| Lung cancer,the main cause of global cancer death,has been widely watched.Lung nodules are an early manifestation of lung cancer,and in the process of its detection,computer aided diagnosis technology（CAD）is an important link between computer and medical image processing.Compared with doctors only diagnosing lung nodule through their own experience,the results are more accurate and faster.The use of computer-aided diagnostic techniques not only reduces the workload of doctors,but also prevents doctors from causing the misdiagnosis of the nodes by working fatigue,and provide patients with more accurate detection results.With the development of artificial intelligence technology,more and more scholars have put their energy into the research of computer-aided diagnosis.The basic segmentation of lung is the first step in the study of computer aided diagnostic technology,and the precise segmentation of the substance of the lung can effectively help doctors to diagnose lung nodules in their post-work work.Because of the problem of boundary segmentation and under-segmentation in the previous lung actual segmentation,the random walk algorithm is a graph theory of segmentation algorithm,which can achieve better results when dealing with weak edges,Therefore,the random walk algorithm is used in lung parenchyma segmentation in this study.The traditional random walk segmentation algorithm has great dependence on users.In this paper,an automatic random migratory lung agent segmentation algorithm is proposed.First,lung parenchyma was preliminarily segmented by OTSU.Secondly,the corrosion boundary points and expansion boundary points of the initial segmentation results were obtained by mathematical morphology method,which were used as the front and background seed points of the random walk algorithm to realize the automatic acquisition of seed points.Finally,to further improve the accuracy of lung parenchyma segmentation,the shortest distance factor between sets was added based on the original random walk weight function calculation.Experimental results show that the proposed method not only achieves automatic segmentation of lung parenchyma,but also outperforms the manual interactive random walk algorithm in segmentation accuracy.In the process of image-assisted diagnosis of Lung nodules,the input data of traditional machine learning methods are artificially defined characteristics of lung nodules.These characteristics are unitary and subjective,leading to poor recognition effect of lung nodules.Therefore,this paper proposes an image recognition method of lung nodules based on deep feature fusion.Firstly,the spatial texture features of lung nodules were extracted by Volume Local Direction Ternary Pattern,and the gray scale features were extracted by Gray Histogram,and the correlation analysis was carried out.Secondly,the relevant texture features and gray features of lung nodules were fused with the depth features extracted from VGG16 network model in series.Through the training and testing of the model,pulmonary nodules can be recognized.Finally,the Confusion matrix is used to evaluate the effectiveness of the model.Experimental data were used in this study from the Lung CT Image Database（LDC-IDRI）,which is collected by the National Cancer Institute（NCI）.Based on this database,the segmentation of lung parenchyma and the evaluation of the effect of the image-assisted diagnosis model of lung nodules are mainly realized.The experimental results show that the improved automatic random walk method is better than the original random walk algorithm,and the average segmentation accuracy is 97.84%.Meanwhile,the recognition accuracy of lung nodules image based on VGG16 neural network reached 97.5%,sensitivity reached 96.81%,specificity reached 98.11%,accuracy reached 97.85%,and F1 score reached 97.3%.Compared with the single depth feature,the VGG16 lung nodule recognition network model with multi-feature fusion has higher recognition accuracy. |
PDF Full Text Request