Deep Learning based Gestational Age Estimation with Outlier Elimination in Blind Sweep Fetal Ultrasound

Deep learning based methods have shown promise in estimating gestational age (GA) from blind sweep ultrasound scans, but their performance can be compromised by outliers. We present a method to address this issue by incorporating an uncertainty measure, specifically the estimated variance of the GA output, into a deep learning based GA regression model. Our approach, trained on a large dataset of 8197 exams, filters out bad estimates or outliers (estimates with > 14 days error) while preserving accurate estimates. We evaluate our method on a test set of 1278 exams, demonstrating improved Mean Absolute Error (MAE) and reduced outliers after variance filtering. Our results suggest that variance filtering can be a safe and effective way to increase trust in deep learning based GA models and achieve safer clinical outcomes in low-resource settings.