---
license: gpl-3.0
---

![](https://huggingface.co/Hemaxi/3DArtificialVision/blob/main/icon3dav_final.png)

# 3DArtificialVision: A Deep Learning Tool for Analyzing 3D Mouse Retinal Vasculature in Microscopy Images

## 🎓 PhD Project  
**PhD Program:** Biomedical Engineering  
**Researcher:** Hemaxi Narotamo  

**This project was conducted at:** 
- Instituto Superior Técnico (IST) – Instituto de Sistemas e Robótica (ISR)  
- Gulbenkian Institute for Molecular Medicine (GiMM)
- Católica Biomedical Research Centre (CBR) 

**Supervisors:** Professor Margarida Silveira (IST,ISR) and Professor Cláudio Franco (GiMM and CBR)

## Abstract

Blood vessels supply oxygen and nutrients to tissues, and their dysfunction contributes to diseases such as retinopathies and cancer. Understanding the mechanisms involved in angiogenesis, the process of forming new blood vessels from pre-existing ones, is essential for identifying key components of disease progression and for creating new therapeutic approaches.

Studies in mouse retinas indicate that endothelial cells (ECs), which line the inner surface of blood vessels, polarize and migrate toward chemokines during vessel growth and against blood flow during vessel remodeling. EC polarity was computed as a vector between the nucleus and the Golgi complex centroids. These findings rely on 2D analysis neglecting the 3D vascular structure. Furthermore, previous studies required manual nucleus-Golgi vector annotation and semi-automatic vessel segmentation, which are laborious and expert-dependent tasks.

To overcome these limitations, we developed 3DArtificialVision, an automated framework composed of deep learning methods to compute nucleus-Golgi vectors and analyze blood vessels in 3D mouse retinal microscopy images. First, we present 3DCellPol, a convolutional neural network designed to detect and group nuclei and Golgi centroids. 3DCellPol outperforms previously proposed multi-stage methods while requiring much less training supervision, only 3D images and centroid pairs. Then, we demonstrate that training 3DCellPol using a combination of real data and synthetic data, generated by a generative adversarial network, improved its performance compared to using real data only.  The final part presents 3DArtificialVision, which segments 3D blood vessels using the cycleGAN model and quantifies vascular parameters such as vessel density, branch length, vessel radius, and branching point density. It efficiently segments vessels and detects phenotypes similar to those observed through manual analysis.

The 3DArtificialVision framework is presented as a free user-friendly graphical interface suitable for researchers without programming experience. This open-source tool will significantly improve the ability to study 3D vascular networks in both health and disease.