Evaluation of ensemble strategy on the development of multiple view ankle fracture detection algorithm

Severance
DOI: 10.1259/bjr.20220924 Publication Date: 2023-03-17T17:52:58Z
ABSTRACT
To identify the feasibility and efficiency of deep convolutional neural networks (DCNNs) in detection ankle fractures to explore ensemble strategies that applied multiple projections radiographs.Ankle radiographs (AXRs) are primary tool used diagnose fractures. Applying DCNN algorithms on AXRs can potentially improve diagnostic accuracy detecting fractures.A was trained using a trauma image registry, including 3102 AXRs. We separately anteroposterior (AP) lateral (Lat) Different methods, such as "sum-up," "severance-OR," "severance-Both," were evaluated incorporate results model different view.The AP/Lat model's individual sensitivity, specificity, positive-predictive value, accuracy, F1 score 79%/84%, 90%/86%, 88%/86%, 83%/85%, 0.816/0.850, respectively. Furthermore, area under receiver operating characteristic curve (AUROC) 0.890/0.894 (95% CI: 0.826-0.954/0.831-0.953). The sum-up method generated balanced by applying both models obtained an AUROC 0.917 0.863-0.972) with 87% accuracy. severance-OR resulted better sensitivity 90%, severance-Both high specificity 94%.Ankle fracture AXR could be identified algorithm. selection methods depend clinical situation which might help clinicians detect efficiently without interrupting current pathway.This study demonstrated AI view optimize performance various needs.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES (35)
CITATIONS (7)