📄 medicine

Evaluating the Practical Compatibility of Two Different Facial Soft Tissue Depth Datasets: A Recognition-Based Evaluation Using the Combination Method

This study demonstrates that facial approximations generated using different facial soft tissue depth (FSTD) measurement methods (2D and 3D) are practically compatible, as observer recognition tests confirmed that the resulting approximations remained recognizable above chance levels.

Original authors: Gülçin Coşkun-Boileau

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gülçin Coşkun-Boileau

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a cold case. You have found a skull, but you have no idea who it belonged to. To help identify the person, you need to build a face on that skull, a process called facial approximation. It's like sculpting a portrait from a skeleton.

To do this accurately, sculptors need a "ruler" to tell them how thick the skin and muscle should be at every point on the face. This ruler is called a Facial Soft Tissue Depth (FSTD) dataset.

The Big Question

For years, scientists have been arguing about the best way to measure these "rulers." Some use 2D methods (like looking at flat X-ray slices on a screen), while others use 3D methods (like measuring directly on a digital 3D model of the skull).

The big question this paper asks is: Does it actually matter which "ruler" you use? If you build a face using the 2D measurements versus the 3D measurements, will people still recognize the person, or will the differences make the face look like a stranger?

The Experiment: A "Face-Off"

The researcher, Gülçin Coşkun-Boileau, decided to test this practically, not just with math, but with real human eyes. Here is how she set up the experiment:

The Subjects: She picked three real people (two men, one woman) and used their archived CT scans.
The Double-Scultping: For each person, she made two clay faces.
- Face A: Built using the 2D measurement data.
- Face B: Built using the 3D measurement data.
- Note: She used the same skull for both, so the only difference was the "ruler" she used to decide how thick the clay should be.
The "Blind" Sculptor: The person making the faces didn't know who the people were. This ensures the sculptor didn't accidentally "cheat" by trying to make the face look like a specific celebrity they knew.
The Audience: She showed photos of these clay faces to 55 regular people (the "observers"). These people didn't know the subjects either.
The Test: The observers were shown a "Reference Face" (one of the clay models) and asked to pick which of the other photos looked most like it. They also rated how similar they looked on a scale of 1 to 5.

The Results: "It's All in the Eyes"

The results were surprisingly positive.

High Recognition: The observers were able to match the "Reference Face" to its twin (the one made with the other measurement method) at a rate far higher than random guessing.
- For the first person, 98% of people got it right.
- For the second, 83% got it right.
- For the third, 96% got it right.
The Verdict: The study concludes that using a 2D ruler or a 3D ruler doesn't ruin the final picture. Even though the measurements were slightly different, the resulting faces were still recognizable as the same person.

Why Was This Hard? (The "Missing Pieces")

The paper admits the experiment wasn't perfect, and here's why that matters:

The "No-Mouth" Rule: The clay faces didn't have mouths or ears because the skull models were missing those parts. It's like trying to recognize a friend when they are wearing a mask that covers their mouth.
The "No-Age" Rule: The sculptor didn't add wrinkles or gray hair because they didn't know the exact ages of the people. This made an older person look younger, which might have confused the observers.
The "Newbie" Sculptor: The person building the faces was new to the job. The paper notes that an expert artist might have done a better job, suggesting that skill matters just as much as the data.

The Takeaway

Think of the 2D and 3D measurement methods as two different brands of paint. This study found that even if you use Brand A or Brand B, the final painting still looks like the same person.

The main conclusion is simple: You don't need to panic if you have to switch between 2D and 3D measurement data. As long as you follow the rules, the face you build will likely still be recognized by the public. However, the paper warns that the skill of the person building the face is just as important as the data they use.

Note: The paper specifically states these results apply to manual (hand-sculpted) methods. It does not claim these results apply to computer-generated faces, suggesting that future studies are needed to see if computers behave the same way.

Technical Summary: Evaluating the Practical Compatibility of Two Different Facial Soft Tissue Depth Datasets

Problem Statement
Forensic facial approximation is a critical tool for identifying human skeletal remains when positive identification methods fail. The accuracy of this process relies heavily on the application of Facial Soft Tissue Depth (FSTD) datasets. While various methods exist to establish these datasets—including needle puncture, two-dimensional (2D) imaging (radiographs, MRI, CT), and three-dimensional (3D) digital modeling—there is limited practical research comparing the compatibility of these different measurement approaches. Previous studies have noted statistical differences between datasets derived from 2D and 3D methods, yet the practical impact of these variations on the final facial reconstruction and subsequent recognition by observers remains unclear. This study addresses the gap between statistical variance in FSTD data and the practical outcome of facial recognition.

