ABSTRACT
With the rapidly increasing interest in machine learning based solutions for automatic image annotation, the availability of reference annotations for algorithm training is one of the major bottlenecks in the field. Crowdsourcing has evolved as a valuable option for low-cost and large-scale data annotation; however, quality control remains a major issue which needs to be addressed. To our knowledge, we are the first to analyze the annotation process to improve crowd-sourced image segmentation. Our method involves training a regressor to estimate the quality of a segmentation from the annotator's clickstream data. The quality estimation can be used to identify spam and weight individual annotations by their (estimated) quality when merging multiple segmentations of one image. Using a total of 29,000 crowd annotations performed on publicly available data of different object classes, we show that (1) our method is highly accurate in estimating the segmentation quality based on clickstream data, (2) outperforms state-of-the-art methods for merging multiple annotations. As the regressor does not need to be trained on the object class that it is applied to it can be regarded as a low-cost option for quality control and confidence analysis in the context of crowd-based image annotation.
ABSTRACT
PURPOSE: With the recent trend toward big data analysis, neuroimaging datasets have grown substantially in the past years. While larger datasets potentially offer important insights for medical research, one major bottleneck is the requirement for resources of medical experts needed to validate automatic processing results. To address this issue, the goal of this paper was to assess whether anonymous nonexperts from an online community can perform quality control of MR-based cortical surface delineations derived by an automatic algorithm. METHODS: So-called knowledge workers from an online crowdsourcing platform were asked to annotate errors in automatic cortical surface delineations on 100 central, coronal slices of MR images. RESULTS: On average, annotations for 100 images were obtained in less than an hour. When using expert annotations as reference, the crowd on average achieves a sensitivity of 82 % and a precision of 42 %. Merging multiple annotations per image significantly improves the sensitivity of the crowd (up to 95 %), but leads to a decrease in precision (as low as 22 %). CONCLUSION: Our experiments show that the detection of errors in automatic cortical surface delineations generated by anonymous untrained workers is feasible. Future work will focus on increasing the sensitivity of our method further, such that the error detection tasks can be handled exclusively by the crowd and expert resources can be focused on error correction.
