Journal of Pediatric Orthopaedics - 2026-04-07 - Journal Article
Automated Detection of Pediatric Slipped Capital Femoral Epiphysis: A Deep Learning Approach Using Anatomically Informed Attention Guidance.
Chakladar S, Pereira DE, Shariati J, Hosseinzadeh P
Topics
Key Takeaway
An attention-guided deep learning model (U-Net++ segmentation + EfficientNet B1 classifier) detected SCFE on pelvic radiographs with AUC 0.893, 93.3% sensitivity, and 90.0% specificity on a 35-image test set.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study developed and validated a two-stage deep learning pipeline to classify pelvic radiographs as SCFE or no-SCFE, addressing known inter-observer variability of Klein's line. A U-Net++ model first segmented the femoral head to define an anatomic region of focus, then an attention-guided EfficientNet B1 model performed binary classification. On the 35-image holdout test set, the model achieved AUC 0.893, accuracy 91.4%, sensitivity 93.3%, and specificity 90.0%, with Grad-CAM confirming anatomically appropriate feature localization.
Key Limitation
A test set of 35 radiographs is insufficient to characterize performance across SCFE severity subgroups (mild vs. severe Southwick angle, stable vs. unstable by Loder classification), making generalizability to clinical deployment unknown.
Original Abstract
BACKGROUND
Slipped capital femoral epiphysis (SCFE) is an adolescent hip disorder that is often missed on initial presentation due to subtle radiographic findings, leading to significant complications. Traditional diagnostic methods like Klein's line are subjective and prone to inter-observer variability. This study aimed to develop and validate an attention-guided deep learning model to automatically identify SCFE on pediatric pelvic radiographs.
METHODS
A total of 174 pelvic radiographs (139 training, 35 testing) were retrospectively collected. A two-stage model was developed: first, a U-Net++ model segmented the femoral head to define an anatomic region of focus (ROF). Second, an attention-guided EfficientNet B1 model, trained to focus specifically within the ROF, classified radiographs as "SCFE" or "no SCFE." Model performance was evaluated on the testing set using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.
RESULTS
The SCFE detection model demonstrated an AUC of 0.893, accuracy of 91.4%, sensitivity of 93.3%, and specificity of 90.0%. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed that the model's predictions were based on the anatomically relevant ROF.
CONCLUSIONS
The attention-guided deep learning model detected SCFE with high diagnostic accuracy. The anatomically informed attention mechanism provides interpretability, improving its utility within clinical settings. This automated tool demonstrates potential as a clinical decision-support system to reduce diagnostic delays and improve the consistency of SCFE detection.
LEVEL OF EVIDENCE
Level III.