Methodology
The study employed a recognition-based evaluation using the "Combination method" of 3D manual facial approximation.

Subjects: The material consisted of head CT scans from three individuals (two males aged 55 and 76, one female aged 28).
Data Generation: Two distinct FSTD datasets were derived from the same CT scans for each subject: one using a 2D measurement method and the other using a 3D measurement method. Both datasets were based on adult Greek population standards, adjusted for age and sex.
Reconstruction Process:
- 3D Printing: CT scans were segmented to create 3D skull models. Two copies of each skull were 3D printed (using Ultimaker 3 and Creality Ender-3 printers) to facilitate parallel reconstruction.
- Approximation: The author, who had no prior experience in facial approximation, performed the reconstructions under blind conditions (unfamiliar with the subjects' identities). Wooden markers corresponding to the specific FSTD values from either the 2D or 3D datasets were glued to the skulls at 15 cranial landmarks.
- Features: Facial muscles and soft tissue were modeled using air-dry clay. Eyeballs were placed according to standard protocols, and the nasal tip and width were developed using established standards. Due to missing lower skull portions, mouths were not reconstructed, and ears were omitted as the evaluation focused on the frontal view. No age-related features (wrinkles) or hair were added to prevent identity cues.
Evaluation: Six facial approximations (three subjects × two datasets) were photographed. These were presented to 55 observers (22 male, 33 female, ages 18–67) in a "face pool" and "resemblance rating" test. Observers were asked to identify a target individual from a set of images and rate the resemblance on a scale of 1 (no resemblance) to 5 (great resemblance).
Statistical Analysis: Data were analyzed using SPSS. Chi-square tests determined if identifications occurred above chance, and frequency analysis was used for resemblance ratings.

Key Results

Recognition Rates: The study achieved high recognition rates across all three subjects, significantly above chance levels ( $p < 0.01$ $p < 0.01$ ).
- Approximation 1 (76-year-old male): 98.18% correct identification.
- Approximation 2 (55-year-old male): 83.63% correct identification.
- Approximation 3 (28-year-old female): 96.36% correct identification.
Dataset Compatibility: The use of two different FSTD measurement methods (2D vs. 3D) did not compromise facial recognition. Observers successfully recognized the target individuals regardless of which dataset was used for the reconstruction.
Resemblance Ratings: The majority of observers rated the resemblance as "close" (Score 4) or "great" (Score 5). No observers selected "no resemblance" (Score 1) for any of the approximations.
Demographic Factors: Female observers generally demonstrated higher recognition rates than male observers, consistent with existing psychological literature on facial recognition biases.

Key Contributions

Practical Validation: The study provides empirical evidence that statistical differences between 2D and 3D FSTD datasets do not necessarily translate to practical failures in forensic facial approximation.
Methodological Rigor: By using the same CT scans, the same practitioner, and the same skulls for both datasets, the study isolated the variable of the FSTD measurement method, controlling for image quality and practitioner subjectivity in the data generation phase.
Blind Reconstruction: The use of a practitioner with no prior experience in facial approximation highlights the robustness of the method, even when executed by a novice, provided standard protocols are followed.

Significance and Claims
The paper concludes that for the 3D manual Combination method, the choice between 2D and 3D FSTD datasets does not jeopardize the recognition of target individuals. The results suggest that the practical impact of the differences between these datasets is minimal regarding facial recognition outcomes.

However, the authors maintain a modest stance regarding the broader implications:

Limitations: The study acknowledges that the lack of mouths and ears in the reconstructions, along with the practitioner's inexperience, may have influenced the specific resemblance ratings. The absence of age indicators (wrinkles) may have affected the perception of older subjects.
Scope: The findings are specific to the 3D manual Combination method. The authors explicitly state that similar results cannot be assumed for 3D computerized methods without further investigation.
Practitioner Dependency: The study reinforces that while methods and standards are clearly described, the accuracy and realism of the final approximation remain dependent on the practitioner's experience and artistic skill.

In summary, the paper demonstrates that different FSTD measurement methodologies are practically compatible for manual facial approximation, supporting the continued use of diverse datasets in forensic contexts without fear of compromising identification potential.

The Big Question

The Experiment: A "Face-Off"

The Results: "It's All in the Eyes"

Why Was This Hard? (The "Missing Pieces")

The Takeaway

More like this