3D MRU-Net: A novel mobile residual U-Net deep learning model for spine segmentation using computed tomography images

The efficient and accurate segmentation of the spine shows the basis of spine analysis including visual insights, malfunctions, and fractures. The spine is made up of 33 vertebrae, 23 inter-vertebral discs, the spinal cord, and connecting ribs. Several existing deep learning models are used for the segmentation of the spine but required higher computational costs and hardware resources for the training process. In this research, a novel 3D MRU-Net (Mobile Residual U-Net) is introduced for the segmentation of the spine using CT scan images. The proposed model is an encoder–decoder-based architecture in which MobileNetv2 with residual blocks is used for feature extraction. MobileNetv2 is a lightweight model that decreases the computational cost with less trainable parameters while the residual block helps to learn deep features. Three separate modified MobileNetv2 are used for training on three different CT scan views (axial, coronal, sagittal). The output of these networks is concatenated to form a 3D feature map. The 3D U-Net is used as a decoder for spine segmentation. The verse 20 and verse 19 datasets are used for validating the proposed model. The result shows that the proposed model achieves a higher dice score with minimal computational cost as compared to the state-of-the-art methods.