SVVD-NET: A framework with relative position constraints for vertebral vertex detection

Abstract

Vertebral vertex detection is a fundamental step for subsequent spine image analysis and X-ray image-based spine disease diagnosis. Existing CNN-based frameworks for landmark detection can be directly used for automatic vertebral vertex detection. However, challenges such as overlapping vertebrae and vertebrae misalignment often arise when applying existing methods to vertebral vertex detection. To address the issues, we propose a sequential vertebral vertex detection network (SVVD-Net) that fully utilizes the regularity of vertebral alignment to generate vertebral vertices with relative positional constraints. Leveraging information about previously predicted vertebrae to identify the next one, the SVVD-Net could make sequential predictions and effectively avoid vertebrae overlapping and misalignment. We design an anatomy-aware encoder based on external attention mechanism, to address the anatomical information regarding the similarities in shape and alignment of vertebrae among samples. Structured mask is used in the decoder to reduce the direct influence of one vertebra upon its immediate neighbor and accordingly accommodate occasional subtle misalignment between two adjacent vertebrae. We evaluate the performance of SVVD-Net on two datasets of X-ray images of the spine. The results indicate that the proposed SVVD-Net consistently outperforms state-of-the-art methods. Ablation experiments further support the effectiveness of involved sequential landmark generation, anatomy-aware encoder and structured mask. Accordingly, this study presents a successful attempt to incorporate anatomical priors into medical image analysis.