🧬 biology

Semi-automated sex estimation from cochlear shape using clinical CT imaging and OTOPLAN segmentation

This study demonstrates that a semi-automated pipeline using clinical CT scans and OTOPLAN segmentation can achieve moderate accuracy (79%) in sex estimation by analyzing cochlear shape descriptors, offering a non-destructive and scalable alternative for forensic anthropology when traditional skeletal methods are unavailable.

Original authors: Priya VISHNUMURTHY, Mary DAVAL, Luke HELPARD, Ismael OUZZINE, Ralph HADDAD, Nikola IVANOVIC, Vincent JOURDES, Justin MICHEL, Stéphane GARGULA

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Priya VISHNUMURTHY, Mary DAVAL, Luke HELPARD, Ismael OUZZINE, Ralph HADDAD, Nikola IVANOVIC, Vincent JOURDES, Justin MICHEL, Stéphane GARGULA

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the only clue you have left is a tiny, hard, spiral-shaped shell hidden deep inside a skull. This shell is the cochlea, the part of your inner ear that helps you hear. Usually, when forensic experts (the "detectives" of the dead) need to figure out if a skeleton belonged to a man or a woman, they look at big bones like the pelvis or the skull. But what if those big bones are missing, burned, or broken? That's where this study comes in.

Here is a simple breakdown of what the researchers did, using some everyday analogies:

The Big Idea: The "Ear Shell" Has a Secret

The researchers wanted to see if they could guess a person's sex just by looking at the shape of their cochlea. Think of the cochlea like a snail shell. Even though all snail shells look roughly the same, if you measure them very carefully, you might find that male shells are slightly bigger or twist slightly differently than female shells.

The challenge? Doing this manually is like trying to measure a snail shell with a ruler while it's spinning in a jar. It's slow, hard, and requires a microscope (or a very high-tech scanner).

The New Tool: The "Robot Measurer"

Instead of doing it by hand, the team used a piece of software called OTOPLAN.

What it is: Imagine a robot that automatically traces the outline of the snail shell on a computer screen. It was originally built to help doctors plan ear implants, but the researchers realized it could also be used to measure the shell's shape for their mystery.
The Process: They took CT scans (like 3D X-rays) of 87 adults' ears. The software automatically traced the "path" of the cochlea and measured its size and how much it twisted.

The Experiment: Teaching a Computer to Guess

The researchers fed this data into a computer program (a "machine learning" model). Think of this like teaching a dog to distinguish between two types of balls.

The Training: They showed the computer thousands of examples of male and female cochleae.
The Test: They then gave the computer a new set of ears it had never seen before and asked, "Is this a man or a woman?"

What They Found

The computer got it right about 79% of the time.

The "Magic" Clue: The most important thing the computer looked at was the diameter (how wide the shell is). Men's shells were generally wider.
The "Twist" Clue: The second most important clue was the ratio of twisting to bending. It's not just about size; it's about how the shell curves.
The Catch: If they only looked at the size (the diameter) and ignored the shape, the computer got confused and failed when tested on new people. It was like a student who memorized the answers to a practice test but couldn't solve a new problem. They needed both the size and the shape to get it right.

Why This Matters (According to the Paper)

The paper emphasizes that this method is not a magic wand that will solve every case perfectly.

It's a "Backup Plan": It's designed for those tough cases where the usual clues (pelvis, skull) are gone. If you only have a piece of the ear bone left, this method gives you a better-than-random guess.
It's Non-Destructive: You don't have to break the bone to look at it. You just scan it.
It's Fast and Repeatable: Because a robot does the measuring, anyone can use it, and they will get the same result every time.

The Bottom Line

The researchers proved that you can use a standard hospital CT scan and a robot software to guess a person's sex based on their inner ear shape. It's not perfect (it's about 8 out of 10 correct), but in a situation where you have almost no other clues, being 80% right is a very helpful tool for the detective.

Important Note: The paper explicitly states this is a "proof of concept." It shows the method works in a controlled setting, but it is meant to be used as an additional tool alongside other methods, not as the only way to identify someone.

