Deep learning for pediatric femoral neck fracture detection in a multicenter study

Summary

Pediatric femoral neck fractures (FNFs) are uncommon but may result in severe complications if undiagnosed. This study developed a deep learning model for automated detection and localization of FNFs on pediatric hip radiographs. The model was trained on 2,594 hip radiographs from 2,116 patients across eight centers. The optimal model (YOLOv11s) achieved a mean average precision at 0.5 IoU threshold (mAP@0.5) of 90.6% and an AUC of 0.921 on the internal test set, and a mAP@0.5 of 96.8% and an AUC of 0.968 on the external test set. To our knowledge, this represents one of the most comprehensive multicenter AI diagnostic studies for detecting pediatric FNFs. In a single-center reader study, AI assistance significantly improved diagnostic performance among emergency department orthopedic surgeons, particularly those with limited experience. These findings suggest the potential clinical utility of this model for supporting decision-making in emergency settings.

Graphical abstract

Highlights

• •The model automatically detects femoral neck fractures in pediatric hip radiographs
• •The model was trained and validated using 2,594 pediatric hip radiographs
• •The model exhibited high performance in both internal and external test sets
• •AI assistance significantly improved diagnostic accuracy for junior surgeons

Health sciences

Introduction

Femoral neck fractures (FNFs) in children are a highly disabling injury, accounting for approximately 1% of all pediatric fractures.^1,2 Failure to diagnose these fractures is strongly associated with several severe complications, including femoral head necrosis, premature epiphyseal plate closure, and hip deformities.^3,4,5 These complications substantially impair hip function and increase socioeconomic burden.^6,7 Therefore, early and accurate diagnosis of FNFs is critical for guiding treatment strategies and improving prognosis.^8,9

Although computed tomography (CT) and magnetic resonance imaging (MRI) offer superior diagnostic accuracy, radiography remains the first-line modality for initial evaluation of pediatric FNF in emergency settings.¹⁰ However, interpreting pediatric hip radiographs poses particular challenges for emergency physicians. The dynamic appearance of the epiphyseal plate on pediatric hip radiographs may be misidentified as fracture line, as illustrated in Figure 1A, and the subtle signs of nondisplaced fracture (see Figure 1B) are prone to being missed.^11,12,13 Furthermore, the scarcity of resources in primary healthcare institutions exacerbates the challenges of accurate diagnosis. These challenges are further compounded for clinicians with limited experience in pediatric musculoskeletal imaging and in resource-limited settings.

Figure 1.

Representative radiographs of pediatric femoral neck fracture

(A) The fracture line adjacent to the epiphyseal growth plate of the femoral head (red arrow). White arrows indicate the normal epiphyseal growth plates of the femoral head and greater trochanter.

(B) Impacted femoral neck fracture with fragment overlap (red arrow). White arrows mark the normal epiphyseal plate lines.

In recent years, artificial intelligence (AI) has advanced rapidly. Deep learning, in particular, has demonstrated significant potential in automating fracture detection on radiographs.^14,15,16 Some studies have shown that deep learning can perform at an expert level in diagnosing adult fractures at various anatomical sites.^17,18 However, translating these achievements to pediatric populations presents distinct challenges.¹⁹ Existing AI research on pediatric fractures remains in its early stages, with an excessive focus on upper limb injuries, particularly wrist and elbow fractures.^20,21 The critical and rare anatomical regions, such as the hips receive insufficient attention. Furthermore, most models have been developed using single-center datasets, necessitating further validation of these models’ generalizability through multicenter datasets.^22,23 These limitations constrain their practical application value in time-sensitive pediatric emergency departments.

Based on the current state of research, this study integrates multicenter pediatric hip radiograph data to develop and validate an AI model for automated detection and localization of FNFs. The study aims to evaluate the model’s performance in detecting and localizing these fractures and to quantify how AI assistance improves the diagnostic efficacy of emergency department orthopedic surgeons at various experience levels.

Results

Demographic characteristics

This multicenter study enrolled 2,116 patients with 2,594 pediatric hip radiographs from eight centers, including 1,127 pediatric FNF radiographs and 1,467 normal hip radiographs. Data from six hospitals (1,940 patients; 2,348 radiographs) were used for model development, while images from two additional hospitals (176 patients; 208 radiographs) formed the external validation set (see Figure 2). The patient age ranged from 0 to 18 years, with a peak age of 10–14 years (45.1%, n = 1,170). This finding aligns with the documented high incidence of pediatric FNFs.²⁴ The proportion of patients aged 0–4 years in the external validation set was significantly lower than in the internal validation set (4.8% vs. 20.8%). Male patients accounted for 61.4% of the total cohort (n = 1,594), consistent with the epidemiological characteristics of pediatric FNFs. The demographic analysis of the multicenter datasets is presented in Table 1.

Figure 2.

