ImplantFormer: Vision Transformer based Implant Position Regression Using Dental CBCT Data

TL;DR

This work tackles automatic dental implant position estimation from CBCT by predicting implant location from a 2D axial crown view and deriving the root position via a centerline projection. It introduces ImplantFormer, a Vision Transformer–based regression network with a convolutional stem, multi-level feature fusion in a convolutional decoder, and a Gaussian heatmap regression head trained with a focal loss and an offset term. A crown-to-root projection algorithm and post-processing map crown predictions back to root coordinates, enabling 3D consistency without heavy computation. Experiments on a clinical CBCT dataset show crown-based training yields superior accuracy over root-based approaches, and ImplantFormer outperforms state-of-the-art detectors, supported by attention visualization and slice-view validations.

Abstract

Implant prosthesis is the most appropriate treatment for dentition defect or dentition loss, which usually involves a surgical guide design process to decide the implant position. However, such design heavily relies on the subjective experiences of dentists. In this paper, a transformer-based Implant Position Regression Network, ImplantFormer, is proposed to automatically predict the implant position based on the oral CBCT data. We creatively propose to predict the implant position using the 2D axial view of the tooth crown area and fit a centerline of the implant to obtain the actual implant position at the tooth root. Convolutional stem and decoder are designed to coarsely extract image features before the operation of patch embedding and integrate multi-level feature maps for robust prediction, respectively. As both long-range relationship and local features are involved, our approach can better represent global information and achieves better location performance. Extensive experiments on a dental implant dataset through five-fold cross-validation demonstrated that the proposed ImplantFormer achieves superior performance than existing methods.

Figures (9)

• Figure 1: The whole procedure of the proposed implant position prediction method, which consists of training and inference mode.
• Figure 2: Visual comparison of the 2D slice of tooth crown (first row) and tooth root (second row). The white number represents the serial number of slice. Areas covered in yellow and blue represent the total area of neighboring teeth at tooth crown and tooth root, respectively; and arrows show the distance between the neighboring teeth.
• Figure 3: The network structure of the proposed ImplantFormer.
• Figure 4: Some sample images in our dataset. The red points denote the implant position annotation.
• Figure 5: Visual comparison of detection results of the proposed component. The red and yellow circles represent the predicted implant position and ground truth position, respectively.