PET/CT Image Fusion Of Lung Tumors Based On Deep Learning

China is one of the countries with the largest number of lung cancer patients in the world.According to the annual report of China Tumor Registry in 2020,there were about 851,000 new cases of lung cancer in China and about 695,000 deaths,still ranking first in the incidence and mortality rate of lung cancer in China.Thus,effective prevention and treatment of lung cancer is imperative.In the diagnosis of lung cancer,Positron Emission Tomography(PET)provides organ metabolism information while Computed Tomography(CT)provides highresolution anatomical structure information.The fusion of these two images can synthesize their multi-modal complementary information and provide a more comprehensive basis for clinical medicine,which has important application value.Based on the theories and methods of medical image fusion and deep learning,this study studied the multi-modal lung tumor image fusion algorithm based on PET and CT.The main work is as follows:In view of the problem that traditional medical image fusion algorithms do not fully consider the visual priorities of different positions and different information of multiple modes,this study proposed a collaborative learning based convolutional neural network PET/CT image fusion algorithm.Firstly,the convolutional neural network is used to obtain the specific features of the multimodal images,and then the fusion map of spatial changes reflecting the relative importance of the spatial positions of the multimodal features is generated based on the features.The fusion map was then multiplied with the multimodal specific feature map to obtain multimodal complementary information at different locations for better analysis of multimodal lung tumor images.In view of the problem that traditional medical image fusion algorithms are mostly forward propagation algorithms and cannot adjust fusion weight through feedback,this study proposed a generative adantagonistic network PET/CT image fusion algorithm based on visual geometric group.The powerful feature extraction capability of the algorithm reduces the requirement of source image registration,avoids the manual design of fusion rules,and improves the fusion efficiency.The loss function is designed based on Graeme matrix,which helps the fusion image retain more details and reduce noise interference,ensuring the stability of training and the quality of fusion image.According to the characteristics of multi-modal medical images,the model is optimized continuously according to the loss function,and the fusion weight is adjusted by the back propagation algorithm to improve the quality of the fusion image.The experimental results show that the two algorithms proposed in this study can better integrate the metabolic information of PET and the structural information of CT,and generate high-quality lung tumor fusion images.The proposed algorithm has good subjective and objective performance,which can assist clinicians in the diagnosis and treatment of lung tumors and reduce the storage capacity.