3DCellPol: joint detection and pairing of cell structures to compute cell polarity

Cell polarity is essential for tissue structure and cell migration, and its dysregulation is linked to diseases such as cancer and vascular disorders. Understanding the associations between molecular mechanisms, such as genetic defects, and abnormal cell polarization can provide clinicians with valuable biomarkers for early disease diagnosis and lead to more targeted therapeutic interventions. Here, we present a deep-learning framework for cell polarity computation based on the association between pairs of objects. Our approach, named 3DCellPol, is trained to detect and group the centroids of two distinct objects. To demonstrate the potential of 3DCellPol, we use it to compute cell polarity by pairing two cell organelles: nuclei and Golgi. The vectors between nuclei and Golgi define the front-rear polarity axis in endothelial cells. 3DCellPol was evaluated on 3D microscopy images of mouse retinas. It detected 71% of the nucleus–Golgi vectors and outperformed previous methods while requiring much less supervision. Moreover, incorporating synthetic data generated by a generative adversarial network further improved detection to 78%. We additionally demonstrated our model's adaptability to 2D images by applying it to a public dataset of cervical cytology images, where polarity is defined based on the cytoplasm-nucleus vectors. In this dataset, our model detected over 90% of vectors. 3DCellPol's ability to robustly compute cell polarity is crucial for understanding mechanisms of diseases where abnormal polarity plays a key role, and it may contribute to improved diagnostics and enable targeted therapies. Hence, it is a valuable open-source tool for both biomedical research and clinical practice.