Sampling and augmentation methods have been widely used to generate sufficiently large numbers of representative training images for supervised segmentation methods, such as convolutional neural networks (CNN). There is a need to understand the impact of sampling and augmentation methods on the resulting estimate of generalizable segmentation accuracy. This paper presents quantitative evaluations to determine the impact of sampling and augmentation on image segmentation accuracy estimation over very large image collections. The contribution lies in demonstrating a methodology for selecting sampling and augmentation parameters in addition to reporting estimated segmentation accuracies with high confidence. Based on quantitative evaluation of sample sizes between 10 and 100 from two large microscopic image collections of cells, we improved the segmentation accuracy estimate by 6% through automatically selecting augmentation models and generating 100 augmentations per image.
BioImage Computing Conference 2018
September 8-14, 2018
European Conference on Computer Vision
CNN, Image Segmentation, Augmentation, Segmentation Uncertainty