Improving Vertebrae Segmentation Using a Centroid Detection-Guided Transformer-Based Network

Sevde Aydodgu, Ka-Wai Yung, Danail Stoyanov, Deepak Kalaskar, Evangelos B. Mazomenos

Abstract: Segmentation and identification of vertebrae are crucial tasks for diagnosing spinal deformities and treatment planning. However, past methods have often treated these tasks separately, neglecting their inherent relationship. This paper proposes a single-stage 2D centroid-detection guidance segmentation network (CD-VerTransUNet) that utilizes global information between vertebrae and the relationship between the two tasks. Moreover, a resampler module enhances the segmentation of rare (e.g. T13/L6) vertebrae. The proposed model demonstrates state-of-the-art segmentation performance for 2D models on the VerSe’20 dataset, achieving a dice-coefficient (DSC) of 75.15% for sagittal and 71.16% for coronal plane. Our novel multitasking approach even shows comparable performance to 3D architectures, yielding a DSC of 77.02% on the VerSe’20 and 71.75% on a scoliotic dataset.