Augmentation for Generalization

Sammanfattning

To be able to automatically segment cells in microscopic images would give biologist a new tool to gather crucial data in quantities not possible by manual work. This is however not a trivial problem and has proven to be very difficult, especially if the images are in 3D. One of the major challenges is the ability for the methods to generalize beyond the data it has previously been presented with. This thesis investigates how image augmentation can be used to mitigate this issue in the domain of 3D microscopy. It does so by training two state of the art deep learning models, Plantseg and Cellpose, with different augmentations and then test their ability to generealize on three data sets which can be considered typical for the field. The results show that augmentations have a small but positive effect on the models. If the un-augmented model is completely unable to segment the image, augmentations will not improve the results. However, if the model is performing poorly, but is still able to segment some cells, augmentation can greatly improve the results. No augmentation by itself stood out as having a greater effect than others. Instead the combination of all the augmentations gave the best results over all the experiments. Furthermore, the ability to generalize was strongly correlated with the difference in shape and size of the different data sets. Further research into the shape and size augmentations are hence encouraged. Implementation for the experiments can be found on Github here or the full link in the foot note 1.