Artificial Intelligence for Rhinoplasty Design in Asian Patients

Li R, Shu F, Zhen Y, Song Z, An Y, Jiang Y.

PubMed 37580565 Publisher (DOI)Free full text (PMC)

What this paper says

Trained on 209 pairs of original and manually designed three dimensional face scans, an artificial intelligence model produced automated rhinoplasty designs the authors describe as similar to the manual ones.

Overview

Preoperative design in rhinoplasty has no uniform standard, and three dimensional simulation still depends on the individual surgeon's aesthetic judgment and experience. The authors set out to automate it. They collected three dimensional facial images from 209 patients, keeping both the original face and the face as manually designed in simulation software, converted the images into point clouds, and trained a modified FoldingNet neural network model.

Sections of note

Design: machine learning model development study. Level of evidence IV.
Dataset: three dimensional facial images from 209 patients, each with an original and a manually designed version.
Method: images converted to point clouds, then a modified FoldingNet model trained using PyTorch 1.12.
Result: the trained model performed aesthetic design automatically with results the authors describe as similar to manual design.
The authors analyzed 1027 facial features captured by the model.
They describe two apparent modes in the model: one resembling human aesthetic reasoning, one solving the task in a machine specific way.
The authors state this is the first artificial intelligence model for automated preoperative three dimensional rhinoplasty simulation.
No numerical similarity scores, surgeon ratings or patient outcomes appear in the abstract.

What it means for a patient

The output is a simulated preview of a possible result, not a surgical plan and not a promise of what surgery will achieve.
The model learned from designs made by particular surgeons on Asian faces, so it reproduces those preferences rather than any objective ideal.
Similarity to manual design is reported without a stated measure, so how close the outputs actually were cannot be judged.
No patients were operated on based on these designs, and no outcomes are reported.

Why this paper matters

Preoperative simulation shapes patient expectations, which is a major driver of satisfaction, yet producing it is slow and depends on the operator. Automating it would make simulation available in more practices. Whether automated designs are acceptable to patients and achievable in surgery is untested.

Terms

Three-dimensional simulation: a computer preview of how a face might look after surgery.
Point cloud: a set of coordinate points representing a three dimensional surface.
Neural network: a machine learning model built from layers of adjustable connections.
FoldingNet: a neural network architecture designed to reconstruct three dimensional point clouds.
Preoperative design: planning the intended result before surgery.

Summary written by rhinoplasty.cc from the abstract, 2026-09-09; not medical advice. The authors' own abstract follows.

Abstract

Background: Rhinoplasty is one of the most challenging plastic surgeries because it lacks a uniform standard for preoperative design or implementation. For a long time, rhinoplasties were done without an accurate consensus of aesthetic design between surgeons and patients before surgery and consequently brought unsatisfactory appearance for patients. In recent years, three-dimensional (3D) simulation has been used to visualize the preoperative design of rhinoplasty, and good results have been achieved. However, it still relied on individual aesthetics and experience. The preoperative design remained a huge challenge for inexperienced surgeons and could be time-consuming to perform manually. Therefore, we adopted artificial intelligence (AI) in this work to provide a new idea for automated and efficient preoperative nasal contour design.

Methods: We collected a dataset of 3D facial images from 209 patients. For each patient, both the original face and the manually designed face using 3D simulation software were included. The 3D images were transformed into point clouds, based on which we used the modified FoldingNet model for deep neural network training (by pytorch 1.12).

Results: The trained AI model gained the ability to perform aesthetic design automatically and achieved similar results to manual design. We analysed the 1027 facial features captured by the AI model and concluded two of its possible cognitive modes. One is to resemble the human aesthetic considerations while the other is to fulfil the given task in a special way of the machine.

Conclusion: We presented the first AI model for automated preoperative 3D simulation of rhinoplasty in this study. It provided a new idea for the automated, individual and efficient preoperative design, which was expected to bring a new paradigm for rhinoplasty and even the whole field of plastic surgery.

Level Of Evidence Iv: This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266 .

Abstract as indexed by PubMed; the article is open access (PubMed Central).