Self-supervised pretraining for transferable quantitative phase image cell segmentation

• DOI: 10.1364/boe.433212
• Publication Date: 2021-09-06T10:30:06Z
• Authors: Tomas Vicar; Jiri Chmelik; Roman Jakubicek; Larisa Chmelikova; Jaromir Gumulec; Jan Balvan; Ivo Provaznik; Radim Kolar
• Keywords: Transfer of learning

In this paper, a novel U-Net-based method for robust adherent cell segmentation quantitative phase microscopy image is designed and optimised. We evaluated four specific post-processing pipelines. To increase the transferability to different types, non-deep learning transfer with adjustable parameters used in step. Additionally, we proposed self-supervised pretraining technique using nonlabelled data, which trained reconstruct multiple distortions improved performance from 0.67 0.70 of object-wise intersection over union. Moreover, publish new dataset manually labelled images suitable task together unlabelled data pretraining.