The flowchart of whole study

Flowchart illustrates the methodology and workflow of the study. Multicenter pediatric hip radiographs, including both fracture and normal cases, were initially collected following strict inclusion and exclusion criteria. After image preprocessing and annotation, datasets were partitioned into training, tuning, and internal and external test sets. Models such as faster R-CNN (with Vgg-16 and ResNet50), YOLOv8s, YOLOv11s, and YOLOv12s were trained, tuned, and validated. YOLOv11s model was selected as optimal model based on comparative performance. A multi-reader study using the external test set (n = 208) evaluated diagnostic performance with and without the assistance of the YOLOv11s model.

Table 1.

Patient demographics of datasets

Characteristics	Training set (n = 1899)	Tunning set (n = 242)	Internal test set (n = 245)	External test set (n = 208)	Total (n = 2594)
Age, (y)
0–4	346 (18.2)	46 (19.0)	51 (20.8)	10 (4.8)	453 (17.5)
5–9	513 (27.0)	65 (26.9)	64 (26.1)	80 (38.5)	722 (27.8)
10–14	856 (45.1)	108 (44.6)	107 (43.7)	99 (47.6)	1,170 (45.1)
15–18	184 (9.7)	23 (9.5)	23 (9.4)	19 (9.1)	249 (9.6)
Sex
Male	1,156 (60.9)	165 (68.2)	151 (61.6)	122 (58.7)	1,594 (61.4)
Female	743 (39.1)	77 (31.8)	94 (38.4)	86 (41.3)	1,000 (38.6)
Fracture categories
Fracture-positive	806 (42.4)	107 (44.2)	110 (44.9)	104 (50.0)	1,127 (43.4)
Fracture-negative	1,093 (57.6)	135 (55.8)	135 (55.1)	104 (50.0)	1467 (56.6)

Note. All cohorts present in this table correspond to imaging samples, where “n” denotes the number of hip X-ray radiographs in each dataset. Qualitative variables (age group, sex, and fracture categories) are reported as raw counts followed by percentages in parentheses.

Model performance

We performed a comparative analysis of five object detection models using both the internal test set (n = 245) and external test set (n = 208). Diagnostic performance was visualized via receiver operating characteristic (ROC) and precision-recall curves. These curves revealed distinct differences among the models. On the internal test set, the optimal model (YOLOv11s) achieved a mean average precision at a 0.5 IoU threshold (mAP@0.5) of 90.6% and a localization accuracy (Loc Acc) of 87.2%. The classification performance demonstrated notable robustness, with an area under the ROC curve (AUC) of 0.921 (95% CI: 0.883–0.952), a sensitivity of 87.2%, and a specificity of 89.5%. On the external test set, YOLOv11s achieved an mAP@0.5 of 96.8%, a Loc Acc of 94.3%, and an AUC of 0.968 (95% CI: 0.944–0.989), with sensitivity and specificity of 94.3% and 91.3%, respectively (Figure 3; Table 2).

Figure 3.

ROC and P-R curves of five models on test sets

Diagnostic performance of the five models visualized via ROC and P-R curves.

(A–D) ROC curve of five models on the internal test set (n = 245); (B) P-R curve of five models on the internal test set; (C) ROC curve of five models on the external test set (n = 208); (D) P-R curve of five models on the external test set. Detailed five models metrics are presented in Table 2.

ROC, receiver operating characteristic; AUROC, area under the ROC curve; P-R, precision-recall; AUPRC, area under the P-R curve; 95%CI, 95% confidence interval.

Table 2.

Diagnostic performance of five models in detection of pediatric femoral neck fracture

	Faster R-CNN with VGG16		Faster R-CNN with ResNet50		YOLOv8s		YOLOv12s		YOLOv11s
	Internal test set	External test set	Internal test set	External test set	Internal test set	External test set	Internal test set	External test set	Internal test set	External test set
Accuracy (%)	74.0% [67.8–79.8]	74.0% [67.8–79.8]	73.1% [67.4–78.4]	81.7% [76.0–87.0]	76.7% [71.8–82.0]	81.7% [76.4–87.0]	86.2% [81.6–90.6]	88.6% [84.1–92.8]	88.5% [84.1–92.2]	92.8% [88.9–96.1]
Sensitivity (%)	80.9% [73.2–87.8]	90.5% [84.6–95.8]	86.4% [79.8–92.4]	94.2% [89.4–98.2]	83.8% [76.8–90.4]	95.1% [90.7–99.0]	88.1% [81.8–93.8]	95.2% [90.8–99.0]	87.2 [80.8–92.9]	94.3% [89.8–98.1]
Specificity (%)	66.7% [60.0–74.2]	57.5% [48.1–66.4]	62.2% [53.9–70.4]	69.1% [60.0–77.9]	71.0% [63.6–78.5]	68.3% [59.4–77.1]	84.7% [78.1–0.9072]	81.9% [74.3–89.0]	89.5% [84.2–94.4]	91.3% [85.3–96.8]
PPV (%)	66.4% [58.6–74.1]	68.0% [59.9–76.1]	65.1% [57.4–72.9]	75.3% [67.6–82.2]	70.1% [62.1–78.3]	75.0% [67.7–82.3]	82.4% [75.2–89.4]	84.0% [76.9–90.4]	87.1% [80.2–92.9]	91.5% [85.5–97.0]
NPV (%)	81.1% [73.0–87.7]	85.9% [77.8–93.5]	84.8% [77.6–91.3]	92.3% [86.0–97.6]	84.3% [77.2–90.7]	93.4% [87.1–98.7]	89.8% [84.3–94.9]	94.5% [89.2–98.8]	89.6% [84.6–94.3]	94.1% [89.4–98.1]
mAP0.5	83.1%	86.9%	88.6%	95.0%	83.7%	90.4%	87.7%	92.1%	90.6%	96.7%
Loc Acc	80.9% [73.2–87.8]	90.5% [84.6–95.8]	86.4% [79.8–92.4]	94.2% [89.4–98.2]	83.8% [76.8–90.4]	95.1% [90.7–99.0]	88.1% [81.8–93.8]	95.2% [90.8–99.0]	87.2 [80.8–92.9]	94.3% [89.8–98.1]
AUC	0.814 [0.744–0.881]	0.897 [0.836–0.948]	0.880 [0.825–0.931]	0.946 [0.907–0.979]	0.868 [0.818–0.912]	0.940 [0.900–0.972]	0.906 [0.862–0.941]	0.955 [0.925–0.980]	0.921 [0.885–0.955]	0.968 [0.942–0.989]

