions may metastasize and spread to lymph node areas or other body parts. The detection of a single site is not conducive to the doctor's comprehensive diagnosis of the condition. Multi-site lesion detection can detect lesion metastasis and treat it earlier, and can also explore the relationship between different lesions. In response to this situation, this thesis uses the Deeplesion dataset to establish a general lesion detection framework that can detect possible lesions through CT images of different parts of the body. Compared with the existing single-path computer-aided diagnosis system, this thesis studies and implements a general lesion detection system to explore the relationship between different lesions. This will help doctors to make a comprehensive clinical diagnosis and visualize the results. This thesis based on the Faster R-CNN network model. First it denoises and enhances the CT images. Then, the VGG16 network is used for feature extraction, and the feature map is obtained through the RPN network to obtain candidate suggestion regions. In view of the misdetection of missed detection, this thesis introduces a Gaussian weighted penalty function to improve the non-maximum suppression. Finally, Tkinter is used to create a GUI visualization interface for doctors to compare clinical diagnosis.