A pseudo-3D coarse-to-fine architecture for 3D medical landmark detection

The coarse-to-fine architecture is a benchmark method designed to enhance the accuracy of 3D medical landmark detection. However, incorporating 3D convolutional neural networks into the coarse-to-fine architecture leads to a significant increase in model parameters, making it costly for deployment in clinical applications. This paper introduces a novel lightweight pseudo-3D coarse-to-fine architecture, consisting of a Plane-wise Attention Pseudo-3D (PA-P3D) model and a Spatial Separation Pseudo-3D (SS-P3D) model. The PA-P3D inherits the lightweight structure of the general pseudo-3D and enhances cross-plane feature interaction in 3D medical images. On the other hand, the SS-P3D replaces the 3D model with three spatially separated 2D models to simultaneously detect 2D landmarks on axial, sagittal, and coronal planes. In comparison to the conventional coarse-to-fine architecture, the proposed method requires only approximately a quarter of the model parameters (60% reduced by PA-P3D and 40% reduced by SS-P3D) while simultaneously improving landmark detection performance. Experimental results demonstrate the effectiveness of the proposed method, achieving state-of-the-art performance on both a public dataset for mandibular molar landmark detection and a private dataset for cephalometric landmark detection. Overall, this paper highlights the potential of the coarse-to-fine method for cost-effective model deployment, thanks to its lightweight model structure.