RPM-Net:融合多路注意和混洗的胸腔多器官分割算法

RPM-Net: thoracic organs segmentation algorithmbased on multiway attention and shuffle

为了提高肿瘤周围健康器官分割的准确性,提出了基于U-Net结构,结合多路注意和像素混洗的网络模型:RPM-Net.网络使用融合转换残差模块,来捕获器官完整的空间背景.像素混洗模块作为上采样部分以得到高分辨率图像信息.在解码层使用多路注意融合模块,进一步提取器官的判别特征.在ISBI 2019 SegTHOR挑战赛中,对40个胸部多器官训练样本进行分割,以Dice系数和HD距离作为主要评判标准,该方法在测试样本中食道、心脏、气管和主动脉的Dice系数分别达到0.824 1、0.942 7、0.904 2和0.929 9,HD距离分别为0.401 1、0.176 1、0.224 7和0.271 6.实验结果表明:该算法在胸部多器官分割效果上更具竞争力.

To improve the accuracy of segmentation of healthy organs around tumors,a segmentation algorithm based on U-Net structure combined with pixel shuffling and attention fusion is proposed.The network uses the aggregated residual transformations module to capture the complete spatial background of the organ.Pixel shuffling module is used as an up-sampling part to obtain high-resolution image information.The multi-channel attention fusion module is used in the decoding layer to further extract the discriminative features of the organ.We segment forty thoracic multi-organs training samples in the ISBI 2019 SegTHOR Challenge and take Dice coefficient and HD distance as the main criteria.The Dice coefficients of the esophagus,heart,trachea,and aorta in the test samples are 0.8241,0.9427,0.9042,and 0.9299,respectively,and the HD distances are 0.4011,0.1761,0.2247,and 0.2716 respectively in this method.Experimental results show the algorithm is more competitive in segmentation of thoracic organs at risk.