# Traditional and Artificial Intelligent Methods in Predicting Maxillofacial Soft Tissue Morphology After Orthognathic Surgery: A Narrative Review

## TL;DR

This review compares traditional and AI methods for predicting soft tissue changes after facial surgery, highlighting AI's potential and areas needing improvement.

## Contribution

The paper systematically evaluates AI methods for predicting post-surgery soft tissue morphology and suggests strategies to enhance their accuracy and usability.

## Key findings

- AI methods, especially deep learning, offer higher accuracy and speed compared to traditional 2D prediction methods.
- Deep learning models like DNN, PointNet, and Transformer show promise but face challenges like data insufficiency and lack of interpretability.
- Improvements through large databases, algorithm optimization, and multimodal data integration are recommended for better prediction.

## Abstract

The soft tissue morphology after orthognathic surgery has always been difficult to predict in preoperative planning. This narrative review aims to evaluate the current prediction methods and point out the direction for future research to improve predicting accuracy and speed. PubMed was the database for the review. The keywords for searching include: orthognathic surgery, prediction methods, simulation methods, artificial intelligence (AI), machine learning (ML), deep learning (DL), soft tissue, and surgical planning. Sixty-seven articles were found and 60 relevant articles were evaluated. The review finds out that with the transition from two-dimensional (2D) prediction to three-dimensional (3D) prediction, AI algorithms, especially DL methods, are brought forward promisingly, such as deep neural network (DNN), PointNet, and Transformer. DL methods have the advantages of high accuracy, fast speed, and simplified use, therefore have broad development prospects. However, DL methods need to be further improved, mainly in data insufficiency, overfitting, and the lack of interpretability. To conclude, the review proposes methods that could improve the accuracy and availability of prediction by DL, such as learning from large and high-quality databases, establishing separate databases for groups with different features, optimizing and simplifying the algorithm structure, utilizing integrated multimodal data, and implementing visualization techniques.