Detection Of Pulmonary Nodules Based On Convolutional Auto Encoder Neural Network

Lung cancer has become the highest mortality rate in the world.The most effective way to improve the survival rate of lung cancer is early detection,early diagnosis and early treatment,and the chest CT scan image provides the possibility to achieve this goal.However,it is not only time consuming,but also very difficult to detect a nodule image from a large number of CT images.Although some computer aided diagnosis system(CAD)have been introduced,but these CAD need professionals to spend a lot of time mark in the CT image of lung nodules in CT image,which need huge manpower cost,and the segmentation of lung nodules from the ROI region may lose important boundary information of pulmonary nodules.In addition,the feature extraction of CAD system is done by hand,and the method of manual feature extraction is difficult to determine the effective features.We studied the detection of lung nodules in chest CT images and put forward a deep learning model of convolutional autoencoder neural network(CAENN)to detect pulmonary nodules.Our method fully takes into account that the manual annotation of ROI region of lung nodules needs huge human cost and unannocated CT images is very rich,so we first use a large number of unlabeled CT image patch for unsupervised learning of image features,then use a small amount of labeled data for supervised fine-tuning to learn a good image classification system for detection CT images with pulmonary nodules.The whole model only needs to use a small amount of CT images of the lung nodules in the ROI area,thus saving a lot of labor costs.Moreover,the deep learning model can automatically learn the features of the image without the need for manual feature extraction,so as to avoid the loss of important information.The CT images of the patients of the experiment that we used come from LIDC/IDRI,are being used to generate 50000 unlabeled image patches for unsupervised learning of image feature and 5500 labeled image patches for supervised fine-tuning.The final classification accuracy rate reached 91%,Especially when the training data has only 2200 samples,the classification accuracy has reached 84%.Experimental results show that,compared to image classification algorithm like ordinary convolution neural network,our algorithm models,not only has improved the performance,but also outperforms when there are few labeled data.