Deep Learning in Scaphoid Fracture Detection and Healing Prediction: A Systematic Review of Artificial Intelligence...

Scaphoid fractures are the most common carpal bone injuries and continue to pose diagnostic and therapeutic challenges due to their tendency for delayed union or non-union. Interpretation variability and delayed detection remain key contributors to missed injuries and adverse outcomes. Deep learning (DL) models, particularly convolutional neural networks (CNNs), have shown strong performance in musculoskeletal imaging and offer potential to improve diagnostic accuracy and prognostication in hand surgery. A systematic review of PubMed, Embase, and PMC was conducted covering January 2015 to March 2025, following Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) 2020 guidelines. Studies were eligible if they applied DL algorithms to detect or classify scaphoid fractures or to predict union and non-union, and reported at least one diagnostic metric (accuracy, area under the curve (AUC), sensitivity, or specificity). Traditional machine-learning approaches, cadaveric studies, and non-peer-reviewed publications were excluded. Fourteen peer-reviewed studies met the inclusion criteria. Twelve evaluated DL for fracture detection, and two assessed fracture-healing prediction. Across detection studies, CNN-based models reported accuracies of 81-96% and AUCs up to 0.97. Transfer-learning architectures (DenseNet, ResNet, EfficientNet) consistently outperformed custom CNNs, particularly in multicentre datasets. Multi-view fusion of anteroposterior and lateral radiographs improved recall by approximately 12 percentage points compared with single-view analysis. Segmentation-enhanced models showed notable gains in occult fracture detection, identifying up to 41% of occult injuries compared with 6.8-13.7% by clinical experts. AI-augmented decision support improved novice radiologist performance, increasing AUC by 9-14 percentage points. For healing prediction, a custom CNN achieved 93.6% accuracy for post-surgical union, while a YOLOv5-ResNet-50 system classified union, non-union, or osteonecrosis with 91% accuracy (AUC 0.96). DL models demonstrate radiologist-level performance for scaphoid fracture detection and show encouraging potential for predicting union. Approaches incorporating segmentation, transfer learning, and multi-view inputs appear particularly promising for clinical workflows. Although early results support integration of AI-assisted tools into diagnostic pathways, robust multicentre validation and explainability frameworks remain essential before routine clinical implementation.