Pulmonary Nodule Detection Algorithm Based On Deep Learning

Early lung cancer screening is an important means to reduce mortality,and pulmonary nodule detection is the first step in lung cancer screening.Currently,computed tomography is the main way to screen early lung cancer.In early lung CT images,the pulmonary nodules are small in size,low in contrast,high in shape heterogeneity,and often interfered by tissues such as blood vessels,trachea,and air bubbles,making the detection of small nodules in the lungs facing enormous challenges.At present,the application of cutting-edge technologies such as big data and artificial intelligence in the medical field has become a trend.In order to solve the difficulty of lung nodule detection,based on the public dataset LUNA16 and three-dimensional convolutional neural network,this paper proposes a two-stage pulmonary nodule detection algorithm: pulmonary nodule candidate detection algorithm and candidate nodule false positive reduction algorithm.In view of the difficulty in accurately detecting small pulmonary nodules,this paper uses the U-Net-like encoder-decoder structure and the fully convolutional network form,combined with the idea of the anchor of the RPN network in the Faster R-CNN object detection algorithm.The three-dimensional pulmonary nodule candidate detection algorithm is proposed.Three algorithms with the same network framework but different basic convolutional blocks are designed for comparison,including common three-dimensional convolutional blocks and residual connections based convolutional blocks.And convolution block based on the dual path connection,and adding coordinate information of the feature map in the original image to the algorithm,and detecting the different size nodules by setting the anchor value corresponding to the nodule size.Aiming at the problem of high false positive rate of candidate detection,this paper proposes four three-dimensional multi-level context information network(3-D MLci-Net)algorithm to extract the characteristics of multi-input and multi-scale volume data of CT data.In this paper,two algorithms are proposed from different training methods of the algorithm,and the posterior probabilities of the sub-networks are weighted and combined to obtain the final classification results.Then,according to the different ways of feature map fusion,two different algorithms are proposed.In order to get richer image features,and to compare the impact of different feature fusion methods on the results.In this paper,the experiments are carried out on public datasets,and the experimental results are compared with the results of other algorithms.The effectiveness of the algorithm is verified.The self-contrast experiments verify the improvement of the algorithm and the rationality of the network structure design.In this paper,the experiments designed in the candidate detection stage compare the effects of different convolutional blocks on the algorithm,and conclude that the dual path connection is more effective for feature extraction of nodules.The sensitivity of the algorithm can achieve 93.33%.And more adequate training will make the sensitivity of the algorithm is improved.The experimental data shows that the data enhancement increases the sensitivity of the algorithm by3.8%.The experimental results of the algorithm proposed in the false positive reduction stage are analyzed.The algorithm obtained by the feature fusion method is superior to the posterior probability fusion method.And the algorithm for independently training the sub-network parameters is better than the training together.From the experimental results,the sensitivity of 3-D MLci-Net-4 is improved by 2.14% relative to the benchmark algorithm.