Supervisors: Alvaro Gomariz, Firat Ozdemir, Prof. Orcun Goksel
Current most performant deep learning techniques for image segmentation rely on big datasets of annotated images. However, manual annotation requires a fastidious and long work from biomedical experts, especially when there are a lot of instances in images. In this thesis, we study both the quantification of uncertainty in microscopy images and the consistency of an active learning pipeline that would suggest the samples to be annotated in priority (using the uncertainty quantification methods mentioned above and similarity) in order to reach an optimal quality of segmentation with minimum annotation effort. The main challenge is to estimate the uncertainty of the model for deep learning methods not initially tailored for this. We test the active learning pipeline with three uncertainty estimation techniques: bootstrapping, random transformations and Monte Carlo dropout. This work has visually compared different approaches for uncertainty estimation and quantitatively evaluated them for the problem of active learning using Data Science Bowl 2018 dataset.