Research On Automatic Detection And Intelligent Classification Of Lung Nodules Based On Deep Neural Network

The incidence and mortality of lung cancer rank first among malignant tumors,and the early symptoms are extremely hidden,and patients often miss the best treatment period.Early detection and classification of lung nodules is the key to the diagnosis of early lung cancer.Computed tomography can completely show the location,texture,shape and other characteristics of the lesion.However,hundreds of gray-scale images are generated for each scan.Manual reading not only wastes a lot of time but also prone to visual fatigue,leading to false detection and missed detection of lung nodules.At the same time,the classification of lung nodules depends on their own inherent heterogeneity.It is difficult for radiologists to define it based on visual parameters and the diagnosis is subjective.Therefore,the study of an intelligent computer-aided diagnosis system for lung nodules is of great significance to improve the early screening rate of lung cancer.This paper studies the automatic detection algorithm of lung nodules based on the attention multi-scale convolutional neural network and the classification algorithm of benign and malignant lung nodules based on the three-dimensional cross-fusion convolutional neural network,and realizes the accurate detection of lung nodules in CT images and benign and malignant.Accurate classification of lung nodules.The main research contents and contributions are as follows:(1)In the CT image,the lung nodules have complex backgrounds and different shapes,and the micro-pulmonary nodules are similar in shape to blood vessels,resulting in inaccurate detection of lung nodules.This paper proposes an automatic lung nodule detection algorithm based on attention multi-scale convolutional neural network.The dense connection structure is used to ensure the transfer and transmission of multi-scale fusion features in the network,and reduce the loss of information from small nodules.According to the characteristics of different characteristics,spatial attention and channel attention are introduced respectively to explore the difference between lung nodules and other tissue characteristics and the primary and secondary relationship between key features.Enter a large number of false positive and negative samples to reduce network parameters and speed up model inference.Validated by the LIDC-IDRI data set,the average detection accuracy of lung nodules is 94.98%,and the recall rate is 95.47%.(2)Aiming at the problem that the two-dimensional slice data loses the three-dimensional information of lung nodules and the low-level features of lung nodule density,grayscale,sphericity and other low-level features are insufficient to represent the heterogeneity of nodules,this paper proposes a lung nodule based on a three-dimensional cross-fusion convolutional neural network Benign and malignant classification algorithm.The three-dimensional nodule data with a small amount of background information and zero-filled margins is intercepted,and the three-dimensional characteristics of the nodule are fully extracted,and a large number of zero values accelerate the calculation of the three-dimensional convolution.Cross fusion to obtain a multi-scale mutual information feature group,which enhances the expression ability of high and low-level semantic information.Multiple classifiers simulate multiple doctors with different experiences to make joint decisions to improve the accuracy of benign and malignant classification.Validated by the LIDC-IDRI data set,the classification accuracy of this algorithm is 90.96%,and the AUC value is 94.95%.