Research On Deep Convolutional Networks Applied To Lumen And Media-Adventitia Border Detection In Intravascular Ultrasound Images

Currently being the leading causes of human death worldwide,accounting for 31%of all deaths,cardiovascular diseases(CVDs)include disease of the heart,vascular diseases of the brain and diseases of blood vessels.Coronary atherosclerotic heart diseases,often referred to as coronary heart diseases(CHDs)are caused by atherosclerotic lesions of coronary arteries,causing stenosis or occlusion of blood vessels,causing myocardial ischemia and hypoxia or necrosis.It is a comprehensive result caused by kinds of coronary artery diseases,among which 95%～99%are the coronary atherosclerosis,known as ’the first killer’.At present,two main medical imaging modalities applied for clinical diagnosis and treatment of CHDs include X-ray coronary angiography(CAG)and intravascular ultrasound(IVUS).In the clinical analysis and medical research applications of IVUS images,the delineation of media-adventitia and lumen borders in coronary arteries and the measurement of those related key parameters are not only important steps of the quantitative assessment of coronary atherosclerosis and vulnerable plaques,but also necessary prerequisites of the clinical diagnosis of coronary artery diseases and intervention treatment.The detection of media-adventitia and lumen borders in IVUS images is usually dependent on the clinician or the experienced researcher for manual recognition by naked eye,and delineation by hand.In clinical diagnosis and treatment,clinicians often need to analyze IVUS images ranging from a few frames to several hundred frames.The workload of manually delineating and recognizing lumen and media-adventitia borders is very large,and the segmentation results are easily influenced by the subjective factors of clinicians and the objective factors from the environment,therefore resulting in the analysis results of very large error between observers or within them and it’s difficult to form a uniform standard.In conclusion,automatic computer-aided IVUS image analysis based on machine learning,computer vision and other intelligent technologies is of great significance to deepen the researchers’ understanding of CVDs,improve clinicians’ work efficiency and clinical diagnosis and treatment of CHDs.The goal of this work is to propose two automatic algorithms for detection of the lumen and media-adventitia borders from IVUS images based on deep convolutional networks to help clinicians diagnose and treat coronary atherosclerotic heart disease.In view of problems of the existing automatic algorithms for detecting lumen and media-adventitia borders,this work first simulates the process of clinicians identifying the lumen and media-adventitia borders,and proposes an automatic algorithm combining a priori shape information of coronary artery and deep fully convolutional networks(DFCNs).In addition,in view of the fact that deep learning algorithms require a large amount of labeled data,another automatic detection algorithm for key tissue border delineation from IVUS images is proposed,which combines stacked hourglass networks(SHGNs)and generative adversarial learning.The above two border detection algorithms include the following four key components:1、Region segmentation in IVUS images based on deep fully convolutional networks(DFCNs).In chapter 3,deep convolutional networks as feature extractors and deep deconvolutional networks as shape reconstructors are used to construct DFCN-1 and DFCN-2 with the same architecture for lumen and media-adventitia borders in IVUS images.DFCN-1 used manual segmented images that contain lumen and non-lumen regions and original IVUS images as the training dataset for detecting lumen border;DFCN-2 applied manual labeled images that contain vessel and non-vessel regions and input IVUS frames as the learning dataset for delineating media-adventitia border.2、Border refinement based on a priori information of coronary artery.In view of the inevitable error problems in above region segmentation such as lumen or media-adventitia highland and depression,and other misclassification at the pixel level or small region level,this paper used a priori information of vessels--ellipse shape,to construct disk structures with adaptive radius,combining morphological close or open operations to remove depressions and highlands,respectively.Other misclassification is eliminated by some simple medical image processing methods.3、Multi-region segmentation in IVUS images based on stacked hourglass networks(SHGNs).We performed multi-region segmentation in IVUS frames by stacking two encoder-decoder structure hourglass networks,combining intermediate output and original IVUS image as input of next level network,jointly optimizing region segmentation and border refinement steps.During training,reconstruction loss based on L1 distance is suitable for detection of outline of media-adventitia,and L2 distance for lumen border.In addition,compared with Pix2Pix model using fully convolutional encoder-decoder structure or U-Net as generator in generative adversarial networks,stacked hourglass networks proposed in this work has some advantages such as less model parameters and high precision of border detection.4、Adversarial learning based on generative adversarial networks(GANs).On the basic of reconstruction loss,introducing adversarial loss,the above stacked hourglass networks not only generate IVUS image segmentation map that is closer to the clinicians’ manual result,but also relieve the difficulty of deep convolutional networks requiring large-scale labeled medical ultrasound images.The main experimental data from clinical practice in this paper are 435 20MHz IVUS images from Set B,an international standardized publicly available IVUS database.In cross-dataset validation and clinical practice analysis,we used 100 IVUS images from the Department of Cardiology in Guangzhou General Hospital of Guangzhou Military Command.Based on the linear regression and Bland-Altman analysis,the evaluation metrics of JM,PAD,Hausdorff distance(HD)and Average distance(AD),the automatic detection algorithm for IVUS target borders on the basic of deep convolutional networks was quantified.The experimental results show that the correlation between the detection results of the algorithm and the results of manual tracing reaches up to 0.94.Over 94.51%of results is within 95%confidence interval.The AD errors between automatic and manual methods are 0.07mm and 0.08mm,while the HD errors are 0.21mm and 0.30mm,respectively.JM and PAD metrics of lumen border are 92%and 5%.JM and PAD indexes of media-adventitia border are 93%and 4%.Compared with the existing international algorithms,this method improves the recognition of various plaques,acoustic regions and collateral vessels,and accurately and reproducibly detectes the key target borders in the IVUS image without being affected by the ultrasonic speckle noise and other factors.For further evaluation of border detection algorithm based on the conditional generation adversarial networks(C-GANs),the statistical results are:JM 93%,PAD 3%,HD 0.19mm for lumen border,and JM 95%,PAD 3%,HD 0.16mm for media-adventitia border.These metrics meet clinical diagnostic requirements and show that the proposed method outperforms existing eight algorithms in recent years,as well as Pix2Pix model based on C-GANs and double sparse auto-encoders based on patch classification.In conclusion,the two lumen and media-adventitia border detection methods based on deep convolutional networks proposed in this work are simple but effective,and accurately detect borders of interst and importance in IVUS images.