Last modified: 2017-05-13
Abstract
Accurate cap thickness quantification is of fundamental importance for vulnerable plaque research. A segmentation method for intracoronary optical coherence tomography (OCT) image based on least squares support vector machine (LS-SVM) to characterize plaque lumen surface, segment borders and fibrous cap for plaque cap thickness when image quality is not high enough, especially at the location of bifurcation.
In vivo intravascular ultrasound (IVUS) and OCT coronary plaque data were acquired from one patient with informed consent obtained. Manual segmentation in OCT images based on the combination of VH-IVUS image and OCT image were given by experts as the gold standard. Processed OCT images were trained and tested via LS-SVM by two methods (M1 and M2). In M1, 500 pixels were randomly selected from each lipid class and vessel tissue class for 9 OCT images. The training data set would be the feature vectors from 9000 pixels. In M2, a procedure similar to leave-one-out cross validation was employed as any 8 out of 9 images were used as training data while the remaining one as the testing data. Borders and lipid contours were extracted from prediction results for cap thickness. Virtual histology (VH) IVUS data were processed with minimum cap thickness set as 50 and 180 micron to generate IVUS50 and IVUS180 data sets, respectively. Cap thickness from manual segmentation, predictions from M1 and M2 based LS-SVM, IVUS50 and IVUS180 data sets were compared.
The accuracy of M1 and M2 were above 76%. Average of mean cap thickness (unit: mm) from 9 images was 0.561 (manual), 0.470 (M1), 0.463 (M2), 0.128 (IVUS50) and 0.204 (IVUS180). Average of minimum cap thickness (9 slices) was 0.390 (manual), 0.288 (M1), 0.282 (M2), 0.040 (IVUS50) and 0.165 (IVUS180). IVUS50 and IVUS180 underestimated cap thickness. The mean cap thickness from prediction were close to manual results (error<18%). The point-point cap thickness from five groups showed that the prediction based LS-SVM had agreement with manual segmentation.
Conclusion. The segmentation methods based on LS-SVM provided reasonable accuracy for plaque cap thickness quantification. More data sets and better gold standard are needed for further improvement.