Deep learning for hip fracture classification: external validation and reader-assistance evaluation

Development used 4,493 images from two institutions; external validation used 675 images from another institution.

Development (4,493), external validation (675) and reader comparison (341) images are shown separately. The 341 are part of the external set; do not add the bars. Table III's 0.83–0.89 values are averages for three doctors and are not labelled as standalone AI sensitivity because descriptions of the evaluator are inconsistent.
Bone & Joint Journal (2025). DOI: 10.1302/0301-620X.107B2.BJJ-2024-0791.R1 · Original explanatory chart based on reported data
Evidence & Original Source Image use terms

The research question

Can AI Classify Four Types of Hip Fracture and Assist Clinicians in Reading Radiographs?

Participants and Methods

CNN classification and DAMO-YOLO preprocessing and augmentation were used. Development took place at two institutions, with external validation at an independent institution. Reading with and without AI assistance was also compared. 5,168 AP radiographs: 4,493 inside and 675 outside.

Main Findings

Development used 4,493 images from two institutions; external validation used 675 images from another institution. Table III presents external sensitivity values 0.83, 0.89, 0.88 and 0.85 in the average column for three doctors (A, B and C). The table footnote and abstract's description of model performance disagree, so these are not presented as standalone AI sensitivity.

Implications for clinical and surgical practice

The model assists diagnosis and fracture-type classification. It does not evaluate the effects of selecting particular operations.

Limitations

This is a retrospective image-classification study. Sensitivity and specificity for individual fracture types are different from overall accuracy, and the findings do not directly demonstrate improvements in patient recovery or surgical outcomes. The abstract/results descriptions conflict with Table III's reader footnote. The table number and the evaluator must be distinguished.

Let's put it briefly.

Which clinical question does this address?

The model assists diagnosis and fracture-type classification. It does not evaluate the effects of selecting particular operations.

Keywords

Hip Fractures
Deep Learning
Retrospective Studies
Sensitivity and Specificity
Neural Networks, Computer
Radiography
Aged, 80 and over
AI

Abstract · Original text

Zhiqian Zheng, Byeong Y Ryu, Sung E Kim, Dae S Song, Seong H Kim, Jung-Wee Park, Du H Ro. Deep learning for automated hip fracture detection and classification : achieving superior accuracy. The bone & joint journal. 2025;107-B(2):213-220. doi:10.1302/0301-620X.107B2.BJJ-2024-0791.R1.

AIMS

The aim of this study was to develop and evaluate a deep learning-based model for classification of hip fractures to enhance diagnostic accuracy.

METHODS

A retrospective study used 5,168 hip anteroposterior radiographs, with 4,493 radiographs from two institutes (internal dataset) for training and 675 radiographs from another institute for validation. A convolutional neural network (CNN)-based classification model was trained on four types of hip fractures (Displaced, Valgus-impacted, Stable, and Unstable), using DAMO-YOLO for data processing and augmentation. The model's accuracy, sensitivity, specificity, Intersection over Union (IoU), and Dice coefficient were evaluated. Orthopaedic surgeons' diagnoses served as the reference standard, with comparisons made before and after artificial intelligence assistance.

RESULTS

The accuracy, sensitivity, specificity, IoU, and Dice coefficients of the model for the four fracture categories in the internal dataset were as follows: Displaced (1.0, 0.79, 1.0, 0.70, 0.82), Valgus-impacted (1.0, 0.80, 1.0, 0.70, 0.82), Stable (0.99, 0.95, 0.99, 0.83, 0.89), and Unstable (1.0, 0.98, 0.99, 0.86, 0.92), respectively. For the external validation dataset, the sensitivity and specificity were as follows: Displaced (0.83, 0.94), Valgus-impacted (0.89, 0.90), Stable (0.88, 0.95), and Unstable (0.85, 0.99), respectively. The overall means (Micro AVG and Macro AVG) for the external dataset were Micro AVG (0.83 (SD 0.05), 0.96 (SD 0.01)) and Macro AVG (0.69 (SD 0.02), 0.95 (SD 0.02)), respectively.

CONCLUSION

Compared to human diagnosis alone, our study demonstrates that the developed model significantly improves the accuracy of detecting and classifying hip fractures. Our model has shown great potential in assisting clinicians with the accurate diagnosis and classification of hip fractures.

Read abstract on PubMed

The original paper.

Deep learning for automated hip fracture detection and classification: achieving superior accuracy

Zhiqian Zheng, Byeong Y Ryu, Sung E Kim, Dae S Song, Seong H Kim, Jung-Wee Park, Du H Ro
Bone & Joint Journal · 2025

DOI · 10.1302/0301-620X.107B2.BJJ-2024-0791.R1 PubMed Based data

Reviewed academic journal articles

Results explained with study numbers