Compare diagnostic performance of five models on the internal test set (n = 245) and the external test set (n = 208), where “n” denotes the number of hip radiographs in each test set. Diagnostic performance metrics include mAP0.5, Loc Acc, AUC, accuracy, sensitivity, specificity, PPV and NPV. Five models including faster R-CNN with VGG16, faster R-CNN with ResNet50, YOLOv8s, YOLOv12s, and YOLOv11s. With the exception of AUC, all other diagnostic performance metrics are reported as percentages. Values in square brackets are 95% confidence intervals (CIs), which were calculated using the bootstrap method.

The other models exhibited comparatively lower performance. For instance, the Faster R-CNN model with a VGG-16 backbone achieved an AUC of 0.814 (95% CI: 0.744–0.881) on the internal set and 0.897 (95% CI: 0.836–0.948) on the external set. The YOLOv8s and YOLOv12s models showed intermediate performance, with the YOLOv12s model ranking second to the YOLOv11s model. The YOLOv11s model demonstrated enhanced consistency in its performance across both internal and external test sets under the experimental conditions of this study, as evidenced by its higher mAP@0.5 and AUC metrics when compared to other architectures that were evaluated in this study (Table 2).

To comprehensively evaluate the model’s robustness and operational adaptability, a multifaceted evaluation was conducted. The analysis involved stratified performance analysis (see Table S1), operating threshold characterization (see Table S2), model reliability plots (see Figure S1), and decision curve analysis (see Figure S3). The model demonstrated good calibration, with a Brier score of 0.0526 (95% CI: 0.0306–0.0792).

Visualizations of model predictions and error analysis

Visualization of model predictions confirmed precise fracture localization. The corresponding HiResCAM heatmaps²⁵ reveal regions of high model activation, which align closely with both the predicted bounding boxes and the ground-truth (see Figure 4). After an in-depth analysis of the false-negative and false-positive cases with incorrect predictions by the model, it was found that false-negative cases often involved subtle fractures near the epiphyseal growth plate or mimicking the appearance of the epiphyseal growth plate. False-positive predictions were frequently triggered by normal anatomical structures, such as the sclerotic intertrochanteric line or normally developing epiphyseal plates. These structures were misinterpreted by the model as fracture lines (see Figure 5).

Figure 4.

Model predicted bounding box and HiResCAM visualizations for FNFS

(A–C) Original radiographs with YOLOv11s predicted bounding boxes and confidence scores highlighting suspected of FNF (light red and arrows).

(D–F) Corresponding HiResCAM heatmaps overlaid on the original radiographs. Regions with warm colors (red/yellow) indicate areas of high attention during the model’s decision making, co-localizing with fracture lines.

Figure 5.

False-negative and false-positive predictions of the YOLOv11s model

(Top row) case 1: nine-year-old male patient; (A and B) left hip radiograph: subtle displaced FNF. The yellow arrow highlights cortical angulation, a key radiographic sign. The deep red box indicates the actual fracture region which is the ground truth as annotated by experts; (B) false negative prediction by YOLOv11s: model misses fracture (no bounding box). (bottom row) case 2: fifteen-year-old female.

(C) Left hip radiograph: true negative case (no fracture present). No annotation as the ground truth.

(D) False-positive: model incorrectly marks region as fracture (light red box with a confidence score of 0.63).

Reader performance

The diagnostic performance of the five emergency department orthopedic surgeons (EDOSs1–EDOS5) from a single tertiary-care institution was evaluated on the external test set (n = 208). The readers had 2, 3, 5, 8, and 11 years of clinical experience, respectively. Without AI model assistance, the AUC values for the five EDOSs ranged from 0.755 (95% CI: 0.690–0.812) to 0.938 (95% CI: 0.902–0.967) (see Figure 6). Overall accuracy, sensitivity, and specificity exhibited a positive correlation with clinical experience. AI assistance was associated with improved junior surgeons’ (EDOS1-3, with 2–5 years of experience) diagnostic performance. In this group, AUC increased by 0.053–0.125. EDOS1 and EDOS2 showed the most notable improvements, with accuracy increasing by 10% and 11.9%, sensitivity by 9% and 9.2%, and specificity by 16.4% and 7.9%, respectively. EDOS3 also demonstrated an AUC increase of 0.053, alongside a 5.4% gain in accuracy and a 7.7% increase in sensitivity. However, the two senior surgeons (EDOS4-5, with 8 and 11 years of experience, respectively) maintained high baseline performance, with minimal changes following AI assistance (Table 3; Figure 6).

