| At present,cardiovascular disease is one of the highest mortality diseases in humans,the most common of which is coronary heart disease,intravascular ultrasound(IVUS)uses ultrasonic reflection to obtain cross-sectional gray-scale maps of blood vessels,manually annotates the inner and outer membranes of blood vessels through experts,enables quantitative diagnosis of the area of stenosis as well as judgment of the nature of vascular plaques.To arrive at a reliable preoperative diagnosis,it usually requires that experienced interventional specialists spend a significant amount of time manually annotating the images,so automated and precise detection algorithms are increasingly required.However,IVUS images usually present with varying degrees of noise interference,resulting in extreme blurring of the inner membrane boundary,and automated boundary detection algorithms remain a challenging subject.In this paper,we systematically elaborate on the historical development and advantages and disadvantages of cardiovascular intimal boundary detection algorithms for analyzing IVUS images,and propose new methods for the automatic detection of vascular intimal boundary on IVUS images based on the shortcomings and difficulties of existing methods: combining artificial features and higher-order semantic features,the most significant feature subset is screened out by the modified cuckoo search,which input to dictionary learning for classification and boundary detection,specifically contains the following three parts:1.Dictionary learning based vascular intimal boundary detection on IVUS images.Considering the relatively similar pixels in lumen and non-lumen regions on IVUS images,and the characteristics that there are many noise interferences in the inner membrane boundary: using dictionary learning and sparse coding,and using radial basis kernel function to map the features into high-dimensional space,it is better to distinguish the non-linear similarity,meanwhile excluding the noise interference outside the echogenic bright ring by a series of efficient preprocessing,better enhancement of detection accuracy.2.U-net based cardiovascular intimal boundary detection.In the presence of noise interference such as vessel bifurcation,the artificial features have slightly poor discrimination ability to similar pixels,while the deep network has strong semantic expression ability,and after contrast of the three semantic segmentation networks suitable for medical image analysis,u-net was used to perform intravascular boundary detection on IVUS images.3.Endovascular boundary detection based on u-net semantic features and artificial features.Artificial features are difficult to mine the hidden information that people cannot perceive,and insufficient u-net training pictures need the assistance of artificial features,so more artificial features are adopted,and higher-order semantic features of u-net are extracted,and the two are combined into mixed features.In order to prevent the problems of redundancy,slow computation and low accuracy caused by too many features,the subset of features with the largest contribution was selected by the modified cuckoo search,input to dictionary learning for classification,and edge segmentation and postprocessing were performed to obtain the final vessel intima boundary.In this paper,published datasets presented at the MICCAI Congress and three experimental outcome assessment methods were used: Jaccard(JACC),Hausdorff distance(HD)and percentage of the area difference(PAD).The experimental results(JACC: 0.88,HD: 0.36,pad: 0.06)were better than all the other participants(JACC:0.77-0.84,HD: 0.38-0.51,PAD: 0.11-0.16),and the results of the semiautomatic segmentation were almost the same(JACC: 0.88,HD: 0.34,PAD: 0.06),at the same time and compared with the methods on the same dataset in recent years,which also achieved the best results on some indicators.Meanwhile,through self-contrast experiments,the subset after feature selection improved JACC by 5% and PAD by 3%compared with the full mixture of features;Compared with u-net,JACC improved by7%,HD by 1%,and PAD by 8%.In conclusion,the proposed method for automated detection of vessel intima borders on IVUS images enables accurate and effective detection of vessel intima borders on IVUS images. |
