Machine Learning Model for Diagnosis of Non-Displaced Fractures On Emergency Department Hip Radiographs
K Drukker*, A Siegel, B Anderson, N Sundaram, L Lan, M Giger, University of Chicago, Chicago, IL
Presentations
PO-GePV-M-51 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Non-displaced hip fractures are easily overlooked in radiographs acquired in the emergency department setting and machine learning methods may be able to triage radiographs for further scrutiny.
Methods: The dataset contained hip radiographs of 130 patients. One radiograph was analyzed per patient and a region-of-interest (ROI) of variable size including the proximal femur (femoral head, femoral neck as well as the intertrochanteric region) was manually identified by a radiology resident with 3 years of experience. The dataset was split into a training/validation set (100 cases: 18 displaced fractures, 69 non-displaced fractures, 13 no-fracture cases) and an independent test set (30 cases: 20 non-displaced fractures and 10 no-fracture cases). We fine-tuned the last 3 layers of a pretrained GoogLeNet using the training/validation set with data augmentation for the task of classifying radiographs for the presence of a fracture in three separate analyses: one using the radiograph ROIs as input and two using image texture feature ROIs, standard deviation and entropy, as input. Performance was assessed on the test set using receiver operating characteristic (ROC) and precision-recall analyses.
Results: The areas under the ROC curve in the task of classifying radiographs in the independent test set for the presence of a non-displaced fracture were 0.94 [0.78; 0.99], 0.92 [0.66; 1.0], and 0.91 [0.74; 0.98], respectively for the original, standard deviation, and entropy ROIs. The corresponding areas under the precision-recall curves were 0.92 [0.83; 0.96], 0.89 [0.70; 0.95], and 0.90 [0.82; 0.96]. We failed to find a benefit from combining the analyses of different image types through soft-voting.
Conclusion: This pilot study demonstrated promising performance in the classification of hip radiographs for the presence of a non-displaced hip fracture. Future studies will be performed with a larger dataset and employ automated ROI selection.
Funding Support, Disclosures, and Conflict of Interest: Supported by the National Center for Advancing Translational Sciences (NCATS) of the National Institutes of Health (NIH) through Grant Number 5UL1TR002389-04 that funds the Institute for Translational Medicine (ITM). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.
Keywords
Taxonomy
Contact Email