Figure 6.

ROC curves of emergency department orthopedic surgeons with and without AI assistance

ROC curves for five emergency department orthopedic surgeons (EDOS1-EDOS5, representing 2, 3, 5, 8, and 11 years of experience, respectively) interpreting the external test set, with and without the assistance of the AI (YOLOv11s) model. Dashed lines represent performance without AI support. Solid lines represent their performance with YOLOv11s assistance. AUC values quantify diagnostic accuracy AI-driven improvement is most pronounced for junior surgeons (EDOS1-3).

Table 3.

Diagnostic performance of five emergency department orthopedic surgeons (EDOS) on the study’s test sets (n = 208)

Metric	EDOS1	EDOS2	EDOS3	EDOS4	EDOS5
Experiences (years)	2	3	5	8	11
AUC (unaided)	0.755[0.690–0.812]	0.788[0.731–0.843]	0.841[0.788–0.88.9]	0.933[0.897–0.965]	0.938[0.902–0.967]
AUC (AI-aided)	0.880[0.836–0.923]	0.875[0.833–0.916]	0.894[0.847–0.934]	0.938[0.903–0.968]	0.942[0.907–0.972]
ΔAUC	+0.125	+0.087	+0.053	+0.005	+0.004
Accuracy (unaided)	75.5% [70.2–81.3]	78.7% [72.6–84.6]	84.1% [78.9–88.9]	93.3% [89.9–96.2]	93.7% [89.9–96.6]
Accuracy (AI-aided)	88.0% [83.7–92.3]	87.5% [82.7–91.4]	89.5% [85.1–93.8]	93.7% [89.9–96.6]	94.2% [90.9.97.1]
ΔAccuracy	+12.5%	+8.7%	+5.4%	+0.4%	+0.5%
Sensitivity (unaided)	81.8% [73.9–88.7]	84.5% [76.7–91.0]	87.6% [80.9–93.9]	93.4% [88.1–97.8]	95.2% [90.6–99.0]
Sensitivity (AI-aided)	90.5% [85.2–95.7]	94.2% [89.7–98.1]	95.3% [90.7–99.0]	95.2% [90.7–99.0]	96.1% [92.2–99.1]
ΔSensitivity	+8.7%	+9.7%	+7.7%	+1.8%	+0.9%
Specificity (unaided)	69.2% [60.2–77.7]	72.9% [63.7–81.6]	80.6% [72.6–88.0]	93.2% [88.0–97.2]	92.2% [86.5–96.9]
Specificity (AI-aided)	85.6% [78.4–91.9]	80.8% [72.6–88.0]	83.6% [761–90.2]	92.3% [86.3–96.8]	92.7% [86.8–97.0]
ΔSpecificity	+16.4%	+7.9%	+3.0%	−0.9%	+0.5%
PPV (unaided)	72.6% [64.6–80.5]	75.6% [67.5–83.3]	81.8% [74.4–88.6]	93.2% [88.3–97.2]	92.5% [87.3–97.1]
PPV (AI-aided)	86.3% [79.7–92.6]	83.1% [76.0–89.4]	85.3% [78.9–91.6]	92.3% [86.5–96.8]	92.5% [87.5–97.1]
NPV (unaided)	79.1% [70.9–87.2]	82.5% [74.7–89.6]	86.7% [79.1–93.2]	93.3% [88.2–98.0]	95.1% [90.6–98.9]
NPV (AI-aided)	89.9% [83.5–95.4]	93.3% [88.0–97.7]	94.7% [89.7–98.9]	95.1% [90.7–99.0]	96.0% [91.9–99.1]

Presents the diagnostic performance of five EDOS on the study’s test set (n = 208), where “n = 208” denotes the number of hip X-ray radiographs in the external test set. EDOS1–EDOS5 correspond to individual EDOS, and “year(s)” indicates each surgeon’s clinical experience.

EDOS = emergency department orthopedic surgeon. Δ = AI-assisted diagnostic metric value minus the corresponding unassisted value. Values in square brackets are 95% confidence intervals (CIs), which were calculated using the bootstrap method.

Discussion

In this multicenter study, we developed and validated a deep learning model for the automated detection and localization of FNFs in pediatric hip radiographs. The optimal model (YOLOv11s) exhibited robust diagnostic performance across both internal and external test sets. In a single-center reader study, AI assistance significantly enhanced the diagnostic accuracy of EDOSs, with the most pronounced benefits observed among those with limited clinical experience.

