Development of Convolutional Neural Networks for Classification and Characterisation of Proximal Humerus Fractures on Computed Tomography.

BACKGROUND: Agreement between surgeons on classification, characterization and choice of treatment for proximal humerus fractures (PHFs) is poor, leading to subjective surgical decision-making and poor inter-surgeon reliability. Machine Learning for classification and characterization of PHFs on radiographs performed insufficiently. More detailed three-dimensional (3D) configuration of PHFs on Computed Tomography (CT) scans may improve performance. This study aimed to 1) develop and internally validate a Convolutional Neural Network (CNN) on CT-scans, 2) externally validate the model, and 3) compare characterization performance with orthopedic surgeons. METHODS: A 3D DenseNet was trained and internally validated on 581 Australian PHF patients with CT-scans, and externally validated geographically on 122 Dutch patients. Ground truth was established through multirater consensus. Fractures were classified as: a) none- to minimally displaced; b) two-part; c) multipart; or d) glenohumeral dislocation, and characterized on: a) greater tuberosity (GT) displacement ≥1 cm; b) varus malalignment (neck-shaft angle ≤100°); c) shaft translation; and d) articular involvement. All fractures were annotated with a bounding cube. The algorithm's performance was assessed with accuracy, area under the receiver operating characteristic curve, sensitivity, specificity, and negative and positive predictive values. RESULTS: Diagnostic accuracy for fracture classification was 78.6% (AUC's 0.87-0.99). The DenseNet could accurately characterize GT displacement (accuracy 80.3%, AUC 0.88), varus malalignment (accuracy 87.2%, AUC 0.91) but performance for subclasses of shaft translation (accuracy substantial translation 16.7%) and articular involvement (accuracy head-split 0%) was insufficient. The DenseNet was on par with orthopedic surgeons for GT displacement and varus angulation, and achieved higher overall accuracy than surgeons on shaft translation and articular involvement (p<0.001). Notably, algorithm performance in some subclasses (substantial and entire shaft translation, head-splits) was worse. CONCLUSION: This open-source CNN accurately classified and characterized PHFs for GT displacement and varus malalignment on CT-scans. AI performance for shaft translation and articular involvement was better than surgeons, but had insufficient areas under the curve. Performance of PHF classification and characterization of greater tuberosity displacement and varus angulation is strong and ready for prospective evaluation. For articular involvement and shaft translation, adding more data is needed to improve model performance. The code is available through https://github.com/Richardqiyi/classification-and-characterisation-of-PHFs. LEVEL OF EVIDENCE: Basic Science Study, Computer Modeling using Machine Learning.