Segmentation Of Coronary Plaques In CT Image Based On 3D-Unet

Segmentation of Coronary Plaques in CT image can play a good role in the clinical diagnosis of coronary artery disease.It has very important significance and good application prospect to study the plaque segmentation technology of 3D coronary CT image based on 3D-Unet neural network.Due to the varied shapes of coronary arteries in CT images,it’s necessary to extract the centerline of coronary arteries and rebulid it in multiple planes to a straight line.Because the centerline of straightened coronary artery is not smooth,a method of median filter and cubic spline interpolation is proposed to smooth it.Two segmentation methods are designed for calcified and non calcified plaques: for calcified plaques that are easy to be segmented,3D-Unet network is used for direct segmentation;for non calcified plaques that are difficult to be segmented,3D-Unet network is used for rough segmentation and location first.Then an improved 3D UNET network is designed to segment the non calcified plaque more accurately.According to the location results,the CT images of coronary artery with non calcified plaque are cut nine times.Finally,the non calcified plaque was segmented.Attention mechanism is introduced into 3D UNET network structure to improve the recognition ability of non calcified plaque.In the new loss function,we consider various forms of non calcified plaque,so as to improve the segmentation accuracy of the model.Finally,the segmentation results are mapped back to the original CT image.The experimental data show that the 3D-Unet network has a good segmentation effect on calcified plaque,and the average Dice coefficient is 0.8706.The average Dice coefficient of the improved 3D-Unet network in the non calcified plaque segmentation is also 0.6642,which is significantly improved compared with that before the improvement,verifying the effectiveness of the improved network.