Source-linked AI summary
Automatic Hip Fracture Identification and Functional Subclassification with Deep Learning
Justin D Krogue, Kaiyang V Cheng, Kevin M Hwang, Paul Toogood, Eric G Meinberg, Erik J Geiger, Musa Zaid, Kevin C McGill, Rina Patel, Jae Ho Sohn, Alexandra Wright, Bryan F Darger, Kevin A Padrez, Eugene Ozhinsky, Sharmila Majumdar, Valentina Pedoia
TL;DR
Hip fractures require accurate, timely radiographic diagnosis and classification because delays to surgery are associated with worse outcomes and fracture type informs operative planning. This study develops an automated deep-learning system for hip fracture diagnosis and subclassification, finding at least expert-level performance and improved resident interpretation when used as an aid.
Problem
Accurate and timely hip-fracture identification and classification are needed to reduce diagnostic errors and delays to operative intervention, while radiographic classification informs implant choice and surgical planning.
Method
The study labeled individual hips from hip and pelvic radiographs and used automated bounding-box placement with a 169-layer DenseNet and attention pooling to classify six hip categories.
Results
The model achieved at least expert-level binary and multiclass classification, outperformed residents under tested image conditions, and improved resident performance when used as an aid to approximate fellowship-trained experts.
Takeaways & Limitations
A fully automated deep-learning system may support hip-fracture diagnosis and functional subclassification while reducing diagnostic error, advanced imaging use, and delays to surgery.
Takeaways & Limitations
The model considers only a single image per prediction, unlike human interpreters who may use multiple views, including lateral images that can reveal subtle femoral neck fractures.
Abstract
from arXiv · showhide
Purpose: Hip fractures are a common cause of morbidity and mortality. Automatic identification and classification of hip fractures using deep learning may improve outcomes by reducing diagnostic errors and decreasing time to operation. Methods: Hip and pelvic radiographs from 1118 studies were reviewed and 3034 hips were labeled via bounding boxes and classified as normal, displaced femoral neck fracture, nondisplaced femoral neck fracture, intertrochanteric fracture, previous ORIF, or previous arthroplasty. A deep learning-based object detection model was trained to automate the placement of the bounding boxes. A Densely Connected Convolutional Neural Network (DenseNet) was trained on a subset of the bounding box images, and its performance evaluated on a held out test set and by comparison on a 100-image subset to two groups of human observers: fellowship-trained radiologists and orthopaedists, and senior residents in emergency medicine, radiology, and orthopaedics. Results: The binary accuracy for fracture of our model was 93.8% (95% CI, 91.3-95.8%), with sensitivity of 92.7% (95% CI, 88.7-95.6%), and specificity 95.0% (95% CI, 91.5-97.3%). Multiclass classification accuracy was 90.4% (95% CI, 87.4-92.9%). When compared to human observers, our model achieved at least expert-level classification under all conditions. Additionally, when the model was used as an aid, human performance improved, with aided resident performance approximating unaided fellowship-trained expert performance. Conclusions: Our deep learning model identified and classified hip fractures with at least expert-level accuracy, and when used as an aid improved human performance, with aided resident performance approximating that of unaided fellowship-trained attendings.
A PREPRINT
The preprint lists the study authors, including Justin D. Krogue and collaborators.
- Justin D. Krogue is listed as the first author.
- Kaiyang Cheng (Victor), Kevin M. Hwang, and additional collaborators are listed among the authors.
- Wright, Bryan F. Darger, Kevin A. Padrez, Eugene Ozhinsky, and Sharmila Majumdar are also listed.
1 Introduction
Hip fractures carry substantial morbidity and mortality, and timely radiographic classification matters because fracture type informs implant choice and operative planning. The study proposes an automated deep-learning system to identify and classify fractures and assess its effect on physician performance.
- Hip fractures cause substantial morbidity and mortality, with reported one-year mortality rates as high as 30%.
- Each 10-hour delay from admission to surgery is associated with a 5% higher odds of 1-year mortality.
- Radiographic fracture classification informs implant choice, while initial images may contain enough information to begin definitive surgical planning.
