Multimodality Medical Image Classification Based On Convolutional Neural Network

Background Deep learning is one of the fast-developing branches in artificial intelligence,it provides technical support for intelligent development.Convolutional neural network is one of the typical models in deep learning,it has good application in computer vision.Lung cancer is one of the malignant tumors with rapid increase in morbidity and mortality,which threatens people’s life and health.Computer aided diagnosis is very important for early detection and treatment of lung cancer.It can accurately analyze the patient’s focus of infection and quickly provide the results of assistant diagnosis and treatment for doctors so as to achieve the consistency of image and pathological interpretation.Computer aided diagnosis can also improve the working efficiency of medical staff,and reduce the rate of misdiagnosis and missed diagnosis.Objectives In this paper,three modals of lung tumor（CT,PET,PET/CT）are taken as research object,and the lung tumor images are recognized by traditional convolution neural network model-LeNet-5,to achieve computer aided diagnosis of lung tumor,it can help patients,assist clinical medical staff to diagnose accurately,reduce the intensity of work,improve the speed of diagnosis,and realize the application of convolutional neural network in clinical practice.Methods On the basis of LeNet-5 model,two special model structures,random fusion and PSO optimization convolution kernel,are proposed and applied to lung cancer image recognition.Firstly,three single-mode convolution neural networks are constructed on different modes of lung cancer images,CT-CNN,PET-CNN,PET/CT-CNN.Fully connected feature vectors and weights of three single-modal CNNs are fused randomly,to identify lung tumors by reconstructing full-connectivity feature vectors.Secondly,the convolution kernel is extracted and the particle swarm optimization algorithm is used to constantly update its speed and position to find the global optimal result for initialization,so the global optimization of PSO is combined with the local optimization of back-propagation.the network model is trained by lung tumor data to obtain a better recognition effect.In this paper,six indicators including training time,sensitivity,specificity,recognition rate,MCC and F₁Score were used to comprehensively evaluate the performance of different convolutional neural network structures.Results For the study of randomized fusion and CNN multi-mode lung tumor image recognition,the experiment discussed the influence of randomized feature fusion on the recognition results from three dimensions of iteration numbers,batch sizes and network layers.The results show that it is feasible to recognize lung tumor images with randomized feature fusion and CNN,and the changes of iteration number,batch size and network layer have a certain impact on target recognition of convolutional neural network.However,the recognition effect of randomized feature fusion is obviously better than that of single mode convolutional neural network,which reflects the superiority of randomized fusion algorithm in identifying lung tumors.For the lung tumor recognition research of pso-convk convolutional neural network,the experiment discussed the influence of PSO optimal convolution kernel on lung tumor recognition from three dimensions of iteration numbers,batch sizes and network layers.The results show that PSO optimal convolution kernel is feasible for CNN to recognize lung tumors,and the change of parameters has a certain impact on the CNN recognition.However,the recognition effect of PSO optimal maximum convolutional layers is obviously better than that of traditional convolutional neural network and gaussian optimal convolutional kernel,which indicates that PSO optimal convolutional kernel has certain advantages in the recognition of lung tumors.