Deep Learning for Wrist X-ray Analysis: Automatic Classification and Radius Segmentation Using ResNet50 and U-Net
Introduction
Fractures of the distal radius are among the most common orthopedic injuries, particularly in two high-risk populations: children under 16 years and postmenopausal women between 60 and 70 years of age. According to national health data, over 175,000 cases of distal radius fractures were treated in Korea in 2018 alone, reflecting the growing burden of wrist injuries in aging societies.
While Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) can provide detailed information on fracture patterns and associated soft tissue injuries, plain radiographs (anteroposterior and lateral views) remain the first-line diagnostic tool due to their accessibility, cost-effectiveness, and speed. However, interpretation accuracy varies widely depending on the physician’s experience. Studies have shown that junior emergency physicians achieve only around 72% diagnostic accuracy when interpreting wrist radiographs.
To address this gap, recent advances in Artificial Intelligence (AI) and deep learning have opened new opportunities for computer-aided diagnosis (CAD) in musculoskeletal imaging. Deep convolutional neural networks (CNNs) and segmentation models such as ResNet50 and U-Net offer powerful tools for automated classification and segmentation of X-ray images, potentially improving diagnostic precision and reducing inter-observer variability.
This column explores a landmark study that applied ResNet50 for wrist X-ray classification (anteroposterior vs. lateral views) and U-Net for radius segmentation, achieving near-perfect performance metrics and demonstrating the clinical potential of AI in orthopedic imaging.
Methods and Model Design
Data Collection
-
Total dataset: 904 wrist X-rays of patients with distal radius fractures.
-
Distribution:
-
472 anteroposterior (AP) images
-
432 lateral (LAT) images
-
-
Annotation: Expert radiologists manually segmented the radius using ImageJ to generate ground truth masks (Fig. 1).
Figure 1. Annotation process for radius segmentation
a) Original X-ray image
b) Ground truth mask created by radiologists
c) Overlay of ground truth mask on the original image
Preprocessing
-
Images were resized to 1024 × 1024 pixels.
-
Zero-padding was applied to equalize dimensions.
-
No data augmentation was used in the initial study.
Classification Model: ResNet50
The ResNet50 CNN was used to classify radiographs into AP or LAT categories (Fig. 2).
Figure 2. ResNet50 architecture showing residual blocks and shortcut connections
|
Segmentation Model: U-Net
The U-Net architecture (Fig. 3) was applied to segment the radius bone in AP and LAT radiographs.
Figure 3. U-Net architecture used for medical image segmentation
|
Results
Classification Performance
The ResNet50 model achieved perfect classification performance for AP vs. LAT images (Table 1).
Table 1. Classification results (AP vs. LAT)
Precision: 100%, Recall: 100%, F1-score: 100%, AUC: 1.0
This indicates that the model correctly identified all radiographs without error.
Segmentation Performance
The U-Net model delivered excellent accuracy in radius segmentation (Table 2, Fig. 4–5).
Figure 4. Radius segmentation in AP images
a) Original image, b) Ground truth, c) U-Net segmentation, d) Overlay
Figure 5. Radius segmentation in LAT images
a) Original image, b) Ground truth, c) U-Net segmentation, d) Overlay |
Table 2. Segmentation results (AP vs. LAT)
| Metric | AP (%) | LAT (%) |
|---|---|---|
| Accuracy | 99.46 | 99.37 |
| Sensitivity | 89.68 | 88.65 |
| Specificity | 99.72 | 99.69 |
| Dice Score | 90.05 | 86.05 |
Despite near-perfect accuracy, performance was slightly lower in LAT images due to the overlap of the radius and ulna.
Bland-Altman Analysis
To validate agreement between manual and AI-based segmentation, Bland-Altman plots were generated (Fig. 6).
Figure 6. Bland-Altman plot comparing manual vs. automated segmentation areas
Results confirmed that >90% of values fell within the 95% confidence interval, proving high reliability.
