Weak supervision using cell tracking annotation and image registration improves cell segmentation

Nadezhda A. Anoshina, Dmitry V. Sorokin

Abstract: Learning-based cell segmentation methods have proved to be very effective in cell tracking. The main difficulty of using machine learning is the lack of expert annotation of biomedical data. We propose a weakly-supervised approach that extends the amount of segmentation training data for image sequences where only a couple of frames are annotated. The approach uses the tracking annotations as weak labels and image registration to extend the segmentation annotation to the neighbouring frames. This technique was applied to cell segmentation step in the cell tracking problem. An experimental comparison of the baseline segmentation network trained on the data with pure GT annotation and the same segmentation network trained on the GT data and additional annotations generated with the proposed approach has been performed. The proposed weakly-supervised approach increased the IoU and SEG metrics on the data from the Cell Tracking Challenge.