Uncertainty-Based Adaptive Data Augmentation For Ultrasound Imaging Anatomical Variations
Edward Chen, Howie Choset, John Galeotti · 2021
Data augmentation remains to be a simple and inexpensive method for generalizing across unseen domains. Current data augmentation methods for ultrasound imaging involve simple image transformations - rotations, flips, skews, and blurs - but are not able to adapt to the current state of the deep learning model. We present the first online adaptive data augmentation method that is able to generate synthetic training data on-the-fly, enabling the model to adapt to countless spatial deformations. Our proposed method leverages prior work on uncertainty quantification to understand the model's weaknesses at any given stage. Our method is also able to then ”spot-augment” subsets of regions within the ultrasound image, all in a real-time manner. We also show that our proposed method is able to perform significantly better in out-of-true-training distributions, when compared against models trained on the same dataset.