Technical Summary: Semi-automated Sex Estimation from Cochlear Shape

Problem Statement
Sex estimation is a cornerstone of forensic anthropology and bioarchaeology, yet classical methods relying on the pelvis and skull are often inapplicable when remains are fragmented, commingled, or degraded. The petrous part of the temporal bone, and specifically the bony labyrinth within it, offers a resilient alternative due to its high density and early maturation. While previous research has demonstrated that cochlear shape carries a sex-specific signal, existing high-accuracy methods (e.g., those by Braga et al., 2019) rely on high-resolution imaging, manual or semi-manual segmentation, and complex geometric morphometrics. These requirements limit their routine applicability in forensic and clinical workflows. The challenge addressed by this study is to determine whether sex estimation can be achieved using a semi-automated pipeline compatible with standard clinical CT imaging and automated segmentation tools, without sacrificing robustness.

Methodology
The study employed a retrospective analysis of 139 cochleae from 87 adult individuals with known biological sex, acquired via routine clinical CT scans (slice thickness ≤ 0.6 mm). The analytical pipeline consisted of the following stages:

Data Extraction: Cochlear segmentation was performed using OTOPLAN, a commercial software platform designed for cochlear implant planning. For each ear, the system exported:
- A 3D lateral wall path represented by 400 landmarks.
- Six scalar metrics: cochlear diameter, width, height, estimated cochlear duct length (CDL), CDL from 0, and number of turns.
Preprocessing:
- Left cochlear curves were mirrored to ensure uniform rotational direction.
- Curves were translated to their centroid and scaled by total norm to isolate shape from size and position.
- Outliers were removed via Principal Component Analysis (PCA).
- Data was split at the patient level (80% training, 20% holdout) to prevent data leakage, ensuring both ears of a single patient remained in the same partition.
Feature Engineering:
- Geometric Descriptors: The normalized curve was divided into 19 segments. Curvature ( $\kappa$ ) and torsion ( $\tau$ ) were computed, along with their ratio ( $\tau/\kappa$ ). Descriptors included mean, standard deviation, and arc-length integrals for each segment.
- Scalar Metrics: The six OTOPLAN-derived scalar variables were retained.
Model Optimization:
- A Bayesian optimization framework (Optuna) was used to jointly optimize preprocessing choices, feature construction, and classification hyperparameters.
- The objective was to maximize the Area Under the ROC Curve (AUC-ROC) via 5-fold cross-validation on the training set.
- Three model configurations were tested: scalar metrics only, geometric descriptors only, and a combined approach.
Classification: The final model was an XGBoost classifier trained on the combined feature set and evaluated once on the independent holdout set ( $n=29$ ).

Key Results

Univariate Analysis: Significant sex differences were found in four scalar metrics (notably cochlear diameter, $r=0.43$ ) and ten curvature/torsion descriptors. However, effect sizes were generally small-to-moderate, and distributions showed substantial overlap, indicating no single metric was sufficient for individual classification.
Model Performance:
- The scalar-only model exhibited marked overfitting (Cross-Validated AUC 0.82 $\rightarrow$ Holdout AUC 0.59).
- The geometry-only model generalized better (Holdout AUC 0.69).
- The combined model (XGBoost) achieved the best performance on the independent holdout set:
  - AUC-ROC: 0.77 (95% CI: 0.55–0.95).
  - Overall Accuracy: 79.3%.
  - Balanced Accuracy: 79.1%.
Feature Importance: Permutation importance analysis revealed that cochlear diameter was the dominant feature. The $\tau/\kappa$ ratio across several spiral segments provided the secondary, complementary geometric signal. The model relied on size information but required geometric context to avoid overfitting.

Significance and Claims
The authors position this work as a proof of concept demonstrating that sex-related cochlear shape information is recoverable from routine clinical CT data using a semi-automated, reproducible pipeline. The study claims the following significance:

Clinical and Forensic Compatibility: Unlike previous high-accuracy methods requiring micro-CT or manual segmentation, this approach utilizes standard clinical CT and automated software (OTOPLAN), making it scalable and non-destructive.
Adjunct Utility: The authors explicitly state that the moderate accuracy (approx. 79%) does not support using this method as a standalone tool for individual identification. Instead, it is proposed as a valuable adjunct for biological profiling in cases where the petrous bone is the only surviving element (e.g., fire victims, mass fatality incidents, or degraded remains).
Non-Destructive Workflow: The method preserves the specimen for subsequent genetic testing, which remains the gold standard for identification in highly degraded cases.
Future Potential: The authors suggest that performance could improve with higher-resolution imaging (e.g., Cone-Beam CT or Photon-Counting CT) and that the current results represent a lower bound under contemporary clinical imaging conditions.

The study concludes that while the method cannot replace established forensic techniques, it offers a viable, non-invasive probabilistic indicator for sex estimation in challenging forensic contexts where conventional skeletal markers are absent.