Image Processing Algorithms Of Central Lung Cancer Based On Neural Network

Lung cancer can be divided into central type and peripheral type according to the location of the lesion,among which central type lung cancer accounts for about 75% of the total lung cancer.In early lung cancer CT images,most of the lesions are small in size,variable in morphology,and adherent to the lining of the chest cavity,leading to a high risk of misdiagnosis through visual observation only,and with the increase of lung CT image information data,it poses a great burden to physicians,so performing efficient and accurate lung cancer detection of computer technology becomes one of the effective methods to solve the existing dilemma.In this thesis,we focus on lung cancer image processing algorithms,which include both lung cancer segmentation and lung nodule benign and malignant recognition.Firstly,the FCMSPCNN lung cancer segmentation model(TA-FCMSPCNN),which combines side suppression and adjustable threshold,is proposed to address the problems of computational complexity,rough edges and details of detection results in the improved PCNN lung cancer segmentation algorithm.The side-suppression mechanism is introduced to obtain more accurate external input stimulus terms and internal activity terms;the adaptive parameter setting method is combined and the correlation parameters in the model are simplified by analyzing the correlation between neurons,and the adjustable threshold value is achieved by adding adjustment parameters;the morphological operation refinement process is performed to obtain refined lung cancer detection results.Experiments show that the proposed method can obtain lung cancer detection results with higher accuracy and more comprehensive detail retention.Then,a nodule detection model with cross-level structure and attention fusion is proposed to address the problems of low generality of TA-FCMSPCNN algorithm and the existing convolutional neural network-like nodule detection methods,such as the difficulty of accurate detection of tiny nodules and loss of details.The cross-level ladder layer structure and multi-scale convolutional kernel are used for different depth feature information extraction to improve the network perception of global features;hybrid attention is added and the high and low level features learned from each layer are fused to make the network locate and identify nodules more accurately;the nodule detection results are output using nonlinear mapping.The proposed model was evaluated on the LIDC-IDRI dataset,and the accuracy rate reached 92.67%.Compared with other algorithms,the detection results achieved by the method in this thesis are more similar to the gold standard and can achieve better detection results for lung nodules of different sizes.Finally,to address the problem that there is more redundant information when using residual networks for lung nodule benign malignancy classification task on the one hand,and on the other hand,it is easy to ignore the importance difference of feature maps of different channels in the classification results,which leads to the limitation of classification accuracy.An improved Res Net50 framework for nodule benign and malignant recognition,namely RE-Res Net,is proposed,which improves the learning efficiency by introducing RC structure to eliminate unnecessary redundant information;furthermore,it combines the effective channel attention mechanism to better capture the importance of feature maps of different channels in the classification task and improve the classification accuracy.The performance of the nodule classification model is validated through comprehensive experiments,and the results show that this improved approach can effectively improve the accuracy and robustness of the residual network-based lung nodule benign-malignant classification task.