Automated skeletal maturity staging from hand-wrist radiographs using deep learning models: A decision support system based on Fishman’s indicators

Abstract

Objective This study aimed to develop an automated system for skeletal maturity staging using deep learning (DL) models of hand-wrist radiographs based on Fishman’s method. Methods In total, 2,318 hand-wrist radiographs of patients aged 8–19 years were retrospectively analyzed. Each radiograph was labeled by experts according to Fishman’s method. This assessment was based on 11 skeletal maturity indicators (SMIs) across six anatomical regions, including the sesamoid bone, third and fifth fingers, and radius. Five different DL models (Faster R-CNN, FCOS, RetinaNet, SSD, and YOLOv7) were trained and evaluated. The model performance was assessed using commonly accepted metrics, including mean average precision (mAP), accuracy, precision, recall, and F1-score. In addition, k-fold cross-validation (k = 5) was applied to ensure the robustness of the results. Results The models achieved mAP values ranging from 0.81 to 0.93, indicating the effective detection of the relevant SMIs. The Faster R-CNN demonstrated the highest overall performance (mAP = 0.93, accuracy = 0.73, precision = 0.74, localization recall = 0.99, and F1-score = 0.83). For individual SMIs, the Faster R-CNN model achieved classification accuracies ranging from 0.45 to 0.90. Similar trends were observed for precision, recall, and F1-score. The ground-truth annotations and model predictions were also presented for visual comparison. Conclusions All evaluated DL models, particularly Faster R-CNN, showed strong potential for the reliable and automated detection of Fishman’s SMIs. This approach may serve as a valuable clinical decision-support tool for orthodontists in growth assessment and treatment planning.