.png){kind=icon}
.png){kind=icon}
Return
MPGLNet: Multi-scale progressive global–local fusion dental point cloud complementation network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2025.108394.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The utilization of 3D scanning devices to capture oral ailments is experiencing a marked increase in popularity, and the quality of the resulting 3D dental models is of particular significance when repairing oral issues. However, the acquired point cloud data is frequently sparse and incomplete, resulting in missing information. To address this issue, we construct a high-resolution dental point cloud completion network named MPGLNet to output dense and complete teeth point clouds. MPGLNet operates in a coarse-to-fine manner. First, we incorporate a Local-Global Guidance Perception module (LGGP) into the encoder to collaboratively learn local and global features, while introducing guided learning to enhance the correlation between points. Next, a coarse and complete teeth point cloud skeleton is generated using a seed generator. Finally, we use a three-stage Progressive Multi-scale Attentional Feature Fusion Upsampling Block (PMA UpBlock) in the decoder to progressively perform cross-scale fusion and reconstruct the complex geometric structures of the teeth. During this process, the upsampling blocks gradually increase the points to predict a fine-grained, complete, high-resolution teeth point cloud. We conduct a series of experiments on the teeth point cloud completion dataset and compare it with representative work in the current field, as well as state-of-the-art algorithms that have performed well in point cloud completion tasks in recent years. Our MPGLNet achieves the best performance compared to other models across four evaluation metrics: CDL1, CDL2, EMD, and DCD. This is evident in terms of completion accuracy, geometric detail restoration and the uniformity of the generated point cloud. This innovative algorithm’s uniqueness lies in its ability to comprehensively perceive changes in point cloud features across different levels, effectively extracting detailed dental point cloud features from all angles. This work not only validates the effectiveness of the multi-scale feature fusion strategy in the task of teeth point cloud completion, but also demonstrates its application potential in practical scenarios such as digital oral medicine, oral restoration, and intelligent diagnosis and treatment modeling, providing new insights for promoting AI-assisted dental reconstruction. Our code is available at https://github.com/labiip/MPGLNet .
Keywords:
3D dental modeling
point cloud completion
multi-scale feature fusion
dental restoration
AI-assisted diagnosis
Journal
IF:
4.9
Papers:
9.8K
Citations:
2.4W
Organization
View more organizations