From Images to Insight: AI and Digital Twins for Retinal and Systemic Diseases
Project Overview
This project creates a digital model of the human eye to help find diseases like diabetes and heart problems. We use computer programs to map blood vessels in retinal photos and then calculate how blood moves through them. This gives a better look at how the heart and brain are working compared to just using a still image. By using fast computer chips, we can give doctors a clear view of a patient’s health and help them choose the right care for each person.
What are the key results of your project?
We successfully developed a blood flow solver using the D2Q9 Lattice Boltzmann Method. This approach allows for the simulation of fluid dynamics within the complex and branching geometries of retinal vessels without requiring manual mesh generation for every individual vessel segment.
The simulation pipeline was implemented in C++ and CUDA, optimized for NVIDIA V100 GPUs. By leveraging GPU hardware, the solver achieves high computational throughput, making patient-specific simulations significantly faster than traditional CPU-based methods.
We successfully integrated the cuDecomp library to enable multi-GPU parallelization. The system can now automatically partition large-scale retinal vascular data across multiple GPUs, managing data communication through MPI to handle simulations that exceed the memory capacity of a single card.
How do you feel you have benefitted from completing this internship?
This internship provided me with experience in taking part in academic research, knowledge in medical image processing and fluid dynamics, and experience in CUDA programming.
It also made me aware that there are still a lot of things in academic research that can be parallelized by GPU. I would consider this path for my future career.