Comparison of 2D, 2.5D, and 3D landmark localization networks for 3D cephalometry in CT images

Accurate landmark localization is important for three-dimensional (3D) cephalometric analysis. Although deep learning has shown promising performance for 3D landmark localization, the high computational burden of processing volumetric data remains a challenge. The 2.5D networks have emerged to provide the good performance while mitigating computational and memory requirements in the medical domain. Therefore, we compared the performance of 2D, 2.5D and 3D network-based landmark localization.

We collected landmark datasets from the volumetric computed tomography (CT) scans of 40 patients. We implemented the 2D, 2.5D and 3D networks for 3D landmark localization. Additionally, we designed a global-to-local loss to mitigate foreground-background imbalance, and employed both soft and hard voting in a network ensemble to improve the robustness. We evaluated each network's performance in terms of accuracy and computational load.

The 2.5D network-based landmark localization achieved a mean radial error (MRE) of 1.19[Formula: see text]0.65 mm and a successful detection rate (SDR) of 86.46% at 2mm, with a favorable computational load. These results outperformed those of the 2D and 3D networks. Furthermore, using the global-to-local loss led to higher performance compared to using the global loss alone. Soft voting proved the most robust performance among voting methods for landmark localization.

Comprehensive experiments demonstrate that the 2.5D network offers an optimal trade-off between computational load and accuracy. These findings highlight the potential for more efficient and reliable 3D cephalometry under limited computational resources.