Automatic orbital segmentation using deep learning-based 2D U-net and accuracy evaluation: A retrospective study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Surgical Planning

Background:

• Orbital fracture surgery often requires 3D anatomical models for reference.
• Creating these models manually from computed tomography (CT) scans is complex and time-consuming due to the thin, intricate nature of orbital bone.

Purpose of the Study:

• To assess the clinical applicability of deep learning (DL)-based automatic segmentation (AS) for fractured orbital CT images.
• To determine if DL-generated segmentation is accurate enough for surgical support.

Main Methods:

• A U-Net deep learning model was trained on 115 CT scans for automatic segmentation (AS) of orbital fractures.
• The AS accuracy was validated using Dice coefficients and average symmetry surface distance (ASSD).
• Four experienced surgeons evaluated 3D-printed models derived from the AS output.

Main Results:

• The DL model successfully performed AS on all 125 CT scans.
• High accuracy was achieved: Dice coefficient of 0.860 ± 0.033 and ASSD of 0.713 ± 0.212 mm.
• Expert surgeons deemed the AS output suitable for surgical support without modifications.

Conclusions:

• The developed DL-based AS algorithm for orbital fractures demonstrates high accuracy and efficiency.
• This method facilitates rapid, low-cost 3D model generation, potentially enhancing surgical safety and precision.
• The findings suggest significant clinical utility for AI in orbital fracture management.