Development and Validation of a Deep Learning Decision-Support Model for Aesthetic Chin-Shortening Surgery

BACKGROUND: Drawer genioplasty (DG) and en-bloc mandibular U-shaped osteotomy (UO) are commonly used options for long-chin reduction, but the choice between them remains subjective. We developed deep learning models to support this decision from routine imaging. METHODS: In this single-center study, 66 adults with long chin deformity (33 DG, 33 UO) had preoperative craniofacial CT and 5 facial photographs. We built 3 pathways: CT-only (3D ResNet and slice-wise 2D ResNet with channel-spatial attention), photograph-only (ResNet-50 and ConvNeXt-Small), and multimodal CT+photograph late fusion with attention and a mandibular region of interest (ATT+ROI). Patient-level outputs were obtained by averaging slice or image probabilities, and the data set was randomly split into training, validation, and test sets. Accuracy, F1-score, and AUROC were calculated; Grad-CAM was used to inspect anatomic focus. RESULTS: 2D ResNet-SEP-LastHalf-ATT reached an accuracy of 0.80, outperforming the 3D CT model (accuracy 0.60). The photograph ConvNeXt model achieved an accuracy and AUROC of 0.90. A multimodal probability-fusion model (CT-3D ResNet-18 plus photo-ResNet-50) showed accuracy 0.80, whereas the ATT+ROI multimodal model achieved accuracy 0.90, F1-score 0.90, and AUROC 0.86 with fewer misclassifications. Grad-CAM indicated consistent attention on the bony chin, mandibular border, and cervicomental junction. CONCLUSIONS: A multimodal attention-guided model based on routine CT and photographs can reproduce the surgeon's choice between DG and UO in a long-chin cohort. To our knowledge, this is the first preliminary deep learning model to support selection between specific chin-shortening osteotomies and may provide a practical decision-support tool for aesthetic chin-shortening surgery.