SpineCLUE: Automatic vertebrae identification using contrastive learning and uncertainty estimation.

Abstract

Vertebrae identification in arbitrary fields-of-view plays a crucial role in diagnosing spine disease. Most spine CT contain only local regions, such as the neck, chest, and abdomen. Existing spine-level methods, which rely on a priori on the specific number of target vertebrae, are less able to cope with this challenge. In this paper, we propose a three-stage vertebra-level method to address the challenges in 3D CT vertebrae identification with arbitrary fields-of-view. In order to integrate contextual prior information during identification, rather than identifying independently at the vertebrae-level, we perform the vertebrae localization, segmentation and identification tasks sequentially, thus making effective use of anatomical prior information about the vertebrae throughout the process. Specifically, to improve the stability of localization and prevent failures caused by abnormal vertebral positions in 3D space, we introduce a dual-factor density clustering algorithm to acquire localization information for individual vertebrae, thereby facilitating the subsequent segmentation and recognition processes. In addition, to tackle the issue of inter-class similarity and intra-class variability, we pretrain our identification network by using a supervised contrastive learning method. To further optimize the identification results, we estimated the uncertainty of the classification network and utilized the message fusion module to combine the uncertainty scores, while aggregating global information about the spine. Our method achieves state-of-the-art results on the VerSe20 challenge benchmark.