ACTIS: improving data efficiency by leveraging semi-supervised augmentation consistency training for instance segmentation

Segmenting objects like cells or nuclei in biomedical microscopy data is a standard task required for many downstream analyses. However, existing pre-trained models are continuously challenged by ever-evolving experimental setups and imaging platforms. On the other hand, training new models still requires a considerable number of annotated samples, rendering it infeasible for small to midsized experiments. To address this challenge, we propose a semi-supervised learning approach for instance segmentation that leverages a small number of annotated samples together with a larger number of unannotated samples. Our pipeline, Augmentation Consistency Training for Instance Segmentation (ACTIS), incorporates methods from consistency regularization and entropy minimization. In addition, we introduce a robust confidence-based loss masking scheme which we find empirically to work well on highly imbalanced class frequencies. We show that our model can surpass the performance of supervised models trained on more than twice as much annotated data. It achieves state-of-the-art results on three benchmark datasets in the biomedical domain, demonstrating its effectiveness for semi-supervised instance segmentation.