Pulmonary Nodule Detection And Recognition Method Based On Neural Network

Lung cancer is one of the deadliest cancers all over the world.The detection of pulmonary nodule at its early stage in the computed tomography(CT)images is crucial for increasing a patient’s survival rate.Medical diagnosis is mostly based on doctors’ experience,and it is time-consuming and laborious,so computer-aided diagnosis(CAD)is gradually applied to imaging.Conventional CAD methods mainly rely on a variety of filters to detect the edge,texture and other characteristics of different types of nodules.Recently,with the help of the powerful image expression ability,CAD system using neural network can accurately and quickly locate and identify the pulmonary nodules.The key point in the pulmonary nodule detection is the accurate detection of nodules,especially early(small)pulmonary nodules.Accurate detection means detecting pulmonary nodules as many as possible while minimizing the number of false positives.At present,most methods of detecting pulmonary nodules using neural network are based on single feature map,and it is not conducive to the detection of small targets.This thesis used the pulmonary CT data obtained from the 2017 TianChi AI competition,and contributed the following on the pulmonary CT image pre-processing method,nodule detection and recognition algorithm:1.Analyze the difference between CT images and normal digital images.Based on the characteristics of the lung CT images,this thesis proposes a data pre-processing procedure that comprises of various image morphological operations.2.Nodule detection task is a single target detection task,and the background of chest CT images are similar to each other.Based on these characteristics,a multi-layer feature map combination method is proposed to optimize detection models(SSD,MDSSD)through feature selection and combination.Comparing with the original multi-class detection,the improved neural network models are more suitable for the detection of pulmonary nodules.3.To balance the positive and negative samples in the training of SSD and MDSSD,a balance parameter is added in the loss function to adjust the weight of positive and negative samples in training.Compared with the loss function used in original models,the optimized loss function makes the modified models converge better,and the confidence of true nodule determination are higher.Finally,in order to reduce the number of false positives in the final output,and also for the integrity of the overall process,this thesis also designs a false positive reduction process based on ensemble learning and AlexNet.Through the above research,this paper optimizes the performance of multi-feature neural network in nodule detection,and its detection sensitivity can reach 95%,which is in the leading level in this area.