2.5D lightweight RIU-Net for automatic liver and tumor segmentation from CT

•Extract inter-slice spatial information in the form of 2.5D.•Proposes a lightweight Inception convolution structure with residual connections to significantly reduce the network’s parameters.•Employ a combination of BCE and Dice loss to achieve fast convergence and low fluctuations in network training.•Evaluate the proposed method on publicly available datasets, LiTS17 and 3Dircadb. One critical factor that restricts the clinical application of computer-aided liver and tumor segmentation is the method's high complexity and low accuracy. Overcoming this limitation is what we are concerned about in this study. This paper presented a new 2.5D lightweight network for fast and accurate liver and tumor segmentation from CT images. The method is grounded in the U-Net framework, which leverages the techniques from the residual and Inception theories. We first adopted the 2.5D training mode for CNN networks to improve the utilization of spatial information. Then, we designed an improved U-type architecture to substantially reduce the parameters by introducing residual block and InceptionV3, named RIU-Net. Finally, a hybrid loss function combined BCE and Dice is employed to speed up the convergence and improve accuracy. We evaluated the proposed method on two publicly available databases, LiTS17 and 3DIRCADb. The performance of our approach is compared with five closely related techniques. Our result outperforms the others on both accuracy and time cost. Specifically, the total number of parameters is reduced by 70% compared to U-Net. Both quantitative and qualitative results demonstrated the superior applicability of our method and thus proved to be a promising lightweight tool for computer-aided liver and tumor segmentation..