The novelty of this study lies in the collection of a multicenter pediatric hip radiographs and uses these rare datasets to construct an object detection architecture, which is specifically designed to detect and localize pediatric FNFs in pediatric hip radiographs. This study diverges from prior AI research on fracture detection in two aspects. The first aspect is previous studies on hip fractures have almost exclusively utilized adult cohorts,^{26,27,28,29,30,31,32} with the majority of participants being elderly individuals. Owing to the higher prevalence of hip fractures in this demographic, the datasets in these studies are typically larger in scale, and the resulting AI models have showed excellent performance. In a systematic review of AI-based intelligent diagnosis for hip fractures, Lex et al.³³ noted that among the 18 included studies, 11 (46.2%) employed AI models to diagnose FNFs from hip radiographs. The models demonstrated sensitivities ranging from 67.0% to 98.0% and specificities from 70.0% to 98.7%, with over 90% of the models attaining both sensitivity and specificity above 80%. In contrast, the present study focuses on the pediatric population, a distinct and understudied group. By integrating scarce hip radiograph data from eight tertiary medical institutions, we developed an AI model capacity of accurately identifying FNFs in pediatric hip radiographs, which feature relatively complex anatomical structures. The model achieved a sensitivity of 87.2% and a specificity of 89.5% in the internal test set and a sensitivity of 94.3% and a specificity of 91.3% in the external test set. Secondly, current AI research on intelligent diagnosis of pediatric fractures has predominantly focused on common upper extremity injuries,^34,35,36 accounting for approximately 60% of all AI-related studies on pediatric fractures,²¹ These AI models developed in these studies have also shown excellent performance. For instance, Choi et al.³⁷ developed a dual-input model to diagnose supracondylar humeral fractures using 2,532 anteroposterior (AP) and lateral elbow radiographs from multiple centers, with achieving an AUC of 0.92, sensitivity of 100%, specificity of 86.1%, and accuracy of 89.5% in the external test set. Zech et al.¹⁹ utilized the faster R-CNN algorithm for intelligent detection of various pediatric wrist fractures, including subtle buckle fractures, using 395 wrist radiographs from a single center. The model demonstrated an AUC of 0.92, sensitivity of 88.0%, and specificity of 89.0%. In comparison, the present study proposes an AI-based intelligent diagnostic system for pediatric hip fractures, and the performance of the developed YOLOv11s model is comparable to that of the aforementioned AI models for pediatric upper extremity fractures. To the best of our knowledge, this is the one of the most comprehensive multicenter AI diagnostic study specifically focusing on pediatric FNFs, thus addressing a research gap in this field.

Notably, this present study demonstrated that all five models exhibited superior performance on the external test set compared to the internal test set (Table 2). A comprehensive evaluation of the test sets reveals that this discrepancy in performance may be associated with variations in the composition of the test sets, and we explicitly acknowledge the potential for case-mix and temporal biases to be contributing factors. Retrospective comparisons between the two test sets revealed that the internal test set contained a higher proportion of diagnostically challenging cases, including a greater share of very young children and a broader distribution of complex Delbet types (see Table S1). This observation aligns with the study of Till et al., who demonstrated that model performance metrics change with the proportion of difficult cases in the test set.³⁸ Therefore, the external set performance is likely to reflect the model’s efficacy in handling common cases encountered in daily practice, while the internal set performance provides a more stringent assessment of its capability against complex presentations. Although systematic stratified analyses by hospital site or imaging device manufacturer were not performed, the independence and heterogeneity of the external test set still support preliminary cross-institutional applicability. It is evident that this has direct implications for clinical deployment. The model demonstrates high potential as a triage tool for common cases, while expert review remains essential for complex cases.

A detailed analysis of the visualized bounding boxes predicted by the YOLOv11s model on the external test set revealed that false-negative and false-positive cases were mainly concentrated near the epiphyseal plate of the femoral head and greater trochanter regions. These regions pose challenges in diagnosing pediatric FNFs in hip radiographs. These challenges arise from anatomical variations across age groups and the strong resemblance between normal developmental features and subtle fractures. High-energy injuries may produce overlapping fracture, which can obscure fracture line. Low-energy injuries may cause non-displaced fractures or minimal displacement fractures that closely resemble the morphology of the epiphyseal plate. Both situations pose considerable diagnostic challenges for less experienced EDOSs, who lack training in interpreting pediatric musculoskeletal images. The AI model, which was trained using a diverse multicenter cohort in this study, can identify these subtle features, highlighting its potential as a decision support tool in emergency clinical settings for pediatric hip fractures.

This study demonstrates the feasibility and performance of a deep learning model (YOLOv11s) for automated detection and localization of FNFs in pediatric hip radiographs using a large multicenter dataset. The model achieved robust performance both on internal and external test sets and provides clinically significant diagnostic improvements, especially for junior orthopedic surgeons in a single-center reader study. These findings highlight the potential of this AI tool to support clinical decision-making and reduce diagnostic errors in pediatric emergency settings.

