Domain Knowledge Adapted Semi-supervised Learning with Mean-Teacher Strategy for Circulating Abnormal Cells Identification

The number of signals in each signal channel in the nucleus of a fourcolor FISH image is the basis for distinguishing between normal cells, deletion signal cells, and CACs. In previous studies, we adopted deep learning for signal detection, which required a relatively large number of voxel-level labeled cells and signal images for training. In this study, we introduce a mean teacher mechanism into the training process and propose an end-to-end semi-supervised object detection method to detect the signal. We also propose Domain-Adaptive Pseudo Labels as a false positive filtering based on the prior knowledge of CAC signal. The experimental results show that the strategies proposed is simple and effective. On the four-color FISH image, when using only 8% of labeled data is used, it can all achieve 0.15% 0.41% 0.55% and 0.85% F1 score improvements compared to the supervised baseline.