Determination of Modified Waldenström Staging in Legg-Calvé-Perthes Disease Using Deep Learning.
Authors
Affiliations (5)
Affiliations (5)
- Division of Orthopaedics, Children's Hospital of Philadelphia, Philadelphia, PA, USA.
- Division of Pediatric Orthopaedic Surgery, Hospital for Special Surgery, New York, NY, USA.
- Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA.
- Gillette Children's Specialty Healthcare, St Paul, MN, USA.
- Department of Orthopedic Surgery, University of Minnesota, Minneapolis, MN, USA.
Abstract
Age and disease stage are critical factors in guiding management in Legg-Calvé-Perthes disease (LCPD). The modified Waldenström system is the most common staging system for LCPD but has only moderate inter-rater reliability. Deep learning (DL) has gained popularity in orthopaedics due to its potential to improve consistency/efficiency in determination of image-based classifications. The current study aimed to develop a DL model for determination of modified Waldenström staging from standard hip radiographs. Paired anteroposterior (AP) and frog-lateral hip radiographs from patients with confirmed LCPD treated at 2 tertiary care children's hospitals were included for model training/validation, hold-out testing, and external validation. Images of patients with skeletal dysplasias or other conditions were excluded. A modified Waldenström stage was assigned by the senior authors for each radiograph. A machine learning classification pipeline leveraging DL-automated extraction of clinically relevant and quantifiable parameters on AP/lateral radiographs was developed. Model performance was assessed on 2 separate LCPD classification schemes: early (Ia-IIa) vs late (IIb-IV) and the full classification. A total of 2,164 images were included from institution 1. The DL pipeline had excellent segmentation metrics (dice coefficient = 0.93) and was able to extract and calculate epiphyseal parameters on AP/lateral radiographs within 12 s. On the hold-out testing set (n = 229 radiograph pairs), the model had an area under the receiver operating characteristic curve (AUROC) of 0.82 in classifying LCPD as either early or late stage. The most predictive parameters driving model decision-making were AP epiphyseal sclerosis and epiphyseal area. On external validation at institution 2 (n = 533 radiograph pairs), the AUROC was 0.75 when categorizing early vs late stages. The DL pipeline provided high accuracy in discerning early vs late stages of LCPD based on imaging parameters and moderate accuracy on the full classification, performing better at more advanced stages. Training the model on larger samples with more representation at earlier stages may improve accuracy. This has the potential to expedite and standardize staging for both clinical encounters and research. (1)Deep learning can accurately classify early vs late-stage LCPD from standard hip radiographs.(2)Automated segmentation provides objective and quantifiable assessment of femoral morphology parameters pertinent to LCPD.(3)Artificial intelligence models could be applied to both clinical practice and research methodology as an automated tool for Waldenström staging.(4)Larger, multicenter datasets with greater representation of early stage LCPD imaging are needed to improve model performance and generalizability. III.