Deformable Deep Network Atherosclerotic Coronary Plaque Recognition of Oct Imaging

Abstract

Cardiovascular disease results great life and economics threat to the patients and their family members around the world. Optical coherence tomography (OCT) image not only obtains higher resolution and faster image modality to assess coronary vessels, but also provides safer guidance for micro scale medical interventions. Deep learning network frameworks have been considered as a promising approach for pathology feature classification and segmentation of computerized diagnosis. In compare with universal dataset for object recognition, atherosclerotic coronary plaque of different patients usually involves different pathology character and formation modality. Limited samples and weakly labeling always cause plaque mis-classification and evaluation. In order to improve recognition accuracy, a novel regional based fully convolutional network with deformable deep networks has been developed to realize plaque object recognition for OCT images. The deformable convolution and regional of interest pooling can adapt to the anchor scales and offsets change to improve classification precision for different pathology character and formation modality of different patients. Recognition experiments from atherosclerotic coronary plaque image have validated the effectiveness and performance of proposed method.