| With the rapid development of medical imaging and computer vision,the application of computer vision methods for processing medical images is becoming increasingly widespread.Medical image analysis typically involves tasks such as object detection and segmentation,with CT imaging being the primary focus of research in medical image analysis.As an important component of medical imaging,medical CT images are three-dimensional images obtained by CT machines that are not limited by depth of field.The image is composed of multiple slices of a particular part of the human body,with the slice spacing determined by the CT machine’s sampling interval.This paper focuses on the task of medical CT image segmentation.The medical CT image segmentation task can be divided into two types: organ segmentation and lesion segmentation.In organ segmentation tasks,the same type of human organ usually has similar size,shape,and density,exhibiting strong commonalities in imaging features.The segmentation accuracy of the vast majority of human organs has reached a high level.Accurately segmenting organs at a high level can assist radiologists in monitoring organ status,detecting organ abnormalities,and providing diagnostic recommendations.In lesion segmentation tasks,target lesions typically exhibit less commonality,varying sizes,and dispersed locations.Currently,there is still significant room for improvement in the segmentation accuracy of lesions.Successful lesion segmentation is crucial for determining lesion types and conditions,monitoring lesion progression,and providing treatment plans.This study designs a segmentation algorithm based on a fusion convolutional neural network and attention mechanism for these two segmentation tasks.Specifically,the main research contents of this study include:1.This paper designs a novel image preprocessing method suitable for three-dimensional convolutional neural networks(3D CNNs)to optimize the performance of 3D medical image segmentation tasks.Due to the high memory consumption of 3D CNNs in processing threedimensional data,it is often necessary to resize the input images.Additionally,CT sampling intervals result in unequal pixel spacing,so this study applies preprocessing methods such as resampling,normalization,and cropping to the images.Moreover,the study introduces an innovative normalization threshold algorithm that compares the Hounsfield Unit(HU)value ranges of lesions and background,aiming to optimize the contrast.This algorithm accurately distinguishes between lesion and background HU values and improves the clarity of boundary information.2.This paper designs an attention gate structure module for the V-Net,a three-dimensional convolutional neural network,that uses deep-level features from the decoder layer to weight shallow-level features from the encoder layer in skip connections.The module uses a crossattention mechanism and sigmoid function to generate cross-attention coefficients,similar to a gate circuit.This method is used to complete lung nodule segmentation and lung bullae segmentation tasks in lung CT images.Experimental results show that this method can improve segmentation accuracy,with a approximately 1% increase in Dice Loss and a 4% increase in recall.3.This paper designs a multi-view network structure that combines a fusion convolutional neural network and Swin-Unet,aiming to comprehensively utilize the ability of convolutional neural networks to extract detailed information and the ability of Swin Transformer to extract overall information.This network structure treats the features obtained by these two feature extraction methods as two views and fuses them into the connection used to pass shallow features through skip connections,thus enriching the semantic features of the network structure.This method was experimentally tested in multiple organ segmentation tasks and achieved an improvement in Dice score.In actual segmentation results,the edge segmentation was improved. |
