Self-guided Multi-attention Network for Periventricular Leukomalacia Recognition

Recognition of Periventricular Leukomalacia (PVL) from Magnetic Resonance Image (MRI) is essential for early diagnosis and intervention of cerebral palsy (CP). However, due to the subtle appearance difference of tissues between damaged and healthy brains, the performance of deep learning based PVL recognition has not been satisfactory. In this paper, we propose a self-guided multi-attention network to improve the performance for classification and recognition. In particular, we first conduct semantic segmentation to delineate four target regions and brain tissues as regions of interest (RoIs), which are pathologically related to PVL and should be focused in terms of the attention of the classification network. Then, the attention-based network is further designed to focus on the extracted PVL lesions when training the network. Moreover, the novel self-guided training strategy can provide comprehensive information for the classification network, and hence, optimize the generation of attention map then further improve the classification performance. Experimental results show that our method can effectively improve the precision of recognizing PVLs.