Abstract
PURPOSE: To use and test a labeling algorithm that operates on two-dimensional reformations, rather than three-dimensional data to locate and identify vertebrae. MATERIALS AND METHODS: The authors improved the Btrfly Net, a fully convolutional network architecture described by Sekuboyina et al, which works on sagittal and coronal maximum intensity projections (MIPs) and augmented it with two additional components: spine localization and adversarial a priori learning. Furthermore, two variants of adversarial training schemes that incorporated the anatomic a priori knowledge into the Btrfly Net were explored. The superiority of the proposed approach for labeling vertebrae on three datasets was investigated: a public benchmarking dataset of 302 CT scans and two in-house datasets with a total of 238 CT scans. The Wilcoxon signed rank test was employed to compute the statistical significance of the improvement in performance observed with various architectural components in the authors' approach. RESULTS: On the public dataset, the authors' approach using the described Btrfly Net with energy-based prior encoding (Btrfly(pe-eb)) network performed as well as current state-of-the-art methods, achieving a statistically significant (P < .001) vertebrae identification rate of 88.5% ± 0.2 (standard deviation) and localization distances of less than 7 mm. On the in-house datasets that had a higher interscan data variability, an identification rate of 85.1% ± 1.2 was obtained. CONCLUSION: An identification performance comparable to existing three-dimensional approaches was achieved when labeling vertebrae on two-dimensional MIPs. The performance was further improved using the proposed adversarial training regimen that effectively enforced local spine a priori knowledge during training. Spine localization increased the generalizability of our approach by homogenizing the content in the MIPs.Supplemental material is available for this article.© RSNA, 2020.