- The study proposes deep learning with a convolutional neural network for automated hip-fracture diagnosis and classification.
- The dataset included 3034 labeled hips from 2004 radiographs, divided into training, validation, and test sets.
- A random 100-image test subset was reserved for comparison with human observers.
2 Materials and Methods
The study retrospectively assembled and labeled hip radiographs, augmented hip images, trained an attention-enhanced DenseNet, and evaluated automated detection and classification against human observers.
- Data collection: The cohort comprised 919 report-identified likely fracture studies and 199 randomly selected studies from 1998–2017 emergency-room radiographs.
- Annotation: Reviewers annotated each hip with bounding boxes and classified it as unfractured, fractured, or containing hardware, with fractures subclassified by type.
- Data processing and augmentation: Training images were resized, laterality-normalized, duplicated at different box sizes, and augmented with contrast changes, cutout, Gaussian-mixture masking, and box wiggling.
- Model architecture: The classifier used a 169-layer DenseNet with attention pooling and a softmax output for each hip class.
- Object detection: The end-to-end system used a single-shot detector with a ResNet-50 feature pyramid to place hip bounding boxes automatically.
- Evaluation: Performance was evaluated using ROC/AUC, accuracy, sensitivity, specificity, confidence intervals, and comparison with fellowship-trained experts and PGY-4 residents.
3 Results
The deep learning system performed strongly in binary hip-fracture detection and multiclass classification, with automated hip localization closely matching manual bounding boxes. It also achieved expert-level performance relative to human observers and improved observer performance when used as an aid.
- Bounding box detection: The model correctly detected every labeled hip with mean intersection-over-union of 0.92, and automated boxes produced 93.4% binary and 90.4% multiclass accuracy.Performance on automatically generated boxes did not differ significantly from manually labeled boxes for either binary or multiclass classification.
- Comparison to human performance: The model achieved statistically superior performance to human observers on model-quality images in both binary and multiclass tasks.With full-quality images, superiority was generally observed but reached statistical significance over residents; experts performed better with full-quality than model-quality images.
- Comparison to human performance: Model assistance tended to improve all human observers, with statistically significant improvement among residents and aided resident performance not significantly different from unaided expert performance.Experts reached essentially equivalent performance to the model when using it as an aid, although residents remained outperformed by the model for binary classification.
4 Discussion
The study reports expert-level automated hip-fracture identification and functional subclassification, with improved resident performance when the model was used as an aid. Key limitations include low sensitivity for nondisplaced femoral neck fractures, single-image prediction, low-resolution inputs, manually generated boxes, and single-institution data.
- Performance: The model achieved at least expert-level binary and multiclass classification of hip radiographs across fractured and non-fractured groups.The study describes classification into six categories and reports this as the first deep-learning report of fracture subclassification.
- Limitations and future work: Model performance may improve with higher-resolution images because statistically significant expert gains occurred between lower- and full-quality images.The authors were restricted to low resolution by the small dataset and need for ImageNet pretraining.
- Human comparison: Aided residents approximated unaided fellowship-trained experts, while the model outperformed residents using both model-quality and full-resolution images.The authors suggest this could decrease diagnostic error and advanced-imaging use in emergency rooms.
- Limitations and future work: Generalizability is limited because all radiographs came from one institution, although images spanned 20 years and multiple scanners.The classification algorithm also depends on bounded-box images, prompting development of an automated detector for end-to-end processing.
- Limitations and future work: The model analyzes a single image rather than multiple views, whereas subtle femoral neck fractures may be best seen on lateral images.The authors identify multi-view prediction as a potential route to improved performance.
- Limitations and future work: Nondisplaced femoral neck fractures were the largest limitation, with 58% detected as any fracture and 46% correctly subclassified.Adjusting the detection threshold increased multiclass sensitivity to 77.0% with specificity of 87.4%.
- Clinical implications: The fully automated system could potentially reduce diagnostic error and advanced-imaging use while decreasing time to surgery.The authors connect these potential effects to improved recovery and reduced morbidity.