Orbital Fracture AI Platform (DenseNet169-UNet)

Brief overview

This study uses DenseNet-169 and U-Net architectures to classify orbital blowout fractures and segment fracture regions from orbital CT images, leveraging 3,016 retrospectively collected scans from a single Chinese tertiary center. The framework distinguishes fracture vs non-fracture cases, lateralization (left vs right eye), and performs pixel-wise fracture segmentation with reported performance of AUC ~0.99 for classification and Dice coefficient ~0.89 for segmentation, alongside IoU ~0.82. Training used ImageNet-pretrained weights with internal 8:1:1 splitting and 5-fold cross-validation, with strong performance consistency across held-out test folds. Preprocessing included OpenCV-based ROI extraction and data augmentation via random rotations, with all annotations generated by consensus of expert radiologists using LabelMe. No code repository, pretrained weights, or deployment interface are reported, and the model remains entirely internally validated without external cohort testing or clinical implementation.