CARNet: Automatic Cerebral Aneurysm Classification in Time-of-Flight MR Angiography by Leveraging Recurrent Neural Networks

Cerebral aneurysms (CAs) detection from unenhanced 3D time-of-flight magnetic resonance angiography (TOF MRA) images is time-consuming, laborious, and error-prone. In this paper we propose a novel architecture, Cerebral Aneurysm Recurrent Classification Network (CARNet), which integrates the spatial information over multi-view Maximum Intensity Projection (MIP) images by exploiting recurrent neural network (RNN). Specifically, CARNet first collects the region of interests (ROIs) around the aneurysms in 3D TOF MRA data via the conventional sliding window strategy. Then it detects CAs in MIP images from ROIs along 9 fixed planes via Maximum Intensity Projection which helps reduce computational cost. Afterwards CNN-GRU Aneurysm (CGA) Discrimination network recursively renews the high-level features of aneurysm extracted by CNN from all MIP images to make a decision by employing the RNN. Finally CARNet was evaluated on 213 patients of 480 samples with aneurysm acquired from the Radiology Department of Wuhan Union Hospital. Experimental results showed that CARNet outperforms the previous methods with a sensitivity of 85%–91%. In addition, the efficiency of CARNet is about twice that of 3D CNN.