Automated image analysis, guided by AI or deep learning algorithms, must first obtain a manually classified data set to train and validate the machine learning models. These data sets are often painfully laborious to create and existing tools fall short of supporting scientific images that can be complex, noisy, and difficult to interpret. The Visual Analytics group discovered this the hard way by getting involved in an image analysis project that required classification of complex biological images as a step in the process. As a result, we decided to build our own tool to make this process as easy and efficient as possible. PixSure is still under development but we are actively seeking use cases to help us define a broad set of requirements useful to many scientific domains.

​

​

Primary reasons to build PixSure:

​

​

• Complex scientific images

• Supports multi-spectral data with multiple RGB views

• User workflows for large images and large image sets

  • Guided workflow/pipeline (“initiator” for getting datasets transformed to ML training data)

• Many ground truth classes with preferred views for each (lens depending on what type of cells)

• Users trace silhouettes not bounding boxes for scientific phenomenon like clouds and blood cells.

• Emphasis on quickly labeling new image sets in response to improvements in imaging technologies. (Instead of focus on building large archives over long time frames)

  • Being able to switch/respond to changes in technology (microscopes)

• Module for being able to keep up to date with quickly moving field of ML