Background: Segmentation of electron microscopic continuous section images by deep learning has attracted attention as a technique to reduce the cost of annotation for researchers attempting to make observations using 3D reconstruction methods. However, when the observed samples are rare, or scanning circumstances are unstable, pursuing generalization performance for newly obtained samples is not appropriate.
New methods: We assume a transductive setting that predicts all labels in a dataset from only partially obtained labels while avoiding the pursuit of generalization performance for unknown data. Then, we propose sequential semi-supervised segmentation (4S), which semi-automatically extracts neural regions from electron microscopy image stacks. This method focuses on the fact that adjacent images have a strong correlation in serial images. Our 4S repeats training, inference, and pseudo-labeling using a minimal number of teacher labels and performs segmentation on all slices.
Result: Our experiments using two types of serial section images showed effectiveness in terms of both quality and quantity. In addition, we experimentally clarified the effect of the number and position of teacher labels on performance.
Comparison with existing methods: Compared with supervised learning when a small number of labeled data was obtained, the performance of the proposed method was shown to be superior.
Conclusion: Our 4S leverages a limited number of labeled data and a large amount of unlabeled data to extract neural regions from serial image stacks in a transductive setting. We plan to develop this method as a core module of a general-purpose annotation tool in our future work.
Keywords: Cell segmentation; Deep neural networks; Pseudo-labeling; Semi-supervised learning.
Copyright © 2021 The Author(s). Published by Elsevier B.V. All rights reserved.