Cell segmentation without annotation by unsupervised domain adaptation based on cooperative self-learning

Abstract

Semantic cell segmentation from microscopic images is essential for the quantitative evaluation of cell morphology. Although supervised deep-learning-based models offer accurate segmentation, their performance degrades for unknown cell types. To address this problem, unsupervised domain adaptation methods based on adversarial training, self-training, or a combination of these approaches have been developed in recent years. These methods train the model using inference labels from the unknown domain as pseudo labels with reliability to resolve the discrepancy between the features of the unknown and known domains. However, conventional methods require a predefined threshold to calculate pseudo-labels reliability, leading to costly hyperparameter tuning. Here, we developed an unsupervised domain adaptation for semantic cell segmentation with cooperative self-learning (CULPICO: Cooperative Unsupervised Learning for PIxel-wise COloring) that does not require predefined threshold of the pseudo-labels reliability. The proposed method consists of two independent segmentation models and a mutual exchange mechanism of inference data. The models infer a label probability at each pixel and generate a pseudo-label as unsupervised learning. The pseudo-labels created by each model are mutually used as ground truth in the other model. Loss function is corrected by considering pixel-level discrepancies between the label probabilities inferred by the two models. The proposed method, despite being an unsupervised learning method, can segment efficiently the unknown cell types without labels with an accuracy comparable to supervised learning models. Our method, which could solve the performance degradation problem without constructing new datasets, is expected to accelerate life science by reducing the cost of extracting quantitative biological knowledge.