Validating AI for Gender Prediction Using Morphometric Analysis of the Mandible

Study Overview and Objective:

1. To evaluate the accuracy of machine learning algorithms (ANN and logistic regression) in gender identification from linear morphometric measurements of the mandible. These measurements will be derived from automated mandibular segmentations from CBCT scans of a sample of Egyptian dental patients. The AI model's accuracy, precision, and sensitivity (recall) will be assessed. The model's predictions will be compared to the known gender from patient records.
2. To evaluate the accuracy of AI (3D UNet) in anatomical point identification and linear measurement. The model's performance will be compared to the radiologist's manual point placement and measurements.

Rationale:

Gender identification is vital in forensic odontology as it makes individual identification easier and faster. Artificial Intelligence can improve accuracy, speed, and efficiency in gender identification processes. The mandible is a sturdy bone with dimorphic features that can aid in gender determination even if other bones are lost. Cone beam computed tomography (CBCT) provides accurate 3D images and therefore accurate measurements. Linear morphometric measurements of the mandible derived from CBCT will provide an objective way of gender determination. AI algorithms can use these measurements to predict gender. Evaluating the accuracy of artificial intelligence in gender determination using morphometric measurements from CBCT-based mandibular segmentations is therefore essential to assess its reliability and potential value as an objective diagnostic tool in forensic practice.

Data Source and Population: Pre-existing, anonymized CBCT scans will be retrieved from the archives of the Oral and Maxillofacial Radiology clinic at the Faculty of Dentistry, Cairo University.

Sample Size Calculation The sample size was calculated to be 385 CBCT DICOMs using the Stats Kingdom statistical tool, based on the primary outcome - accuracy of AI in gender prediction. The power of study was 0.8, the margin of error was 0.04 and the alpha level of significance was 0.05.

Methodology:

Reference standard:

After recruiting a scan in the study, the name of the patient will be removed, and the scan will be given an identification number instead. One of the supervisors will have a dataset with the identification number, age, and the correct gender written before all personal information is erased from the CBCT scan. This will ensure that the radiologist performing the morphometric measurements will be blinded.

The DICOM file will be opened in 3D slicer imaging software (version 5.10.0; MIT, Cambridge, MA, USA) where automatic segmentation of the mandible will be peformed.

The radiologist performing the measurements has 2 years of experience. The following linear measurements will be performed: bi-coronoid width, bi-condylar width, distance from right gonion to menton and left gonion to menton, and distance from right coronoid to right gonion and left coronoid to left gonion.These measurements will be written in a separate dataset with the identification number of the scan with no information about the gender. Interobserver agreement will be validated by a senior radiologist (>15 years experience) who will re-evaluate a 30% sample of the cases. The datasets will be sent to a statistician to assess the normality of data and sexual dimorphism.

The segmented mandibles (NRRD files) and the measurement files (JSON files) will be saved for each case to be introduced in the AI model by the AI engineer.

Index test:

A three-dimensional deep learning model will be developed to perform two tasks:

1. To identify the anatomical points and make the specified linear measurements from these points on the segmented mandibles
2. To accurately predict gender based on these measurements The AI engineer will be responsible for model implementation. For the first task, 3D U-Net model will be employed. The NRRD and JSON files will be introduced to the model for training and testing. For the second task, Artificial neural networks and logistic regression will be used.

Data preprocessing will be performed and the dataset will be divided into training, validation, and testing subsets to ensure balanced and unbiased evaluation. K-fold cross-validation will be used if sample size is small.

Training:

The landmark detection model will be trained using supervised learning techniques .The classification models will subsequently be trained using the extracted feature representations. Each model will be trained using the same training data and preprocessing pipeline.

Testing:

In the testing phase, the trained system will process unseen segmented mandibles to automatically localize keypoints, compute relevant geometric features, and generate gender predictions without further model updates.

Evaluation and Expected Outcome:

Model performance will be evaluated using quantitative localization error metrics for keypoint detection and standard classification metrics for gender prediction.

Statistical Analysis:

Statistical tests will be applied to assess the normality of data and sexual dimorphism of the linear morphometric mandibular measurements before introducing these measurements to the AI model. Mean Euclidean Distance Error (in millimeters), and Root Mean Square Error (RMSE) will be used to assess the accuracy of the model in anatomical point placement in comparison to the radiologist's measurements. Accuracy, Precision, Recall (sensitivity), Specificity, F1-score, and a Confusion matrix will be used to evaluate the prediction models.

No data will be missed considering the retrospective nature of the study