Thoratic Spine Segmentation Based on CT Images

Abstract

Automatic vertebrae localization and segmentation in computed tomography (CT) are fundamental for computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems. However, they remain challenging due to the high variation in spinal anatomy among patients. In this paper, we propose a simple, model-free approach for automatic CT vertebrae localization and segmentation. The segmentation pipeline consists of 3 stages. In the first stage the center line of the spinal cord is estimated using convolution. In the second stage a baseline segmentation of the spine is created using morphological reconstruction and other classical image processing algorithms. Finally, the baseline spine segmentation is refined by limiting its boundaries using simple heuristics based on expert knowledge. We evaluated our method on the COVID-19 subdataset of the CTSpine1K dataset. Our solution achieved a dice coefficient of 0.8160±0.0432 (mean±std) and an intersection over union of 0.6914±0.0618 for spine segmentation. The experimental results have demonstrated the feasibility of the proposed method in a real environment.