Deep Learning-based 3D Oral Cavity Reconstruction Using 2D Intraoral Images

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Medical Imaging & 3D Modeling · Depth: Advanced, long

Summary

A deep learning framework is proposed for 3D oral cavity reconstruction using only ten 2D intraoral images, eliminating the need for expensive hardware like intraoral scanners or Cone Beam CT. This software-based approach addresses patient discomfort, material deformation errors, and high equipment costs associated with conventional methods. The model, trained on 950 upper jaw samples from the Dental3DS dataset, employs MobileNetV2 for image encoding and Multi-head Attention for multi-view feature fusion. It directly predicts 50,000 3D vertex coordinates and achieved an accuracy of 77.49%, measured by nearest-neighbor matching with a distance threshold of 0.035. However, a limitation identified is the uneven distribution of predicted vertices, which tend to concentrate in high-density regions of the ground truth due to the Chamfer Distance loss function.

Key takeaway

For dental practitioners or AI scientists developing accessible 3D modeling solutions, this research demonstrates a viable software-only approach. You can achieve 3D oral cavity reconstruction from just ten 2D intraoral images, significantly lowering equipment costs and patient discomfort. However, you must address the current limitation of uneven point distribution caused by Chamfer Distance to ensure clinical utility. Future work should focus on refining loss functions for more uniform vertex prediction.

Key insights

3D oral models can be reconstructed from ten 2D images using deep learning, reducing cost and discomfort but facing point distribution challenges.

Principles

• Software-based 3D reconstruction reduces hardware dependency.
• Multi-head Attention effectively fuses multi-view features.
• Chamfer Distance can cause uneven point distribution.

Method

An encoder-decoder model uses MobileNetV2 for feature extraction from ten 2D images, fuses them with Multi-head Attention, and a decoder predicts 50,000 3D vertex coordinates.

In practice

• Use 2D intraoral images for cost-effective 3D modeling.
• Explore MobileNetV2 and Multi-head Attention for multi-view tasks.
• Consider loss function impact on point cloud uniformity.

Topics

• 3D Reconstruction
• Deep Learning
• Oral Cavity Modeling
• Intraoral Imaging
• MobileNetV2
• Multi-head Attention
• Chamfer Distance

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Deep Learning-based 3D Oral Cavity Reconstruction Using 2D Intraoral Images — Deep learning can reconstruct 3D oral models from minimal 2D images, overcoming cost and discomfort of traditional methods.
• VGG-T$^3$: Offline Feed-Forward 3D Reconstruction at Scale — VGG-T^3 enables scalable 3D reconstruction by converting variable scene geometry into a fixed-size MLP via test-time training.
• When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT — Higher dimensionality in medical imaging models does not guarantee stable performance, especially with class imbalance.
• IEEE Transactions on Neural Networks and Learning Systems, Volume 37, Issue 3, March 2026 — The issue highlights advancements across neural networks, computer vision, and machine learning applications.
• Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning — Injecting fundamental geometric priors into VLMs directly enhances 3D spatial reasoning more effectively than VQA fine-tuning.
• Real2SAM2Real: Generative 3D Caches as Complementary Context for Video Diffusion — Real2SAM2Real uses editable 3D caches from lifting models to provide robust geometric scaffolds for Video Diffusion Models, improving control and consistency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.