Image Heterogeneity Modelling

Summary

Image heterogeneity does not have a mathematical definition but it is used in describing image regions. We developed a model to automatically and accurately characterize stem cell colonies into three categories. The colony categories are described by cell biologists as heterogeneous, homogeneous or dark, and shown in Figure 1. Three interactive visualisation tools help understand the complex nature of stem cells behavior and evolution.

Description

The stem cell colonies exhibit various texture appearance, such as Homogeneous, Heterogeneous, Dark or a mix of those.

A predictive model has been designed, to automatically assess and characterize the texture of each colony based on expert's inputs. The model works at a multi-resolution level. First, it analyzes local properties of the colonies, and then aggregates them at the colony level. Despite a small number of labels provided by experts, the prediction accuracy is over 97% over a five-fold cross-validation evaluation.

The prediction results are accessible in three browser friendly interactive visualizations accesible via URL. Screen captures of the visualizations are shown below.

Figure 2: The colony predictions' map (interact by hovering mouse and dragging items)

• Clusters colonies with similar texture together
• Offer gradual and smooth transition between the 3 types of extreme colonies
• A full resolution image of each colony is displayed when hovering mouse
• The predictions output are displayed in text and illustrated
• The layout minimizes information overlap on the “prediction" map”

Figure 3: The augmented lineage tree interacts by zooming, panning and hovering

• Displays the texture predictions for every colony & time frame
• Shows a lineage tree to follow colony interactions & temporal evolution

Figure 4: The feature selection optimization process interacts by hovering mouse

• Easily assesses classification accuracy for different feature subset
• Shows accuracies by using a color-coded heat map