Efficient End-to-end Learning for Cell Segmentation with Machine Generated Incomplete Annotations

Automated cell segmentation from optical microscopy images is usually the first step in the pipeline of single-cell analysis. Recently, deep-learning based algorithms have shown superior performances for the cell segmentation tasks. However, a disadvantage of deep-learning is the requirement for a large amount of fully-annotated training data, which is costly to generate. Weakly-supervised and self-supervised learning is an active research area, but often the model accuracy is inversely correlated with the amount of annotation information provided. Here we focus on a specific subtype of incomplete annotations, which can be generated programmably from experimental data, thus allowing for more annotation information content without sacrificing the annotation speed. We designed a new model architecture for end-to-end training using such incomplete annotations. We benchmarked our method on a variety of publicly available dataset, covering both fluorescence and bright-field imaging modality. We additionally tested our method on a microscopy dataset generated by us, using machine generated annotations. The results demonstrated that our model trained under weak-supervision can achieve segmentation accuracy competitive to, and in some cases surpassing, state-of-the-art models trained under full supervision. Therefore, our method can be a practical alternative to the established full-supervision methods.