RCLM-net: A rich-club properties based lightweight deep neural network for multimodal sequence spine MR images segmentation

Abstract

The diagnosis of spinal diseases relies on the comprehensive assessment of multimodal sequence MR images. As a prior foundation for intelligent aid diagnosis, it is important to design an efficient MR image segmentation model. However, previous models for multimodal images tend to have a large number of parameters, which is not suitable for low-computing-power application scenarios, especially for primary medical sites. To address the above problem, this study constructs a lightweight neural network model for multimodal sequence spine MR images segmentation, named RCLW-Net (Rich-Club Lightweight Net), which is inspired by the “Rich-Club” properties of brain functional networks. Specifically, our network effectively reduced the number of model parameters by constructing a compact network structure that satisfies the Rich-Club properties and combining the Recurrent residual Convolution as the base convolution module. Verified by the five-fold cross validation of the three-modal sequence MR spine images containing 200 patients' data, proposed RCLW-Net has only 11 % of the parameter scale of the classical Unet++, but the segmentation performance is better. Moreover, proposed RCLW-Net can be successfully deployed on a lightweight embedded device Jetson Nano B01. The low number of parameters and the high performance of the RCLW-Net show the potential of the application in low-computing power scenarios.