A CNN-GNN Approach for Polarity Vectors Prediction in 3D Microscopy Images

The polarity between nuclei and Golgi is an important aspect of cellular division, migration and signaling. For example, nucleus-Golgi polarity significantly impacts angiogenesis, the physiological process in which new blood vessels develop from pre-existing vessels. Therefore, estimating polarity between nuclei and Golgi is crucial to understanding complex cellular mechanisms.Previous approaches for the prediction of nucleus-Golgi polarity vectors use simple versions of bipartite matching algorithms. These methods, while effective, fail to consider other important interactions between nuclei and Golgi, that can be explored with more advanced approaches, namely in the realm of deep learning.In this work, we propose a novel approach that combines a convolutional neural network (CNN) with a graph neural network (GNN) model. Firstly, we use a CNN for the detection of nuclei and Golgi centroids in 3D microscopy images of mouse retinas. Subsequently, we employ a GNN to predict links between nuclei and Golgi. To our knowledge, this is the first work that proposes a GNN for nucleus-Golgi link prediction. The proposed approach detects 66% of the polarity vectors outperforming the traditional methods by a significant margin. Hence, the presented method establishes a path for the automated and accurate detection of polarity vectors, facilitating the study of various cellular processes.Clinical relevance - Nucleus-Golgi polarity vectors provide crucial information about cell polarization, signaling and migration which is of great importance to unravel the complex mechanisms involved in the process of angiogenesis. The method proposed in this work enables automated prediction of nucleus-Golgi polarity in 3D microscopy images of mouse retinas, reducing the burden of manual shape or vector annotation.