A novel clinical investigation using deep learning and human-in-the-loop approach in orbital volume measurement

Abstract

Orbital volume assessment is crucial for surgical planning. Traditional methods lack efficiency and accuracy. Recent studies explore AI-driven techniques, but research on their clinical effectiveness is limited.

This study included 349 patients aged 19 years and above, who underwent three-dimensional facial computed tomography (3DCT) without orbital trauma or congenital anomalies. To construct an AI training dataset, manual segmentation was performed on 178 patients’ 3DCT using 3D Slicer. The remaining data of 171 patients underwent human-in-the-loop method, resulting in a dataset of 349 annotated samples. Comparative analysis of Dice coefficients and execution speeds was performed between manual and semi-automated segmentations.

Comparing AI-assisted semi-automated segmentation with manual segmentation, all six annotators demonstrated lower average inference times without a significant difference in Dice coefficients (90.31% vs. 88.72%). For 178 patients’ 3DCT, a high average Dice coefficient of 89.9% was observed, and a 38.42-ms inference time was recorded. For the full dataset, the AI model achieved a high average Dice coefficient of 94.1% and a fast average inference time of 32.55 ms per axial slice.

This study demonstrates the potential of AI for maintaining high accuracy and time-efficiency in orbital region segmentation, with wide clinical applications.