Automatic volume classification in intravascular optical coherence tomography images

Intravascular Optical Coherence Tomography (IVOCT) is an emerging modality which could produce high resolution images, and would help cardiologists study coronary artery disease. Current IVOCT studies are mainly focus on classification or identification at frame level, while overlook the information provided by the entire pullbacks. In this work, we propose a volume based classification system which identifies the disease pullbacks automatically. Here, volume is defined as a set of consecutive frames within IVOCT pullback. In the proposed system, we first apply graph search based lumen segmentation method to detect lumen boundary and region-of-interest (ROI). After that, two texture features, Local Binary Patterns (LBP) and Histograms of Oriented Gradients (HOG) are extracted from each frame. Finally, we represent each volume using bag-of-words, and apply linear SVM for classification. A dataset with 9 healthy volumes and 32 unhealthy volumes is constructed to evaluate the system. In our experiment, each volume includes 20 consecutive IVOCT frames. Leave-one-patient-out cross-validation is employed for evaluation. Sensitivity, specificity and mean accuracy are computed to quantitatively evaluate the proposed system. The mean classification accuracy of 0.95 is reported, demonstrating the effectiveness of the proposed system.