Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1873
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-19T01:09:16.965623Z
status: published
lens: p_space
sector: health
headline: A comparative analysis of artificial intelligence generated 3d facial models from 2D photos: A procrustes-based validation of landmark accuracy in Orthodontics
dek: Objectives: The primary aim of this study was to evaluate the accuracy and precision of a deep learning artificial intelligence (AI) model in reconstructing 3D facial geometry from a single 2D photograph. The study specifically measured the coordinate differences of human-placed facial landmarks on the AI-generated models against those placed on 3D models from a gold-standard scanner (Bellus 3D). Secondary objectives included identifying dimensional variations (X, Y, Z) and systematic biases in the AI reconstruc…
gain_title: (none)
problem_title: The AI reconstructions showed systematic downward displacement of chin landmarks such as Menton and Pogonion and larger errors at the lateral and inferior face borders, which could affect diagnosis and treatment planning.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: The AI reconstructions showed systematic downward displacement of chin landmarks such as Menton and Pogonion and larger errors at the lateral and inferior face borders, which could affect diagnosis and treatment planning.
problem_evidence: systematic biases were observed mainly as inferior displacements of key landmarks | Menton (-0.35 units) and Pogonion (-0.30 units)
quick_read: Researchers at the University of the Pacific tested a deep learning model, 3DDFA-V2, that builds a 3D facial model from one 2D photograph. They compared its output with same-day Bellus 3D scanner models for 47 anonymized patient records, using landmarks placed by two trained observers and Procrustes normalization.

The result suggests a low-cost route to 3D facial records in orthodontics, with more repeatable landmark placement than the scanner models. The consistent downward shift of chin landmarks and larger errors at some facial borders mean the algorithm needs refinement before clinicians rely on it for diagnosis and treatment planning.
limitation: The AI models showed systematic biases, particularly a consistent downward shift of landmarks like Menton and Pogonion, and the authors say the algorithm needs further refinement before clinical use.
tag: Evidence-backed problem
key_points: The study was retrospective and used 47 anonymized patient records from the University of the Pacific database. | Mean Procrustes-normalized error across all landmarks was 0.19 ± 0.08 units, about 0.2 mm of proportional deviation. | The largest deviations were at Right Gonion, Trichion and Left Gonion. | Menton and Pogonion were shifted inferiorly, by -0.35 and -0.30 units respectively. | Authors call the tool promising and low cost but say further algorithm refinement is needed.
rundown: Each patient had a 3D model generated from a 2D photograph with 3DDFA-V2, and a reference model captured the same day with a Bellus 3D scanner. Two trained observers marked facial landmarks on both model types.

Coordinates were normalized with Procrustes analysis so shape and landmark position could be compared independent of scale and orientation. The team then statistically analyzed landmark coordinate differences and examined variation along the X, Y and Z dimensions.

Landmark identification on the AI models was less variable than on the gold-standard scans, which the authors say may yield more repeatable measurements. Even so, the authors report systematic bias and tie their findings to clinical use of AI-generated models in diagnosis and treatment planning.
sources:
- peer_reviewed | Seminars in Orthodontics | https://doi.org/10.1053/j.sodo.2025.10.014 | 2025-10-29
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

e973f45503f876d970a18a34cbcccc03a63f5c4ee3e15babd66820dee069ee41

previous

0000000000000000000000000000000000000000000000000000000000000000