Limitations of the study

This study has several limitations. (1) The model was developed and validated solely for pediatric FNFs on AP-view radiographs. Although the AP view is the standard initial imaging projection for hip trauma, this focused scope limits the model’s immediate applicability to other pediatric fracture sites or additional radiographic views. (2) As a multicenter retrospective study, this research is susceptible to selection bias, and the cohort may not fully represent all emergency department patients presenting with hip pain. Furthermore, the external test sets obtained from the other two hospitals underwent manual screening and processing, resulting in a positive-to-negative ratio of 1:1. Such a discrepancy in this ratio may have implications for the generalizability of the model, particularly predictive values in real-world settings, where positive predictive value may be overestimated due to the lower true prevalence of pediatric FNFs. (3) The number of readers in our study was relatively small (n = 5), all of whom were from a single tertiary-care institution. This limitation restricts the generalizability of the observed performance improvement. Future research should focus on expanding the dataset with more cases from diverse sources, conducting prospective multicenter validation in real-world clinical workflows, and extending the model’s capability to other pediatric fracture types, aiming to enhance its generalizability and clinical utility.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Zhewei Ye (yezhewei@hust.edu.cn).

Materials availability

• •An iteration of the trained deep learning model is available upon request to the lead contact.
• •Patient medical images cannot be publicly shared due to institutional regulations and patient privacy restrictions.

Data and code availability

• •We cannot deposit the patients’ medical images used to train the deep learning model in this study in a public repository due to institutional and legal regulations regarding patient confidentiality and imaging. Contact the lead contact to request access.
• •The code used in the development of the deep learning models for this study has been deposited in a publicly available GitHub repository [https://github.com/nulling00/fracture-detection-code]. This repository includes scripts for model training, inference, and key analysis steps. The specific environment details (e.g., Conda/Pip environment files) are included to facilitate reproducibility. If there is any extra information needed to reanalyze the data of this paper, it is available upon reasonable request to the lead contact.

Acknowledgments

We sincerely thank Prof. Zhewei Ye for his invaluable guidance in designing the research and revising the manuscript. We also thank Prof. Xu Zou from the Department of Artificial Intelligence and Automation at Huazhong University of Science and Technology for their technical support in constructing, training, validating, and testing the model, as well as our colleagues at the participating multicenter medical institution for generously providing the pediatric femoral neck fracture dataset.

This work was supported by the National Natural Science Foundation of China (grant no. 82172524), Research Project on Educational and Teaching Reform of Ningxia Medical University (no. NXJY2025057), and Health and Science and Technology Project of Hubei Province (no. WJ2025Q027).

Author contributions

X.C., M.X., and X.W. are co-first authors and contributed equally to this work; Z.Y., T.H., and P.L. are co-correspondents; the contributions of each author were listed using the CRediT taxonomy; conceptualization, X.C., M.X., X.W., L.J., P.L., T.H., and Z.Y.; methodology, X.C., M.X., X.W., T.Z., N.L., H.X., L.H.C., P.L., and Z.Y.; data curation, X.C., M.X., X.W., L.J., P.L., and J.H.; investigation, X.C., J.Y., and X.W.; software, resources, and validation, T.H., L.H.C., T.Z., N.L., H.X., and W.H.T.; formal analysis and visualization, X.C., M.X., T.H., and X.W.; writing – original draft, X.C., M.X., and X.W.; writing – review & editing, X.C., M.X., P.L., T.H., and Z.Y.; funding acquisition, supervision, and project administration, P.L., T.H., and Z.Y.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT (Open AI) solely for language polishing and grammatical refinement of non-scientific content. All critical scientific components—including research design, data analysis, and result interpretation—were exclusively conducted by human authors. The final manuscript has been thoroughly reviewed and approved by all co-authors who take full responsibility for its academic integrity and accuracy.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
Faster R-CNN	Ren et al.39	arXiv:1506.01497
Vgg-16	Simonyan and Zisserman.40	arXiv:1409.1556
ResNet-50	He et al.41	arXiv:1512.03385
YOLOv8s	Ultralytics (2023).YOLOv8	arXiv:2408.15857
YOLOv11s	Ultralytics (2024).YOLOv11	arXiv:2410.17725
YOLOv12s	Ultralytics (2025).YOLOv12	arXiv:2502.12524
PyTorch 1.9.1	Paszke et al.42	arXiv:1912.01703

Experimental model and study participant details

Study design and ethics

This retrospective multicenter study was received ethical approval from the Ethics Committee of the Union Hospital of Tongji Medical College, Huazhong University of Science and Technology. (approval number UHCT250500; informed consent waived). The workflow of whole study is summarized in Figure 2.

• •This study utilized de-identified medical images of patients who underwent hip X-ray examinations in eight medical institutions.
• •The age, gender, and other relevant information of the patient cohort in this study are detailed in Table 1 of the manuscript.

