Algorithm Research Of Lung Nodule Segmentation Based On Deep Learning

Lung cancer is one of the most important diseases that harm people in recent years.The main reason is that when patients are diagnosed,most of them are in the middle and late stage of the disease and have missed the best treatment time.During the initial period of the disease,lung cancer is still in the primary form of lung nodules,and if it is detected early,it is crucial for the diagnosis and treatment as well as the cure of the patient.However,the primary laboratory test used by doctors today is Computed Tomography(CT),which requires doctors to identify hundreds of lung scans,and due to the huge workload of doctors,there is a possibility of misjudgment and miscalculation.Moreover,because the radiological part of the lung scan is not significantly different from the background area,it is particularly similar to other organs in terms of sensory,and the ratio of the area occupied by lung nodules is small and the doctors’ experience varies,the diagnosis of the condition may vary.In order to solve the above problems,this thesis proposes a U-Net-based lung nodule segmentation model: the MRBU-Net-WD model,which is improved as follows:(1)A Multi-scale Densely Connection(MSD)algorithm is proposed.This algorithm fuses the information of features at different scales in the encoding and decoding stages,so that the feature information input to each stage of the model is all the features of the previous stage,which enriches the feature information and improves the segmentation accuracy.As shown by the experiments,this algorithm improves the Dice coefficient in lung nodule segmentation by 4.5% compared with the original U-Net model,which is a significant improvement.(2)A Residual 3D Convolution(R3D-conv)algorithm is proposed to combine the U-Net network with the residual network,which is a good solution to the phenomenon of network degradation when the network is deepening,and the training time of the model is greatly reduced,and also reduces the possibility of gradient disappearance and gradient explosion.As shown by the experiments,this algorithm also has a substantial improvement of 5.3% in Dice coefficient compared with the original U-Net model in lung nodule segmentation.(3)A Bidirectional Feature Pyramid Network(Bi-FPN)network structure is proposed to improve the original jump connection of U-Net model into Bi-FPN,and the output of the feature network is combined with the decoder architecture respectively to improve the feature extraction efficiency at each level of the backbone architecture and enrich the feature vector to obtain the combination of lower-level features and higher-level semantic features.This improvement improves the Dice coefficient by 4% compared to the original U-Net model.(4)A weighted Dice loss function was proposed and compared with Focal loss function and weighted cross entropy loss function.The experimental results showed that compared with Dice loss function,weighted cross entropy loss function and Focal loss function,the weighted Dice loss function increased by 3.0%,6.1% and 3.1%.The experiment proved that the weighted Dice loss function had a better effect,which better solved the problem of pixel imbalance between lung nodules and lung pictures due to the small proportion of lung nodules.The MRBU-Net-WD model propose in this thesis was validated in Lung Image Database Consortium/Image Database Resource Initiative(LIDC-IDRI)and compare with U-Net,U-Net3+,U-Det,and RU-net,and it is experimentally demonstrated that the Dice coefficients of the model propose in this thesis were improved by 17.8%,11.8%,8.1%,and 2.4% compared with the comparison model,and the sensitivity rate and Jaccard coefficient are also significantly improved.The MRBU-Net-WD model propose in this thesis has a high accuracy in lung nodule segmentation,which is useful for the diagnosis of lung cancer and the development of intelligent medical treatment.