Srcnn: Cardiovascular Vulnerable Plaque Recognition With Salient Region Proposal Networks

Vulnerable plaques recognition from IVOCT images is a valuable yet challenging task for computer-aided diagnosis and treatment of cardiovascular diseases. However, most existing supervised methods only used one kind of annotation information, and so they didn't fully and effectively utilize biomedical image information. In this paper, we propose a single, unified salient-regions-based convolutional neural network (SRCNN) to address this challenging task. The proposed SRCNN takes advantage of multi-annotation information (i.e., classification labels and segmentation labels) and combines prior knowledge of cardiologists. Our contributions in this paper are as follows: (i) We employ a bi-branch network combining the annotation information of classification and segmentation to recognize vulnerable plaques in IVOCT images. (2) According to prior knowledge of cardiologists, we construct a salient region proposal network (SRPN) that can propose irregular salient regions different from bounding boxes. (3) We embed SRPN in the bi-branch network through an appropriate merging strategy, and call this new bi-branch network SRCNN. Our proposed SRCNN is evaluated on the 2017 CCCV-IVOCT Challenge dataset. And ablation experiments demonstrate that compared to separate networks, the bi-branch network can improve the performance of classification and segmentation simultaneously. Furthermore, they also show SRPN contributes to extracting more discriminative features and boosting the whole performance of recognizing vulnerable plaques in IVOCT images greatly.