Research On Lung Cancer Recognition Method Based On Multi-scale And Feature Fusion

Lung cancer is a common malignant disease of the respiratory system.It is easy to be ignored in the early stage and usually comes to the middle and late stage after diagnosis.The survival rate is low.If the patients can be effectively detected and treated in the early stage of tumor growth,the survival probability of patients can be greatly improved.Pulmonary nodule detection is a key step in the diagnosis of lung cancer.With the development of image processing and artificial intelligence technology,deep learning has achieved many important results in the detection of pulmonary nodules.However,due to the strong professionalism of labeling CT images,insufficient labeled data,and different sizes and complex signs of pulmonary nodules,it is difficult to detect pulmonary nodules.At the same time,the small-scale input nodule detection network is difficult to obtain the overall characteristics of large lesions,which will affect the recognition of lung cancer.Aiming at the above problems,this paper proposes a multi-scale and feature fusion method based on transfer learning,which is used to predict the probability of lung cancer in the next year according to CT images.The main research contents are as follows:Firstly,a multi-scale input method is proposed to detect pulmonary nodules because it is difficult to obtain the overall features of large lesions in small-scale input networks.According to the size of nodules and lung masses,three different scales of image blocks were input into the nodule detection network.In small-scale input,the network pays attention to the details of lung lesions;When large-scale input is used,the network can capture the overall feature of the lesion.In order to ensure the complementarity of the multi-scale features,the multi-scale features extracted from the network are combined to recognize the lesions.The experimental results show that the proposed method improves the ability of obtaining the overall features of lung lesions and the accuracy of lung cancer recognition.Secondly,because the features extracted from the last layer of the network are single,the detailed features of the lesions cannot be fully described,so the feature fusion strategy is used to recognize lung cancer.Features of bottleneck layer extracted by neural network and features extracted by output layer are fused to predict lung cancer.The experimental results show that the bottleneck layer feature can more fully describe the detailed characteristics of lung lesions and improve the accuracy of lung cancer recognition.Finally,combining multi-scale input and feature fusion,a lung cancer recognition method based on multi-scale and feature fusion is proposed.The fusion features at three scales are recombined to predict lung cancer.Aiming at this method,two multi-scale lung cancer recognition strategies are designed.Experimental results on the Kaggle Data Science Bowl 2017 data set show that the proposed method reduces the loss of lung cancer prediction and improves the accuracy of lung cancer recognition.