Research On Lung Nodule Detection Based On Deep Learning

Detection of lung nodules on CT scans has an important value in improving the five year survival rate of early lung cancer patients.Due to the complexity of the nodules,it is still a challenging task to extract discriminative features and obtain low false positive rates.For this problem,the detection methods of lung nodules were studied in this thesis,the specific work and main achievements are listed as bellows:(1)Aiming at the difficult extraction of adhesion nodules,a method based on maximum entropy segmentation and morphology for extracting suspicious nodules was proposed.Firstly,the lung parenchyma of CT image was extracted and segmented with maximum entropy to obtain the region of interest;Then,morphological methods was used to segment the adhesions between the nodule and lung tissue in the region of interest;Finally,the connected-domain labeling method was applied to extract suspicious nodules and the results were saved as a datasets.Experiments shown that the method can break the adhesion area effectively and extract suspected nodular regions accurately.(2)According to the problem of training deep neural networks using small-scale datasets,a lung nodule detection method based on transfer learning and SVM is proposed.Firstly,the pre-trained VGG of ImageNet was copied as the target mode;Secondly,the model was fine-tuned to make it suitable for pulmonary nodule classification tasks;Then,the datasets in(1)was used to optimize the weight of the model filter by backpropagation;Finally,the deep features extracted by the fine-tuned model was used to train and test the SVM.Experiments shown that the method can detect nodules accurately and has a low false positive rate.(3)For the problem of redundancy caused by features extracted by deep CNN,a lung nodule detection method based on feature optimization was proposed.Firstly,deep features were extracted from the fine-tuned model.Secondly,different dimension reduction algorithms were applied to optimize the deep features.Finally,the SVM was trained and tested by thedimension reduction features.Results shown that the method has a highly accuracy and can significantly reduce the time required to train classifiers.