Research On CTA Carotid Plaque Segmentation Algorithm Based On The Deep Convolutional Neural Network Unet

Stroke is one of the diseases that seriously threaten human health,and carotid atherosclerotic lesions are an important cause of ischemic stroke.Nowadays,a variety of imaging technologies for diagnosing carotid atherosclerotic diseases have been developed.Computed Tomography Angiography(CTA)has been widely used in clinic because of its fast scanning speed,high resolution,and non-invasiveness.The location and segmentation of carotid lumen and plaque based on CTA images is an important step in the assessment of carotid atherosclerotic diseases.Traditional segmentation methods cannot get rid of the dependence on manual labor and have poor robustness,and cannot achieve automatic segmentation.Under the above background,this paper studied the automatic segmentation method of carotid artery lumen and plaque based on deep learning.The main research contents and results are as follows:(1)A process of carotid lumen segmentation from coarse to fine(C2F)was proposed.In view of the serious imbalance between the target area and the background area,a coarse-to-fine segmentation process was designed.First,the carotid artery lumen area was roughly positioned in the coarse segmentation stage,and then the cascaded fine segmentation network refined the edge segmentation of the coarse segmentation results.Register and preprocess the head and neck CTA scans and expert annotations provided by the hospital to construct training set and test set.Based on the constructed data set,the Dice accuracy of five different segmentation networks under the C2 F process had increased by 9.81%-16.11%.It shows the effectiveness of the C2 F segmentation process,which can realize automatic segmentation and avoid resolution loss.(2)Multiplanar D-SEA UNet was proposed in the carotid artery lumen segmentation network.A multi-plane data augmentation preprocessing architecture for 3D images was adopted,and a deep supervision mechanism,a squeeze and excitation module and an attention mechanism were added to the structure of UNet,which were used to fuse multiscale segmentation output and weight feature channels.It could reach 91.51% of the segmentation accuracy,and could ensure the integrity and continuity of the blood vessel.Each evaluation index was better than the comparison segmentation networks.(3)A 3D Trans-IS UNet plaque segmentation algorithm was proposed.In view of the small size of plaque and small sample data sets,a patch-based sampling strategy was adopted for data amplification.The Transformer model was added to the segmentation network for global information modeling,and the intermediate supervision mechanism was integrated to make the network fully trained.At the same time,the loss function design was combined with Dice and cross entropy.According to the results of carotid artery lumen segmentation,plaque segmentation was performed on the left and right carotid artery regions.3D Trans-IS UNet could achieve a segmentation accuracy of74.47%,which was significantly better than other segmentation networks,and could detect and accurately locate plaques.(4)Prospective clinical verification.Forty-seven cases with no plaques or plaques were collected from the hospital for verification,and the accuracy of plaque detection was 80.85%.If a plaque with a volume pixel less than 50 is used as a non-plaque sample without treatment,the detection accuracy is 89.36%,which can assist doctors in initial rapid image screening.There are 33 pictures,12 tables,and 58 references.