Segmentation Of Pulmonary Nodules Based On Deep Learning

With the prevalence of tobacco,air pollution and other factors,lung cancer has accounted for two-fifths of all cancers in the world.In recent years,with the continuous improvement of medical treatment level,lung cancer has attracted more and more attention from people,and the relevant data has become more and more detailed.The early symptoms of lung cancer in lung nodules on medical image,in the treatment of pulmonary nodules,the doctor is in principle will be in accordance with the patient's history,nodule size,shape,and the X-ray or ct before the check to determine,this is likely to some possible into lung nodules leak detection,patients cannot receive timely diagnosis and treatment,missed the best time,do great harm to human beings.Therefore,timely segmentation and detection of pulmonary nodules is of great significance and value for prevention and follow-up treatment plans of doctors.At present,the only early screening method that can reduce the mortality rate is low-dose spiral CT scan.However,due to the concealment of early lung cancer and the lack of cancer awareness in humans,early detection of lung cancer and early treatment are still a long way off.This paper focuses on the deep learning-based lung nodule segmentation algorithm,aiming to provide more accurate and effective auxiliary diagnosis methods for doctors,improve the survival rate of early lung cancer,and play a positive role in the early diagnosis and treatment of lung cancer.The main research contents and achievements are as follows:1.In view of the small data set segmentation difficult problems of medical images,we used the migration study,the method of using VGG-16 in a large amount of data,the characteristics of coarse-grained natural images on learning,knowledge,fitting network parameters,and then moving characteristic information into small data sets,fine-grained pulmonary nodules on image segmentation tasks,and in order to improve the segmentation performance,put forward a kind of block type stack is fine-tuning(BMFT)strategy as auxiliary segmentation.The experimental results showed that the BMFT method achieved the Dice value of 0.9179 on the LUNA16 lung nodule public data set,and its performance was significantly better than that of the mainstream lung nodule segmentation network.2.Has been proposed for semantic segmentation of typical network U-Net,but U-Net web based medical image segmentation,there are some disadvantages: although network category for each pixel prediction accuracy is high,but ignores the relationship between the pixels,the results make segmentation result is not continuous,or the size is different from standard size clearly divided.To solve these problems,a method of lung nodule segmentation based on generative antagonistic network was proposed to improve the segmentation performance.The experimental results showed that the method reached the Dice value of0.9120 on the LUNA16 data set.