Our Project

Predicting cell evolution and learning their growth dynamics with noisy annotations and partial-blurry images

The development of plant cells over time is very important. If we can characterize how a plant cell will grow over time, we can either:

• ignore the plant cells that we expect to die quickly and focus on the plant cells that we expect to divide and grow into new plants,

• perform interventions to transform cells that we expect to die into cells that we expect to divide and grow into new plants.

To characterize how a plant cell will grow over time and to investigate how we can perform interventions to influence the development, one could conduct experiments. For instance, we can apply pressure to cells and monitor which cells divide under which force. Based on the observations, we can formulate a mathematical relationship between cell division and the applied pressure. In turn, this mathematical relationship can be used to predict how a given cell will react to a particular applied pressure. 

One big disadvantage of such an approach is that for humans it is very hard to, for instance, design the mathematical relationship from the experiments. To circumvent the disadvantage, we employ artificial intelligence (AI) to:

• characterize the development of plant cells,

• evaluate the effect of applying pressure to the development of plant cells.

We combine artificial intelligence (AI)  models with mathematical models, such as Optimal Transport (OT) models, to achieve these tasks. 

To support large-scale cell analysis, we apply deep-learning-based computer vision models for cell classification, detection, and segmentation. However, image quality can directly affect model performance. Our dataset contains images with partial out-of-focus blur, as the cells do not lie on the same focal plane. Moreover, during the image capture process, environmental factors such as illumination conditions, microscope setups, etc., affect the data distributions, and thus might negatively impact the generalizability of models. Thus, we aim at developing robust computer vision models under their influence for cell analysis.