Method details

Datasets collection

The overall workflow of whole study is illustrated in Figure 2. This retrospective study collected plain hip radiographs of children aged 0–18 years old with suspected hip injuries who were referred to eight tertiary care hospitals in China between March 2014 and September 2024. Inclusion criteria for the images were:1) age ≤18 years old; 2) standard anteroposterior hip radiograph; 3) final institutional radiology report indicating either normal or femoral neck fracture. Exclusion criteria for the images were: 1) age >18 years old; 2)Images showing implants, cast fixation, or other pathologic lesions in the bone; 3) improper positioning; 4) pathologic fractures or fractures with a high suspicion of child abuse (e.g., metaphyseal angle fractures). In the case of a particular patient who underwent X-ray procedures during several medical visits, only the fracture X-rays taken during the initial visit were incorporated into the study. The development of the model incorporated a total of 2,386 radiographs from 1,940 patients from six hospitals. The 2,386 pediatric hip radiographs were randomly divided into training, tuning, and internal test sets in an approximate ratio of 8:1:1. Additionally, 208 radiographs from two other hospitals served as an independent external test set to verify the generalization ability of the model, including 104 images with fractures and 104 without. The distribution of the dataset and specific demographic information is detailed in Table 1.

Dataset annotation and allocation

All images were anonymized, retaining only basic demographic information (age and sex). The 2,594 radiographs collected in Digital Imaging and Communications in Medicine (DICOM) format were converted and saved as Joint Photographic Panel (JPG) format files. Image review and annotation were performed independently by a senior radiologist (J.L.) and a pediatric orthopedic specialist (L.P.), both with more than 15 years of clinical experience. The labeling procedure was performed as follows: The pre-downloaded LabelMe software was launched according to its operational protocol, and all images meeting the inclusion and exclusion criteria were imported. Fractured areas of the femoral neck were marked with rectangular boxes, using the smallest possible boxes to avoid oversized labels that might compromise the model’s performance. The labeled images were saved with the tag “fracture,” and the final labeled images with rectangular boxes were stored as “json” files alongside the original images. Normal images remained unprocessed. Prior to the adjudication of the final annotation results, an analysis of the inter-reader agreement between the two annotators was conducted. Inter-observer agreement was calculated using Cohen’s kappa statistic for classification and mean Intersection-over-Union (IoU) for localization.The results showed that the inter-observer agreement was substantial, with a Cohen’s kappa of 0.974 for fracture classification and a mean IoU of 0.72 for bounding box localization. For cases with diagnostic discrepancies, the patient’s CT findings served as the gold standard. When CT data were unavailable, the final diagnosis and annotation were determined through a consensus of a pediatric orthopedic specialist with 30 years of clinical experience (see Figure S2). The gold standard was used to evaluate model performance and subquent research once all annotations were completed.

Model development and training

Model architerture

A single-stage YOLO (You Only Look Once) detection framework was developed to automatically identify FNF in pediatric hip radiographs. The model was designed to address the subtle, low-contrast, and occasionally discontinuous cortical features typical of pediatric FNFs, which require both high localization precision and rapid inference for emergency decision-making. To comprehensively evaluate detection performance, five object detection architectures were implemented, including three single-stage YOLO variants (YOLOv8s, YOLOv11s, and YOLOv12s) and two two-stage Faster R-CNN models using VGG-16 and ResNet-50 as backbone networks.

The YOLO models integrate feature extraction, localization, and classification within an end-to-end single-stage pipeline, providing real-time performance and enhanced sensitivity to small and low-contrast targets.⁴³ Each variant follows a consistent structural principle composed of a hierarchical feature-encoding backbone (CSPDarknet) that preserves cortical edge continuity, a multi-scale fusion neck (PAN–FPN) that merges semantic and spatial representations to strengthen small-target detection, and a decoupled head that performs classification and bounding-box regression in parallel to generate confidence-weighted coordinates. In the present study, YOLOv11⁴⁴ introduces two improved modules: the C3K2f structure in the backbone to enhance gradient propagation, and the C2PSA attention module in the neck to improve detection of subtle, low-contrast features. Compared to YOLOv8 (C2f structure) and YOLOv12 (ABlock + RELN modules), YOLOv11 provides a more efficient balance between accuracy and computational cost (see Figure S4).

As two-stage baselines, Faster R-CNN models employed VGG-16 and ResNet-50 as backbone networks. The Region Proposal Network (RPN) was first tasked with generating candidate regions, followed by classification and bounding-box regression (see Figure S5). The purpose of these models was to establish a benchmark for the detection accuracy against computational efficiency relative to the single-stage YOLO series.

Among all models, YOLOv11s achieved the best overall performance and was selected for downstream analysis.

Model training details

All models were trained following a two-stage protocol. In the pre-training phase, 808 fracture-positive images were used for 300 epochs to reinforce cortical feature extraction. In the fine-tuning phase, a balanced dataset containing 808 fracture and 202 normal images (ratio 4:1) was used to refine detection performance. Model training was conducted in a Conda virtual environment configured with Python 3.8.20,PyTorch 1.9.0+cu111, TorchVision 0.10.0+cu111, OpenCV 4.11.0.86, NumPy 1.24.4, and Ultralytics 8.3.143. Data augmentation techniques included Mosaic augmentation, random horizontal flipping (p = 0.5), rotation (±10°), and brightness–contrast adjustments (±20%).

