Deep learning for orbital fracture detection and reconstruction: A systematic review on diagnostic accuracy and surgical planning

Abstract

Objective

To systematically review the efficacy of deep learning (DL) models in detecting and reconstructing orbital fractures based on computed tomography (CT) imaging, assessing their diagnostic accuracy, processing time, and role in surgical planning.

Method

A systematic search was conducted in PubMed, Embase, Web of Science, Wiley, Cochrane, and additional sources. Five studies met the inclusion criteria. Performance was evaluated using accuracy, sensitivity, specificity, area under the curve (AUC), Dice Similarity Coefficient (DSC), and Intersection over Union (IoU).

Results

Deep learning models, particularly U-Net, GAN-based approaches, and SPAK-guided architectures, demonstrated high accuracy in fracture detection and reconstruction. DenseNet achieved the best fracture identification performance (AUC = 0.99). SPAK-based models improved reconstruction precision, reducing geometric errors. Automated segmentation reduced processing time from 25 min to less than 5 min per case, and GAN-based models optimized surgical planning, lowering it to 1.5 min.

Conclusions

Deep learning enhances orbital fracture diagnosis and reconstruction accuracy, significantly reducing segmentation and planning time. However, further comparative studies are needed to standardize methodologies and validate clinical applicability.