Application of Hybrid Network of UNet and Feature Pyramid Network in Spine Segmentation

Spine segmentation is a common task for spinal imaging and spinal surgical navigation. Spine segmentation provides valuable information for the diagnosis, and the segmentation output can also serve as an input for downstream surgical navigation. Unfortunately, spine segmentation is a labor-intensive task. In this study, we applied a deep network combining feature pyramid network (FPN) and UNet to the segmentation of vertebral bodies (VBs), referring as Res50_UNet. Compared with the original UNet, Res50_UNet has the following enhancements: 1) five consecutive spine MRI slices and two coordinate maps are concatenated as the input; 2) the convolutional block from ResNet are used; 3) an FPN architecture is applied to extracting rich multi-scale features and obtaining segmentation output. Experiments were conducted on an annotated T2-weighted MRIs of the lower spine dataset. We have benchmarked Res50_UNet against UNet and other UNet based network structures. It was found that Res50_UNet needs the lowest number of epochs (similar to 1000 epochs) to achieve steady-state performance. The accuracy (AC) of Res50_UNet is higher than 99.5% with only 1000 epochs, which is very impressive. This study demonstrated the feasibility of applying Res50_UNet in spine segmentation. The network integrates the characteristics of FPN and UNet. These results have shown the potential for Res50_UNet in spine MRI segmentation, especially when a low number of epochs is desirable.