For YOLO models, input images were resized to 640 × 640 pixels and normalized to [0,1]. Training employed the Adam optimizer with an initial learning rate of 0.002 following a cosine decay schedule, with a batch size of 32, and a total of 300 epochs. The loss function combined cross-entropy, CIoU, and objectness losses. Automatic Mixed Precision (AMP) and Exponential Moving Average (EMA) mechanisms were enabled to enhance convergence and inference stability. For Faster R-CNN models, input image size was fixed at 600 × 600 pixels, and anchor scales were set to [8, 16, 32]. Training adopted the Adam optimizer (initial learning rate = 1 × 10⁻⁴, decayed to 1 × 10⁻⁶ via cosine scheduling) for 200 epochs. Model checkpoints were saved every 10 epochs, and early stopping was triggered when validation performance showed no improvement for 30 consecutive epochs.

Inference and validation strategy

Model validation was performed every 10 epochs using a held-out validation set. During inference, YOLO models applied a confidence threshold of 0.25 and an NMS IoU threshold of 0.5, while Faster R-CNN models retained predictions with confidence scores greater than 0.5.

All YOLO models converged within 140–180 epochs, whereas Faster R-CNN models converged around 100 epochs. The overall training time ranged from 35 to 45 h, with GPU utilization above 99%. Among all models, YOLOv11s demonstrated the most stable convergence curve and highest generalization ability, achieving both high accuracy and low latency, and was therefore selected as the final model for clinical validation and multi-reader study.

Multi-reader study

A multi-reader study was conducted to evaluate the diagnostic efficacy of an AI model in assisting emergency department orthopedic surgeons across varying experience levels in detecting pediatric FNFs on hip radiographs. An external test set (n = 208) was utilized for this study. The sample size, which was determined by the held-out external test set size, was sufficient for the initial feasibility assessment. Five board-eligible orthopedic surgeons from Emergency Department of Wuhan Union Hospital were recruited and stratified into junior group (EDOS1-3, with two, three, and five years of clinical experience, respectively) and senior surgeons (EDOS4-5, with eight and eleven years of clinical experience). All participants had at least one year working experience in interpreting orthopedic radiographs and declared no conflicts of interest related to the study. During the AI-assisted reading session, readers were presented with the original radiograph alongside the model’s prediction boxes and confidence scores via a separate interface. All readers were informed that AI outputs was reference only and each reader should be made independent judgment based on their own interpretations. Diagnostic confidence for each case was scored using a 5-point Likert scales (1 = definitely fracture-negative, 2 = probably fracture-negative, 3 = uncertain, 4 = probably fracture-positive, 5 = definitely fracture-positive).³⁷ A case was considered a true positive only if the reader’s confidence score was ≥4 and their annotated bounding box achieved an IoU of ≥0.5 with the ground truth; otherwise, it was considered an incorrect diagnosis. To minimize memory bias, the two reading sessions were separated by at least one month, and image order was randomized in each session. Diagnostic performance metrics, including accuracy, sensitivity, specificity, missed diagnosis rate, misdiagnosis rate, PPV, and NPV, were calculated for two groups of readers. The study had a double-blind design, ensuring all readers were blinded to patient’s clinical information and gold-standard diagnoses.

Quantification and statistical analysis

Statistical analysis of data in this study were performed using Python (scikit-learn 1.2.2). All statistical details, including exact “n” values, estimates of performance metrics, and 95% confidence intervals, are reported in the results section, Table 1, and corresponding table legends.

Diagnostic performance was evaluated at the image level using detection, localization and classification performance metrics. The training set, tuning set, internal test set and external test set analyzed in this study are summarized in Table 1, where “n” represents the number of radiographs in each dataset. Demographic characteristics of these sets (age group, sex, fracture categories) are presented as raw counts and percentages (Table 1), which reflect the distribution of qualitative variables in the study.

For detection and localization, we calculated the mean average precision at an intersection-over-union(IoU) threshold of 0.5 (mAP@0.5) and the localization accuracy (Loc Acc). Loc Acc was defined as the proportion of fracture sites accurately localized when the IoU between model-predicted bounding boxes and ground truth annotations was ≥0.5, using a confidence threshold of 0.25. This threshold was selected based on preliminary analysis optimizing the balance between sensitivity and specificity for clinical deployment scenarios. Classification performance was assessed using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC). These performance metrics are reported as point estimates paired with 95% confidence intervals in the results section and corresponding tables. To comprehensively evaluate the model’s robustness and operational adaptability, we further conducted calibration analysis (Brier score and reliability curves) and decision curve analysis.

The statistical analysis incorporated rigorous methods to ensure robust inference. Bootstrap resampling with 1,000 iterations was used to calculate 95% confidence intervals(95% CI). For validating the model’s performance in clinically relevant subgroups, metric evaluations were conducted by age and Delbet fracture type stratification. The same Bootstrap resampling method was applied in stratified analyses to produce confidence intervals for two subgroup. ROC and P-R curves were generated with the corresponding AUC computed using scikit-learn 1.2.2.

In the multi-reader study, diagnostic performance was assessed using the same classification metrics. Reader improvements were quantified as the difference (Δ metrics) between AI-assisted and unassisted values. A dual-criterion approach required both a fracture probability score ≥4 on the 5-point Likert scale and accurate localization (IoU ≥0.5) for a case to be considered a true positive.

Published: December 8, 2025