Deep learning for gender estimation using hand radiographs: a comparative evaluation of CNN models

Area of Science:

• Forensic Science
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate gender estimation is vital in forensic identification, particularly with fragmented remains.
• Traditional methods are often unavailable in mass disasters or decomposed cases.
• Hand radiographs offer a potential source for objective gender classification.

Purpose of the Study:

• To develop and evaluate a deep learning model for gender classification using hand radiographs.
• To provide a rapid and objective alternative to conventional forensic identification techniques.

Main Methods:

• Analysis of 470 adult left-hand X-ray images (ages 18-65).
• Utilized four Convolutional Neural Network (CNN) architectures: ResNet-18, ResNet-50, InceptionV3, and EfficientNet-B0.
• Applied image preprocessing and data augmentation (rotation, flipping, brightness) for model generalization.

Main Results:

• ResNet-50 achieved the highest accuracy (93.2%), with precision (92.4%), recall (93.3%), and F1 score (92.5%).
• CNNs demonstrated the ability to extract sexually dimorphic features from hand radiographs.
• ResNet-50 consistently outperformed other tested models.

Conclusions:

• Deep learning, specifically ResNet-50, offers a robust solution for gender prediction from hand X-rays.
• This method is valuable for forensic scenarios requiring speed and reliability.
• Future research should focus on diverse population validation and explainable AI.