Silver Standard Masks for Data Augmentation Applied to\n Deep-Learning-Based Skull-Stripping

Abstract

The bottleneck of convolutional neural networks (CNN) for medical imaging is\nthe number of annotated data required for training. Manual segmentation is\nconsidered to be the "gold-standard". However, medical imaging datasets with\nexpert manual segmentation are scarce as this step is time-consuming and\nexpensive. We propose in this work the use of what we refer to as silver\nstandard masks for data augmentation in deep-learning-based skull-stripping\nalso known as brain extraction. We generated the silver standard masks using\nthe consensus algorithm Simultaneous Truth and Performance Level Estimation\n(STAPLE). We evaluated CNN models generated by the silver and gold standard\nmasks. Then, we validated the silver standard masks for CNNs training in one\ndataset, and showed its generalization to two other datasets. Our results\nindicated that models generated with silver standard masks are comparable to\nmodels generated with gold standard masks and have better generalizability.\nMoreover, our results also indicate that silver standard masks could be used to\naugment the input dataset at training stage, reducing the need for manual\nsegmentation at this step.\n

Ähnliche Arbeiten