Segmentation and Labeling of Vertebra Using SegFormer Architecture

Vertebra segmentation and labeling in MR images of the spine play a vital role in the identification of diseases or anomalies. MRI captures the tissue structure of a spine accurately, hence it is essential to demarcate and identify the vertebra in the MRI image. There are both supervised and unsupervised methods for vertebra segmentation and labeling. However, the acquisition of requisite data is a challenge to designing methods with very high accuracy. In this work, we have modified a transformer-based architecture called Segformer for semantic segmentation of 3D sliced data. Our method leverages transfer learning on low-population data. With a new advanced masking logic, we achieve 99% accuracy for segmentation and labeling of lumbar spine MR images.