Related Experiment Video
Updated: Aug 6, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Development of Convolutional Neural Networks for Classification and Characterisation of Proximal Humerus Fractures on
Stijn R J Mennes1, Reinier W A Spek2, Xingyuan Zhang3
1Department of Orthopaedics and Trauma Surgery, Flinders Medical Centre and Flinders University, Adelaide, Australia; Shoulder and elbow unit, Department of Orthopaedic Surgery, OLVG, Amsterdam, the Netherlands; Faculty of Behavioural and Movement Sciences, Department of Human Movement Sciences, Vrije Universiteit Amsterdam, Amsterdam Movement Sciences, Amsterdam, the Netherlands.
This study developed a Convolutional Neural Network (CNN) to classify and characterize proximal humerus fractures (PHFs) using CT scans. The AI demonstrated strong performance in identifying greater tuberosity displacement and varus malalignment, outperforming surgeons in some aspects.
Area of Science:
- Orthopedic Surgery
- Medical Imaging
- Artificial Intelligence
Background:
- Surgeon agreement on proximal humerus fracture (PHF) classification and treatment is poor, leading to subjective decision-making.
- Machine learning (ML) on radiographs has limitations for PHF classification and characterization.
- Three-dimensional (3D) CT scans offer potential for improved ML performance in analyzing complex PHF configurations.
Purpose of the Study:
- To develop and validate a Convolutional Neural Network (CNN) for PHF classification and characterization using 3D CT scans.
- To externally validate the developed AI model across different geographical cohorts.
- To compare the AI's characterization performance against orthopedic surgeons.
Main Methods:
- A 3D DenseNet model was trained and validated on 581 Australian PHF CT scans, with external validation on 122 Dutch cases.
- Fractures were classified and characterized for greater tuberosity displacement, varus malalignment, shaft translation, and articular involvement.
- Performance was evaluated using accuracy, AUC, sensitivity, specificity, and predictive values.
Main Results:
- The CNN achieved 78.6% accuracy in fracture classification (AUCs 0.87-0.99).
- Accurate characterization was achieved for greater tuberosity displacement (80.3% accuracy, AUC 0.88) and varus malalignment (87.2% accuracy, AUC 0.91).
- AI performance was superior to surgeons for shaft translation and articular involvement, though insufficient for specific subclasses.
Conclusions:
- An open-source CNN accurately classifies and characterizes PHFs on CT scans, particularly for GT displacement and varus malalignment.
- AI shows promise in improving objectivity and reliability in PHF assessment.
- Further data is required to enhance AI performance for articular involvement and shaft translation, with prospective evaluation planned for validated aspects.
