Few Shot Learning Framework to Reduce Inter-observer Variability in Medical Images

Most computer aided pathology detection systems rely on large volumes of\nquality annotated data to aid diagnostics and follow up procedures. However,\nquality assuring large volumes of annotated medical image data can be\nsubjective and expensive. In this work we present a novel standardization\nframework that implements three few-shot learning (FSL) models that can be\niteratively trained by atmost 5 images per 3D stack to generate multiple\nregional proposals (RPs) per test image. These FSL models include a novel\nparallel echo state network (ParESN) framework and an augmented U-net model.\nAdditionally, we propose a novel target label selection algorithm (TLSA) that\nmeasures relative agreeability between RPs and the manually annotated target\nlabels to detect the "best" quality annotation per image. Using the FSL models,\nour system achieves 0.28-0.64 Dice coefficient across vendor image stacks for\nintra-retinal cyst segmentation. Additionally, the TLSA is capable of\nautomatically classifying high quality target labels from their noisy\ncounterparts for 60-97% of the images while ensuring manual supervision on\nremaining images. Also, the proposed framework with ParESN model minimizes\nmanual annotation checking to 12-28% of the total number of images. The TLSA\nmetrics further provide confidence scores for the automated annotation quality\nassurance. Thus, the proposed framework is flexible to extensions for quality\nimage annotation curation of other image stacks as well.\n