Computer-aided Diagnosis of Peritonitis Using Two-Stream Attention Deep Convolutional Network

Toshiki Kawahara, Akitoshi Inoue, Yutaro Iwamoto, Bolorkh Batsaikhan, Svohei Chatani, Akira Furukawa, Yen-Wei Chen

Abstract: Cine magnetic resonance imaging (MRI) analysis methods are used to evaluate intestinal peristalsis. However, the evaluation of intestinal peristalsis by MRI is subjective, time consuming, and non-reproducible, which are recognized as a critical issue that needs to be resolved. In our previous work, we used deep optical flow network (DOFN) to extract temporal-spatial features of intestinal movements and differentiate peritonitis from intestinal peristalsis. Since the DOFN is based on difference of two neighboring image frames, it lacks texture and spatial information of the small bowels. To solve these problems, we also proposed a Two-Stream Deep Spatial-Temporal Convolutional Network, which consists of optical flow stream (i.e., DOFN) and dynamic image stream. The Two-Stream DCSTCN, which extracts both spatial (texture) and temporal information, achieved an improved accuracy. To further improve the classification accuracy, we propose a Two-Stream Attention Deep Convolutional Networks (Two-Stream Attention DCN). The proposed method is an improved version of our Two-Stream DCSTCN by introducing the attention modules to both streams in order to improve results. Our experiments show that our proposed method improves the performance by about 6%.