Keywords: Deep Learning, Uncertainty, Quality Control, Radiology, Proximal femur fractures
TL;DR: Uncertainty score using Monte Carlo dropout layers as a quality control for automated proximal femur fracture classification using the AO-fracture classification system.
Abstract: Deep Learning methods over the past years provided high-performance solutions for the medical applications. Yet, robustness and quality control is still required for clinical applicability. In this work, the uncertainty of proximal femur fracture classification,was modeled. We introduce a reliability measure to our predictive model using the Monte Carlo Dropout approach. We performed an extensive quantitative and qualitative analysis to validate the results. We further exposed the results to expert physicians in order to get feedback on the model’s performance and uncertainty measures. Results demonstrate a positive correlation between the miss-classification of the model’s prediction and high uncertainty scores. Additionally, the uncertainty measures are mimicking the actual radiologist’s uncertainty for challenging examples reflected on intra- and inter- experts variability.