Table 2 Comparison of DeLTA and Omnipose trained on 3 types of synthetic training data: data containing large cells only, data containing small cells only, and the combination of these two cell types to produce a combined dataset. As expected, specialist models tend to perform better when segmenting their own data type, and worse when segmenting unseen data types. A model trained on a combination of the two datasets provides a good compromise, not needing to train multiple models, while retaining good performance across the entire dataset. Interestingly, the Omnipose model trained on combination data performed as well on exponentially growing cells as the same model trained only on large cell synthetic data, implying that the network has residual capacity to learn. This was also noted in the original Cellpose paper [32]
