Research On Lung Nodule Detection Algorithm Based On Deep Learning

Among many diseases,lung cancer is the one with the highest prevalence and the highest fatality rate among many cancer types.Current early diagnosis of lung cancer usually uses Computed Tomography to perform tomography and imaging of the lungs.During the diagnosis process,there are problems such as a wide variety of lung nodules,small sizes,and the number of CT slices is frequently hundreds of pieces.This makes the manual diagnosis of pulmonary nodules not only valuing the doctor’s experience,but also consuming manpower and material resources.With the growth of algorithms in computer vision in recent years,the technology for computers to learn features and make judgments on their own has become more sophisticated.Artificial intelligence-assisted diagnosis has become a popular technology for lung nodule detection nowadays.With the above-mentioned problems,this paper completed the research of lung nodule detection algorithm based on deep learning.Two parts included in the research are as follows:First,in view of the large number of CT slices in the lung nodule detection algorithm,and the lung nodules are mostly small sizes that are difficult to diagnose,this paper give out a grand new C-E data augmentation scheme.Through the extraction of the sliding window of the lungs and the multiple replication of small lung nodules,the background noise is reduced while the number of positive samples is increased.At the data level of the algorithm,the diversity and stability of the samples are increased,which greatly improves the accuracy of the detection algorithm for small lung nodules and improves the ability to suppress false positive nodules.Second,in view of the various types of lung nodules,and the adhesion and confusion of vessels and organs with similar features,this paper proposes a new multi-scale structure of convolutional neural network in the lung nodule detection algorithm,and combines the new context semantic module CONTEXTn and convolution module RDn for feature extraction and feature processing.Firstly,in terms of structure,the feature pyramid structure is used to achieve multi-layer feature output and reuse;secondly,different numbers of convolution modules are used and then concatenated to strengthen the integration of different receptive field information;finally,a new type of neural network module that combines dense links and residual links is proposed,which ameliorates the efficiency and correctness of feature collection while reducing the amount of parameters.This paper improves the feature extraction and feature utilization capabilities of nodules at the level of the algorithm network by improving the structure and modules of the network.