Intravascular Optical Coherence Tomography Image Segmentation Based On Pattern Recognition Algorithm

Timely and effective diagnosis of coronary artery disease(CAD)has long been a concern in the field of cardiovascular disease.In recent years,the intravascular optical coherence tomography system(IVOCT)has been rapidly accepted by more and more clinicians.The IVOCT system resembles the intravascular ultrasound system(IVUS),but its ultra-high imaging resolution(10-20 ?m)enables it to accurately observe the lumen wall of the coronary artery.Although the American Heart Association has published consensus papers on IVOCT imaging characterization,there are still a lot of challenging issues in clinical diagnoses,e.g.hard to interpret and segment the plaque components,time-consuming in reading the images.This work attempts to automatically identify and segment several typical components of atherosclerotic plaque,such as the fibrous tissue,the calcified tissue,and the lipid tissue.The expected outcome is to provide convenient and supplementary information for clinical diagnostic reading.The pre-processed IVOCT images were used for analysis and segmentation in this work.All image data come from clinical patient examinations.The pre-processing consists of several steps: filtering,removing guide wire artifacts,removing imaging catheter artifacts,using the OSTU algorithm to extract the vessel wall and determining the region of interest(ROI)of the subsequent analysis.A data augmentation was implemented to the preprocessed data.Then,attenuation coefficient and texture features based on the gray level co-occurrence matrix was extracted from each image.With the acquired features,the experiment uses Support Vector Machine(SVM)for tissue recognition and classification.We analyzed the performance of this method according to different parameter settings.Some feature selecting method were used to improve the input of SVM classifier.Limitations of this method were also discussed.The experimental results showed that the 252-dimensional features obtained from the log-transformed image performed better than the features extracted from the image without log transformation.For feature selection,model-based feature selection performs better than other methods.For classifier,random forests acquired an accuracy of 82.5%.Meanwhile,support vector machine gained a 84% accuracy.In conclusion,the segmentation method based on pixel-by-pixel classification is well-adapted to this task and is a considerable way for assisting clinical IVOCT image reading.