| The intelligent diagnosis and treatment system is the surgical cornerstone for the realization of minimally invasive diagnosis and treatment of lung tumors.This research analyzed the key technologies of intelligent image processing in the image navigation robot-assisted lung tumor minimally invasive diagnosis and treatment system,focused on the intelligent detection and diagnosis of lung nodules,and intelligent segmentation of lung tumors and multiple organs.The specific works are as follows:(1)Aiming at the problems of low early detection accuracy of lung nodules and serious misjudgment and missed judgment,the paper proposed an intelligent detection algorithm for lung nodules based on a 3D residual squeeze?and?excitation network.The network integrates the advantages of self-adaptive feature recalibration of squeeze?and?excitation network and residual network feature reuse,which can adapt to the feature requirements of different levels of networks,solve the problem of network degradation,and improve the overall representation ability of the network.The residual squeeze?and?excitation-based region proposal network achieved a sensitivity of 97.55%on the lung nodule detection task in the LUNA16 international public dataset,and effectively detected complex nodules such as isolated,vascular adhesion,and diaphragm adhesion nodules.The proposed method reduced misjudgments and missed judgments,improved the accuracy and efficiency of lung nodule detection,and provided effective technical means for early lung cancer screening.(2)To improve the accuracy of distinguishing benign and malignant lung nodules after detection and avoid the surgical injury caused by invasive biopsy patients,the paper proposed an intelligent diagnosis algorithm for lung nodules based on hybrid features.The algorithm innovatively combines texture features based on local binary patterns,shape features based on the histogram of oriented gradient,and depth features based on a 3D dual-path network to characterize benign and malignant lung nodules,effectively making up for the lack of single feature representation ability and presents strong robustness in benign and malignant classification tasks;the designed feature extraction network based on dual-path network integrates the advantages of aggregated residual transformations for feature reuse and a densely convolutional network for exploring new features,encouraging the model learns highly discriminative deep features;constructed a binary classifier based on gradient boosting machine,and achieved a classification accuracy of 93.78% in the diagnosis of benign and malignant lung nodules in the LUNA16 dataset,providing effective clinical decision-making reference for early screening of lung cancer.(3)If lung nodules develop into malignant tumors,they need to be intervened.In view of the low accuracy of manual segmentation of lung tumors and poor robustness of automatic segmentation,the paper designed an intelligent segmentation network for lung tumors based on the improved 3D recurrent dense UNet.The network innovates the nested connection mode between encoding and decoding features,supports capturing multi-dimensional deep features learned by UNet at different levels,and reduces the semantic feature gap between encoding and decoding.The loop recursion mechanism provides the decoder with more robust deep features;in addition,the thin plate spline method based on spatial flexible transformation was used to realize the expansion and enhancement of the training data,avoiding the over-fitting problem caused by the limited available data;The proposed network achieved a Dice score of 0.8316 on the lung tumor segmentation task based on the TCIA and LUNA16 datasets.The excellent segmentation performance showed important implications for monitoring lung tumor growth changes and developing treatment plans.(4)Accurate lung tumor and multi-organ segmentation results are a key prerequisite for formulating a high-quality radiotherapy plan.In order to break the training barriers between non-overlapping labeled datasets and solve the problem of multi-target joint segmentation of lung tumors and multiple organs,the paper innovatively established a conditional strategy to form a multi-target intelligent segmentation system with conditional nnUNet as the main body;for the first time,a semi-supervised uncertainty-aware teacher model was introduced to effectively utilize unlabeled data.The model encourages the segmentation network to make consistent predictions for the same input under different disturbances,and improves the overall robustness of the multi-objective joint segmentation network;designed personalized combined loss based on deep supervision,with the help of uncertainty-based consistency loss,to maximize the guidance of the network for high-quality multiobjective learning;the proposed method achieved satisfactory performance on the joint segmentation task of lung tumor and multiple organs,which lays a solid foundation for the development of an effective and executable radiotherapy plan.(5)Developed an intelligent diagnosis and treatment system for lung cancer.Designed clinical experiments based on actual cases,performed generalization ability tests on functions in intelligent detection and diagnosis of lung nodules,and intelligent segmentation of lung tumors and multiple organs;designed equivalent human model experiments to evaluate the feasibility and effectiveness of the robot-assisted intelligent diagnosis and treatment system.The experimental results showed that the intelligent diagnosis and treatment system developed in the paper can complete tasks such as early screening and diagnosis of lung cancer patients,and can provide accurate lung tumor and multi-organ segmentation results for the formulation of radiotherapy plans for lung cancer patients,and it has the potential to achieve minimally invasive radiotherapy for lung cancer patients with the help of image-guided puncture surgery robots,which has strong clinical application value. |
PDF Full Text Request