| Cancer is globally recognized as one of the most significantly public health problems.In China,lung cancer ranks first in terms of incidence and mortality of various cancers.Studies have shown that lung nodules are one of the main clinical features of most lung cancers in their early stages,and screening for lung nodules in the early stages of lung cancer can be effective in improving treatment outcomes and increasing survival rates of patients.Meanwhile,since 2023,countries around the world have entered a regular phase of the new crown(COVID-19)epidemic.In clinical practice,patients with COVID-19 may present with ground glass nodules on CT in addition to fever or respiratory symptoms,and therefore the introduction of medical imaging technology into the screening of COVID-19 has largely facilitated the confirmatory screening and timely treatment of patients and suspected patients.In this paper,a convolutional neural network-based approach for pulmonary nodule detection on spatially sequenced 3D CT images is investigated,combining clinical medical priori knowledge and CT data intrinsic characteristics,with the aim of improving the sensitivity and robustness of lung nodule detection.Research elements and innovations are summarized as follows.(1)When making a medical diagnosis,clinicians will first focus on the visual contrast cues and contour features of lung nodules in order to locate the specific spatial location of the nodules.Based on this clinical medical a priori knowledge,this work designs an efficient multi-task dual-branch 3D convolutional neural network for lung nodule detection and segmentation.In the two-branch structure,one branch is designed for candidate region extraction for lung nodule detection,while the other merged branch is used for semantic segmentation of lesion regions of lung nodules.In this paper,a 3D attention-weighted feature fusion module is developed based on the physician’s clinical diagnosis perspective so that the capture information obtained from the designed segmentation branch can further contribute to each other to enhance the effectiveness of the respective detection branches.In addition,a false positive reduction strategy is also proposed,developing a 3D salient region pooling layer for improving the feature maps that produce narrow receptive fields.The strategy shares several feature extraction modules designed to ensure that the network has a narrow receptive field for capturing detailed feature representations.It essentially differentiates itself from region candidates and promotes robustness in lung nodule detection.Experiments and evaluations were conducted on commonly used medical image analysis datasets,and experimental comparisons showed that the designed framework outperformed other typical mainstream methods.(2)In conventional clinical practice,due to the rich spatio-temporal background information of CT scans,physicians have always focused on correlation between multiple consecutive slices of CT images to determine the specific location of lung nodules for the corresponding disease diagnosis.In order to achieve intelligent assisted diagnosis through automatic lung nodule detection in CT scans that takes full advantage of this unique spatiotemporal feature of clinical experience,this paper develops a 3D multi-attentional convolutional neural network and designs a 3D multi-attentional module to simulate the clinical diagnosis process of physicians,which exploits the attentional enhancement of spatio-temporal views of CT slices in multidimensional directions and aims to integrate spatio-temporal contextual information.Through the OHEM(Online Hard Example Mining)loss calculation method,the weights of positive and negative samples can be effectively adjusted,thus solving the uneven distribution of positive and negative sample,and the discriminative features of positive and hard-to-classify samples can be effectively learned,which in turn effectively enhances the generalization ability and classification sensitivity of the model. |
PDF Full Text Request