Research On Automated Detection Of Plaque Erosion By Intravascular Optical Coherence Tomography

Plaque erosion is the second most frequent underlying mechanism of acute coronary syndrome(ACS)and the important pathological basis of acute coronary thrombosis.Optical coherence tomography(OCT),as a high-resolution intracoronary imaging modality,has been proved to be the only clinically available imaging modality that can diagnose plaque erosion in vivo.However,the detection and diagnosis of plaque erosion still demand physicians with rich clinical experience,and are subject to diagnostic errors due to potentially high inter-and intra-observer variability.Therefore,automated detection of plaque erosion in vivo can effectively reduce the difficulty and improve the accuracy,which is of great significance for the clinical diagnosis of ACS.This thesis improves the performance of medical image detection algorithms by incorporating clinically interpretable medical features,and develop a fully automated method for detection of plaque erosion in vivo for the first time,finally achieving a sensitivity of 0.800±0.175 and a precision of 0.734±0.254 compared with human experts.The main research works of this thesis are as follows:1)A plaque erosion dataset was constructed which filled the gap in the field,consisting of 83 OCT pullbacks,29914 cross-sectional images from 83 ACS patients with plaque erosion labeled by three experienced physicians.2)A shaped-encoded convolutional neural network(CNN)Mask RCNN-CK model was proposed to address the limitations of existing deep learning methods in the detection of plaque erosion.First,plaque erosion usually does not exhibit lateral boundaries on the side away from the lumen and has textural similarity with fibrous tissue.Second,plaque erosion on the luminal side shows irregular luminal surface on OCT images,while CNNs are more concerned with texture information of images,the image detection results obtained by using existing deep learning algorithms are not desirable.In this thesis,a novel shape-encoded CNN model was designed to combine two shape features convexity and curvature into the Mask RCNN architecture,and the precision of plaque erosion detection was improved by 18.2%.3)A post-processing algorithm based on domain knowledge was proposed.Considering that CNNs only use the information of original OCT images without combining the medical domain knowledge,the proposed post-processing algorithm further extracted optical and morphological characteristics of tissue,the threedimensional continuity,and others based on preliminary experiments.It generated a significant visual improvement on the final results,which was also reflected numerically by the significant increase in the precision and sensitivity.