Development and validation of a deep learning model for radiographic classification of pediatric femoral neck fractures.

Authors

WenTao Wang, ShengHua He, LiHong Ou, Federico Canavese, Antonio Andreacchio, Chiara Arrigoni, Nunzio Catena, Marco Corradin, Mariabeatrice Damasio, Valentina Camurri, Francesco Candusso, Claudia De Cristo, Camilla De Luca, Sara De Salvo, Cristina Di Grigoli, Roberto Facchi, Andrea Florian, Cesare Gallo, Claudio Gargiulo, Rocco Grillo, Laura Giarratana, Emanuela Lanari, Veronica Lodovici, Giovanni Lucchesi, Ludovico Lucenti, Francesca Magnaguagno, Lorenza Marengo, Vittoria Mazzola, Angelo Musso, Luigi Aurelio Nasto, Vito Pavone, Andrea Pellegrino, Marco Ramella, Francesca Rizzo, Marco Sapienza, Nicola Stagnaro, Gianluca Testa, Andrea Vescio, XiaoLiang Chen, ChongZhi Zhao, DongKe Lai, Hang Liu, ChunXing Wu, ShunYou Chen, SuiGu Tang, QianQian Mei

Published in

Journal of orthopaedics and traumatology : official journal of the Italian Society of Orthopaedics and Traumatology. Jul 08, 2026. Epub Jul 08, 2026.

Abstract

The Delbet-Colonna (DC) classification guides treatment of pediatric femoral neck fractures (PFNFs) but relies on clinical experience. No deep learning (DL) model has been developed and validated to differentiate between PFNFs and proximal femoral growth plates (PFGPs) and classify PFNFs via DC classification, in order to overcome this limitation.
X-ray data including the annotations of 5555 PFGPs, 1306 PFNFs with various DC types, and 257 pediatric subtrochanteric femoral fractures (PSFFs), were prepared to construct a DL model based on the you-only-look-once (YOLO) model with wavelet transform (WT) and attention mechanism (AM) architectures. Two senior-level pediatric orthopedic surgeons (POS) performed the annotations independently by referring to the postoperative X-rays. The annotations were finalized if there were no differences. Otherwise, the two POS discussed and determined the final annotation. Thirty-one POS with different experience assessed the external testing dataset twice, without (first) and with (second) YOLO-WTAM model assistance. The rating performances of the YOLO-WTAM model and POS were evaluated. The kappa value reflecting reliability was obtained using a Fleiss kappa analysis.
According to the internal testing dataset, the area under the curve for different annotations ranged from 0.94 to 0.99. According to the external testing dataset, in the second round, the accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were greater than those in the first (P < 0.001): 79.17-87.16%, 83.30-87.47%, 94.68-96.57%, 74.62-79.44%, and 92.97-95.55%, respectively. Senior-level POS exhibited superior accuracy (P = 0.021), sensitivity (P = 0.013), specificity (P = 0.039), PPV (P = 0.004), and NPV (P = 0.025) in the first round but not in the second. The kappa value improved among residents (+27.36%), junior-level (+17.03%), mid-level (+26.66%), and senior-level (+17.07%) POS.
The YOLO-WTAM model can accurately differentiate between PFNFs and PFGPs and classify different DC types of PFNFs. This improves POSs' rating performance and reduces the need for experience in classifying PFNFs.
Level III.

PMID:
42420692
Bibliographic data and abstract were imported from PubMed on 09 Jul 2026.

Read full publication at:
Please sign in to see all details.

Stats

1-terrible, 9-excellent. How would you rate this publication? Sign in in to submit your rating.

Post a comment

You need to be signed in to post comments. You can sign in here.

Comments

There are no comments yet.