Pulmonary Nodule Recognition Based On Deep Learning

Pulmonary nodules are the early imaging findings of lung cancer,and the recognition of nodules plays a significant role in the prevention and the treatment of lung cancer,but the current medical resources are far from meeting the demand of early screening,as well as the early diagnosis.Thus it is necessary to develop the automatic recognition algorithm or the intelligent recognition algorithm for pulmonary nodules.Before studying the algorithm of pulmonary nodule recognition,this paper also does some research about lung parenchyma extraction for the purpose of avoiding interferences with other normal tissue and improving the effect of pulmonary nodule recognition.A simple and effective algorithm for lung parenchyma extraction is proposed.After the extraction of lung parenchyma,pulmonary nodule recognition can be studied.Traditional algorithms for pulmonary nodule recognition are usually based on threshold,clustering or template matching,the implementation is complicated and the accuracy is low.In recent years,with the development of artificial intelligence(AI)and computer vision(CV),deep learning(DL)and 3D convolutional neural networks(3D CNNs)have gradually become a prevalent choice for pulmonary nodule recognition.In this paper,pulmonary nodule recognition is accomplished by deep learning.During the implementation,two technical routes are tried respectively: the one is based on semantic segmentation(pulmonary nodule segmentation),the other is based on object detection(pulmonary nodule detection).There are two stages in the pulmonary nodule segmentation algorithm:firstly,the region suspected to contain nodules are extracted by semantic segmentation algorithms;then the classification algorithms are used to distinguish the nodules from a large number of candidates extracted in the previous stage,also known as false positive reduction.In the training or the inference phase,the strategies of multi-model fusion,multi-stage combined training and multi-angle inference are applied to improve the performance of the algorithm.The dataset of Tianchi Medical AI Contest is used to verify the effectiveness of the proposed segmentation algorithm.The pulmonary nodule detection algorithm is an one-stage algorithm,the probability and the location of nodules are obtained directly by deep convolution neural networks.Compared with the two-stage segmentation algorithm,the single-stage detection algorithm is more efficient.The dataset of Lung Nodule Analysis 2016(LUNA16)Challenge is used to verify the effectiveness of the proposed detection algorithm.