Deep learning methods are promising avenues for addressing imaging analysis challenges in neuroscience. However, the training of supervised deep learning models requires vast amounts of labeled training data. This can be problematic when the acquisition and labeling of data are costly. For example, for a segmentation task in neurophotonics, an expert must label complex structures on microscopy images of expensive biological samples. Active learning (AL) aims to alleviate this issue by reducing the required amount of data to label by selecting the most informative samples. However, many AL approaches assume a uniform labeling cost for all data points, which may result in selecting samples that would be costly to acquire in practice.

In this work, we tackle the AL problem with non-uniform sample costs.