Deep learning-based cell segmentation is increasingly utilized in cell biology due to the massive accumulation of large-scale datasets and excellent progress in model architecture and instance representation. However, the development of specialist algorithms has long been hampered by a paucity of annotated training data, whereas the performance of generalist algorithms is limited without experiment-specific calibration. Here, we present Scellseg, an adaptive pipeline that utilizes a style-aware pre-trained model coupled to a contrastive fine-tuning strategy that also learns from unlabeled data. Scellseg achieves state-of-the-art transferability in average precision and Aggregated Jaccard Index on disparate datasets containing microscopy images at three biological levels, from organelle, cell to organism. Interestingly, when fine-tuning Scellseg, we show that performance plateaued after approximately eight images, implying that a specialist model can be obtained with few manual efforts. For convenient dissemination, we develop a graphical user interface that allows biologists to easily specialize their self-adaptive segmentation model.
Keywords: Artificial intelligence; Bioinformatics; Cell biology.
© 2022 The Author(s).