Segmentation of MR Spinal Images Using a 2 - Stage Semantic Framework

Segmentation of Spine with the 2 - Stage Semantic framework involves the Vertebrae and the Inter-vertebral discs components of the spine, helps in the identification of the spinal disorders. The methods using the Convolutional Neural Network does not precisely identify the interdependencies within the Vertebral and the Inter-vertebral disc components of the spine. The proposed method helps achieve Segmentation of MR Spinal images in two stages. The Coarse-Grained Segmentation utilizes the 3D-Graph Convolution Segmentation Network and the Segmentation Enhancement is carried out using the 2D-ResUNet. In Coarse-Grained Segmentation, the Spatial Pooling technique transforms the image into a graph-based representation, where individual spinal structures are represented as nodes. This graph undergoes three layers of graph convolutions to extract Semantic Features. Following this, the Inverse pooling process is applied to transform the graph back into a Visual Semantic Feature map. Experiments are carried out using the T2-weighted volumetric MRI data of 215 individuals[1], and achieved DSC values 88.56 ± 3.36%, 87.59 ± 3.54%, and 88.07 ± 5.25% for the Vertebrae, IVDs and the complete set of nineteen spinal components respectively.