Research On Lung Nodule Detection Based On Transfer Learning And Convolutional Neural Network

With the development of information technology,lung cancer has become one of the most malignant tumors that threaten human health.Therefore,the early diagnosis and treatment of lung cancer is of vital importance to human life safety.In medical imaging,the early symptoms of lung cancer are pulmonary nodules,which usually appear as circular or nearly circular low-contrast spots in CT images.It is difficult to distinguish lung nodules from other lungs without special treatment and further tested.Transfer learning and deep learning are two emerging areas of machine learning that have grown rapidly in recent years.In general,transfer learning is to apply the knowledge learned by one system to the learning model in another related field,while deep learning is to train by constructing multiple hidden neural networks and using massive amounts of data.The data extracts useful features to improve the classification accuracy of the model.Both have achieved remarkable results in various fields such as image processing,speech recognition and natural language processing.At present,most of the main lung nodule detection methods are based on traditional machine learning algorithms.The disadvantages of these methods are that they need to manually mark the sample features and then train the classifiers,which leads to the low efficiency and poor adaptability of the trained models.Since transfer learning can transfer knowledge learned from one domain to another,the convolutional neural network(CNN)combines the performance of feature extraction and classification,enabling the model to automatically learn features,which not only reduces the cost of medical image acquisition,but also avoids the complicated artificial feature extraction work.In this paper,lung nodules will be detected using transfer learning and CNN based on the case of a small amount of CT images in the lungs.The main research contents are as follows:(1)Data preprocessing.Firstly,the Gaussian filtering algorithm is used to complete the noise removal of the lung CT image dataset.Then the threshold binarization method is used,and the Kmeans algorithm is used to divide the black and white region of the image after denoising to remove the detailed information in the image and to enhance the contour of the target image,and then use the morphological principle to erode and expand the binarized lung image to achieve a clearer process of segmentation of the lung CT image.Finally,based on the previous work.Further,the measure sub-module is used to cut the non-ROI region in the image,thereby obtaining a series of lung CT sample images that can be used to train the model.(2)Directly training network model.Using the pre-trained convolutional neural network model(AlexNet)from the large-scale natural image classification as the feature extractor to extract the features of the lung nodule image,that is,the random initialization method,and then the lung CT image is detected.(3)Fine-tuning transfer learning.As can be seen from the experimental results of the random initialization in the previous section,the classification results of the directly training network model parameters are not very satisfactory.Therefore,on the basis of the previous section,we chose to use the fine-tuning transfer learning method to automatically learn the lung nodule images.That is to say,using the pre-training model based on feature network for training,this paper uses lung nodule images for training.Through the comparative analysis of the experimental results,the final conclusion is that the introduction of migration learning into the detection of pulmonary nodules can not only make up for the over-fitting problem caused by the small data set,but also use the pre-trained network model parameters for lung nodules to extract image feature can effectively improve the classification accuracy.It is sufficient to demonstrate the effectiveness of using transfer learning and CNN to classify lung nodules.The innovations of this article include the following two points:(1)Most of the previous research scholars used traditional machine learning methods to perform lung nodule detection based on a large number of artificially labeled image feature samples,which is time consuming and labor intensive.Therefore,this paper considers the characteristics of medical image datasets belonging to small datasets,and introduces transfer learning methods to detect lung nodules,thereby reducing the workload of doctors.(2)Most of the researcher's training samples for lung nodule detection are manually labeled,which brings a huge workload to the doctor.Therefore,this paper considers the cost of artificially labeled medical image features too high,and introduces a convolutional neural network model to assist in the extraction of lung nodule image features,so as to better help doctors diagnose.