Discussion
The findings highlight the transformative role of deep learning in musculoskeletal radiology:
-
Clinical Impact
-
Accurate classification of AP vs. LAT images ensures correct orientation for diagnostic assessment.
-
Automated segmentation provides objective measurements of radial inclination and palmar tilt, critical for fracture evaluation.
-
-
Performance Insights
-
AP images: superior performance due to clearer bone boundaries.
-
LAT images: slightly lower Dice score due to bone overlap and complex fracture patterns.
-
-
Limitations
-
The dataset size (904 images) was modest for deep learning.
-
Only AP and LAT views analyzed; oblique views not included.
-
Severe comminuted fractures posed segmentation challenges.
-
-
Future Directions
-
Incorporating data augmentation and larger multicenter datasets.
-
Extending segmentation to oblique radiographs.
-
Integrating with clinical decision support systems (CDSS) for real-time fracture diagnosis.
-
Clinical Applications
-
Emergency departments: Rapid triage of wrist injuries.
-
Orthopedic clinics: Standardized fracture assessment.
-
AI-powered PACS systems: Seamless integration into radiology workflows.
Ultimately, AI-driven models can assist less experienced clinicians, reduce diagnostic errors, and enhance patient outcomes.
Quiz
1 Which deep learning model was used for the classification of AP vs. LAT wrist X-rays?
-
VGG16
-
ResNet50
-
U-Net
-
AlexNet
2. Why did the segmentation performance of U-Net drop slightly in lateral (LAT) images compared to AP images?
-
LAT images had lower resolution.
-
The overlap between the radius and ulna complicates segmentation.
-
U-Net is not suitable for lateral X-rays.
-
Ground truth annotations were missing.
Answer & Explanation
1. Answer: 2. ResNet50. Explanation: The study employed ResNet50 for classification due to its residual learning blocks, enabling deeper network training without vanishing gradients.
2. Answer: 2. Overlap between radius and ulna complicates segmentation. Explanation: In LAT views, bones overlap, making boundary detection difficult for both human experts and AI models.
Conclusion
This study demonstrates that deep learning models (ResNet50 + U-Net) can achieve state-of-the-art performance in wrist radiograph analysis, accurately classifying image orientations and segmenting the radius with high precision.
Such AI-driven approaches represent the future of computer-aided fracture diagnosis, offering scalable solutions to address variability in clinical expertise, improve diagnostic confidence, and support orthopedic decision-making worldwide.
References
[1] C. Brudvik and L. M. Hove, "Childhood fractures in Bergen, Norway: identifying high-risk groups and activities," J. Pediatr. Orthop., vol. 23, pp. 629–634, 2003.
[2] R. A. Owen, L. J. Melton, K. A. Johnson, D. M. Ilstrup, and B. L. Riggs, "Incidence of Colles' fracture in a North American community," Am. J. Public Health, vol. 72, pp. 605–607, 1982.
[3] M. J. Kiuru, V. V. Haapamaki, M. P. Koivikko, and S. K. Koskinen, "Wrist injuries: diagnosis with multidetector CT," Emerg. Radiol., vol. 10, no. 4, pp. 182–185, 2004.
[4] S. U. Cho, J. I. Yang, K. H. Han, et al., "The accuracy of a simple radiologic study for diagnosing intra-articular fractures of the distal radius," J. Korean Soc. Emerg. Med., vol. 21, no. 5, pp. 569–574, 2010.
[5] J. G. Lee, S. Jun, Y. W. Cho, H. Lee, G. B. Kim, J. B. Seo, and N. Kim, "Deep learning in medical imaging: general overview," Korean J. Radiol., vol. 18, no. 4, pp. 570–584, 2017.
[6] O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional networks for biomedical image segmentation," in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Interv., 2015, pp. 234–241.
[7] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 770–778.
댓글
댓글 쓰기