Detection And Analysis Of CTA Coronary Plaque Based On Deep Learning

Calcified plaque is a relatively stable coronary plaque,which has the characteristics of different sizes and structures,and is not easy to break and fall off.Accurate detection of calcified plaque is of great significance to the diagnosis and treatment of cardiovascular diseases.Coronary CTA(Computed Tomography Angiography)images can clearly show the location,size and shape of plaques,and has the advantages of non-invasive,fast and high detection accuracy.However,detection of calcified plaques using coronary CTA images still faces the challenge of class imbalance problem and obfuscation of interference regions.Existing methods usually adopt heart segmentation to remove redundant regions in the image,and then further detect calcified plaques.This approach pays insufficient attention to the coronary vascular region,which may lead to missed or misdiagnosed lesions.In addition,this method also fails to make full use of the lesion information between multiple views,which will reduce the accuracy of detection results.In order to solve the above problems,this thesis first created a joint coronary CTA data set,which is composed of data from Shandong Cancer Hospital and or Ca Score(Coronary Artery Calcium Scoring)challenge data,a total of 109 people.Among them,36 people were marked as normal,and 73 people were marked as containing calcified plaque.Secondly,this thesis uses deep learning technology to design a two-stage algorithm for detecting calcified plaques.(1)In the first stage,redundant interference regions are removed by 2.5D dynamic attention network and fuzzy boundary selection method.First,the coronary CTA was converted to the coronal,sagittal and horizontal plane views,and the dynamic attention network classification method was used to retain the slices containing coronary vessels in the three views.The dynamic attention network can dynamically select the network layer to add the attention mechanism according to different coronary artery slices,improve the calculation efficiency and accuracy of the model,and realize the accurate classification of coronary artery slices.Then,the rough borders of the three views including the coronary vessel slices are obtained by the fuzzy border selection method,and the borders are used to segment the coronary CTA images,effectively removing interference regions such as ribs and descending aorta calcification,and narrowing the region of interest for plaque detection.(2)In the second stage,accurate detection of calcified plaques is realized through a 2.5D multi-view fusion network.Firstly,the segmented coronary CTA images are converted to three-view angle again,and the image texture details are increased by using the image enlargement strategy.Then imitate the doctor’s diagnosis process,take the horizontal plane as the main viewing angle,and use the coronal plane and sagittal plane as the auxiliary viewing angles to construct the main information extraction network and auxiliary information extraction network to extract plaque lesion information of different scales,and finally adopt the multi-view fusion strategy Acquire calcified plaque lesions to achieve accurate positioning of the plaque area.Finally,this thesis designs a prototype system for intelligent detection of coronary CTA plaques.The system realizes fuzzy boundary prediction and plaque intelligent detection functions.Doctors can select cases to upload according to needs,and can return the fuzzy boundary and plaque detection results predicted by the model in real time.The system can provide doctors with an auxiliary diagnosis basis,effectively shorten the doctor’s diagnosis time,and improve the doctor’s diagnosis efficiency.