Lung Nodules Detection On Medical Images Based On Deep Convolutional Neural Networks

With the development of the times,the air pollution has increased and lung cancer has become one of the malignant tumors which are in the highest degree of threating the human lives.We can greatly improve the patients' survival rate while we can treat patients early after detected early.Lung cancer appears to be isolated pulmonary nodules in the medical images in the early stage,and usually appears round or myopic circular and low contrast spot in the chest radiographs.Without special processing,it is difficult to distinguish the lung nodules from other lung soft tissues only with our eyes.Deep learning is the new field of machine learning,and develops rapidly in recent years.Its essence is to use a large number of hidden layer neural networks and training with a large number of data,so as to extract more useful features from the data to improve the accuracy of the prediction or classification.Deep learning has been well applied in many fields such as image processing,speech recognition and natural language processing,and has achieved good results.In this paper,we mainly study the methods of automatic detection of lung nodules in chest radiographs,and explore the application of deep learning method in pulmonary nodule detection in CT images.The main contents are as follows:(1)We proposed a lung nodule detection scheme in chest radiographs based on convolutional neural networks after study the traditional lung nodule detection methods in chest radiographs.The proposed scheme first preprocesses the chest radiographs to enhance the lung nodules signals using USM sharpening method.Then we subsamples the patches cut from preprocessed chest radiographs by sliding window method and then feeds them into the pre-trained convolutional neural network for classification.Then we can get candidate nodules of the whole chest radiograph.Finally,we can exclude lots of false positives by applying an area threshold to the candidates.The proposed method can detect more lung nodules than those in the relevant literature at the same false positives level by evaluation on the JSRT database.(2)We proposed an ensemble convolutional neural network for lung nodules detection in chest radiographs after study the morphological manifestations of subsampled patches of pulmonary nodules.The proposed method first also preprocesses the chest radiographs to enhance the nodules signals using USM sharpening method,and then cuts 229×229 patches from the chest radiographs with sliding window method and subsamples to 12×12,32×32 and 60×60 three different scales and then feeds them into three different pre-trained convolutional neural networks respectively.The final class of the patches is decided by the voting of the output from the three convolutional networks and then got the candidates for the whole chest radiograph.Finally we can exclude a large number of false positives with one area threshold.The experimental results on the JSRT database show that this ensemble method can eliminate a large number of false positives and detect more nodules at the same false positive level than in the relevant literature and that in the previous chapter.(3)We preliminary explored the suspect lung nodules detection in CT images.We proposed a multi-input convolutional neural network model for the detection of suspected pulmonary nodules in CT images after studying the stereoscopic characteristics of lung nodules.This method first preprocesses the CT slices and enhances the nodules signal with USM sharpening method.Then,small patches of the same size are cut from the same position of the adjacent slices and feed into the pre-trained multi-input convolution neural network to obtain the candidate regions in the whole CT sequence.With the area threshold can do the initial false positive screening work.In terms of the experimental results on the subset of databases selected from LIDC-IDRI,the detection rate of nodules meets the requirements and can be used for further research.