Human identification performed with skull's sphenoid sinus based on deep learning

Area of Science:

• Forensic Science
• Medical Imaging
• Artificial Intelligence

Background:

• Human identification is crucial in forensic investigations.
• Traditional methods can be time-consuming.
• 3D sphenoid sinus models offer unique anatomical features.

Purpose of the Study:

• To develop and evaluate a deep learning framework for human identification using 3D sphenoid sinus models.
• To assess the accuracy and speed of the proposed identification method.

Main Methods:

• A retrospective study utilizing 1475 noncontrast thin-slice CT scans from 732 patients.
• Development of a customized convolutional neural network (MVSS-Net) for sphenoid sinus modeling and recognition.
• Training and testing the deep learning model on segmented sphenoid sinus data.

Main Results:

• The deep learning framework achieved Rank 1 accuracy of 93.94% and Rank 5 accuracy of 99.24% on the test set.
• Identification process completed within 55 seconds, demonstrating high efficiency.
• Manual verification of framework results yielded 100% accuracy in the test set.

Conclusions:

• Customized deep learning on 3D sphenoid sinus models provides a reliable and fast method for human identification.
• This approach can significantly assist forensic radiologists in improving identification accuracy.
• The framework effectively excludes low-similarity models and supports visual comparisons.