Research On Segmentation And Detection Method Of Aortic Dissection Image Based On Deep Learning

Aortic Dissection(AD)is a dangerous cardiovascular disease.The cause of its formation is that blood flows from the aortic intimal gap into the aortic wall,which separates the artery wall into two regions,a true lumen and a false lumen.Any delay in treatment or misdiagnosis will aggravate the illness of the patient with aortic dissection,and even lead to the death of the patient.Clinical imaging methods are usually used to diagnose aortic dissection,such as Magnetic Resonance Imaging(MRI),Transthoracic Echocardiogram(TTE)and CT Angiography(CTA)Imaging methods,etc.CTA has fast imaging speed and powerful post-processing functions.It only requires CT scan of the aortic area where the contrast agent is injected.It is currently one of the preferred methods in clinical and emergency situations.Due to the complex clinical manifestations of aortic dissection and the lack of understanding of the disease by primary doctors,it is easy to cause missed diagnosis and misdiagnosis.Automatic segmentation of aortic dissection images can enable medical staff to quickly and clearly locate the aortic lesion area,ignoring the interference of non-lesion areas,and further realize the auxiliary detection of aortic dissection.Based on deep learning related technologies,this paper proposes three deep learning-based aortic CTA image segmentation algorithms,and realizes the detection of aortic dissection images.The specific work is as follows:(1)Aiming at the problem of the lack of public aortic image datasets,by collecting other types of biomedical image datasets,a transferable deep learning model is constructed on this basis.Combining the images of patients with aortic dissection and the control group,the aortic CTA image dataset is established,which lays a resource foundation for further research on image segmentation and detection of aortic dissection.(2)An image segmentation algorithm based on fully convolutional attention network(FCANet)is proposed,which can effectively aggregate the context information of long-range and short-range distances in biomedical images.FCANet adds two types of attention modules: spatial and channel to Res2Net with dilated strategy.The spatial attention module is used to aggregate the features of each spatial location,so that similar features can promote each other on the spatial scale.The channel attention module is used to emphasize the dependence between any two channel graphs.Finally,the weighted summation of the output features of the two types of attention modules preserves the feature information of the long-range and short-range distances,making the biomedical image segmentation more accurate.The effectiveness of FCANet is verified on three public biomedical image datasets,and the algorithm is transferred to the Aortic CTA dataset to realize the segmentation of the true lumen region of the aortic CTA image.(3)A segmentation and detection algorithm for aortic dissection based on CTA image is proposed.This algorithm can improve the detection efficiency of aortic dissection disease while avoiding the problem that manual features are not suitable for other data sets of similar lesions.Construct a semantic segmentation architecture based on U-Net and apply it to the Aortic CTA dataset to achieve automatic segmentation of the aortic true lumen region;perform aortic circularity analysis on the segmented image to obtain slice-level detection results;define the threshold and summarize the results of all slice-levels to further obtain patient-level test results.The proposed algorithm achieves89.60% sensitivity and 86.80% specificity on 1000 test images.The total running time is1 minute and 47 seconds.The average time for each slice image is 107 milliseconds,and the average time to detect each patient is 5.35 seconds.(4)An image segmentation algorithm based on dual dense u-structure network(DDU-Net)is proposed,which has strong generalization ability.The DDU-Net algorithm consists of two encoders and a decoder structure: the first encoder uses pre-trained Dense Net-121 as a fixed feature extractor;the second encoder uses a network structure similar to the first encoder,built and trained from scratch.Both encoders try to encode information on the input image,and each layer is directly connected to the next layer in a feed-forward manner.In the decoder,an up-sampling path suitable for the dual-encoder structure is designed.This path makes the network deeper through densely connected convolutional,and integrates the learning from the two encoders from low-level to high-level semantic features.Experimental results show that DDU-Net is superior to the benchmark U-Net and state-of-the-art(SOTA)segmentation algorithms in the segmentation of aortic CTA images,skin cancer images,nuclear images and lung images.