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

TL;DR

This paper introduces silver standard masks generated via consensus algorithms to augment training data for CNN-based skull-stripping, reducing reliance on manual annotations and improving model generalization across datasets.

Contribution

The study proposes a novel use of silver standard masks for data augmentation in medical image segmentation, demonstrating comparable performance to gold standards and enhanced generalizability.

Findings

Silver standard masks achieve similar accuracy to gold standards.

Models trained with silver masks generalize better across datasets.

Using silver masks reduces the need for manual annotations.

Abstract

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