Research On Computer-Aided Diagnosis Of Lung CT Images Based On Convolutional Neural Network

Lung Cancer is the main reason of high mortality caused by cancer and it is a serious threat to human health.People hope to reduce the mortality rate of lung cancer patients by earlier diagnosis.Computed tomography(CT)is one of the main methods of diagnosis and treatment of lung cancer in modern medicine.But a lung CT often contains little information among the millions of image pixels,which makes the early diagnosis of lung cancer is faced with great difficulties.CAD uses the structural features of the lung regions extracted from CT images as a "second opinion" to assist physicians to diagnosis the pulmonary nodules.It has great clinical significance to detect lung cancer that is curable.At present,most researches on CAD in lung cancer are still at the stage of artificially defining nodular features.This paper proposes to use an improved Convolutional Neural Networks(CNN)to diagnose benign and malignant lung CT slices.The work is mainly carried out in four directions:(1)This paper systematically analyzes current researches of computer-aided diagnosis and CNN at home and abroad.It illustrates the improvements in the existing technologies and the significance of the implementation of this article.(2)The unique medical data files of DICOM format are studied.And the image information of experimental samples are described and analyzed.To improve the model's generalization ability,the number of malignant samples is expanded by data augmentation.(3)The traditional convolutional neural network is improved with a skipping-convolutional neural network structure,which helps full-connected layer to obtain feature expressions from different dimensions.The sample feature dimension could be widen,and the feature loss caused by pooling in the traditional convolutional structure could be weaken.Combined with the ensemble learning strategy,the MSS-CNN(Multi-scale Skipping Convolutional Neural Network)is designed to realize the implicit learning of CT images and avoid the tedious task of manually extracting features.(4)The validity of the MSS-CNN model has been confirmed through experiments.The model's recognition rate of the lung CT can be improved effectively by using the new structure.And compared with the traditional CNN method;the MSS-CNN could make full use of the experimental samples,expand the experimental data and improve the adaptability of the model.Besides,the model's recognition result has higher accuracy,which is helpful for improving the early diagnosis rate of lung cancer.The innovation of this research lies in two aspects:First,in the aspect of data sampling,we obtain three kinds of scale CT slicess as training set through data enhancement,image scaling and other operationgs.In this way,the model can obtain the multi-dimensional features of the samples;Secondly,in the aspect of model structure,different from the serial structure of the traditional convolutional neural network,this research construct three skipping convolution neural network submodels of different scales,and the submodels are fused and enhanced by the ensemble learning.Due to the limitations such as time,experimental environment and personal ability,there are still many aspects of this research that need to be improved and perfected.To make the model operate faster and the results better,optimizing the hardware configuration and adjusting the model parameters could be a part of work that needs to be further strengthened in the next.