Principled Ultrasound Data Augmentation for Classification of Standard Planes

Deep learning models with large learning capacities often overfit to medical\nimaging datasets. This is because training sets are often relatively small due\nto the significant time and financial costs incurred in medical data\nacquisition and labelling. Data augmentation is therefore often used to expand\nthe availability of training data and to increase generalization. However,\naugmentation strategies are often chosen on an ad-hoc basis without\njustification. In this paper, we present an augmentation policy search method\nwith the goal of improving model classification performance. We include in the\naugmentation policy search additional transformations that are often used in\nmedical image analysis and evaluate their performance. In addition, we extend\nthe augmentation policy search to include non-linear mixed-example data\naugmentation strategies. Using these learned policies, we show that principled\ndata augmentation for medical image model training can lead to significant\nimprovements in ultrasound standard plane detection, with an an average\nF1-score improvement of 7.0% overall over naive data augmentation strategies in\nultrasound fetal standard plane classification. We find that the learned\nrepresentations of ultrasound images are better clustered and defined with\noptimized data augmentation.\n