Projects

We work on many projects, and the following is just a snapshot, as of May 2020.

Monte Carlo simulation with GPU parallel computation

Monte Carlo (MC) simulation is a versatile method that is widely used in radiation medicine. It is capable of faithfully describing the physical processes and flexibly handling complex geometries. Over of years, it serves as one of the most accurate methods for radiation dose computation and has increasing applications on radiation track simulation and computation of radiation-induced DNA damages. However, MC simulation is also a computationally demanding task. Computational efficiency has become a bottleneck for CPU-based simulation packages in advanced applications. Parallel computation via graphic processing units (GPUs) can be a cost-effective option to overcome the hurdle, while NVIDIA Compute Unified Device Architecture (CUDA) is especially preferred, considering its good documentation support.

In our group, we targeted on novel GPU-based MC package development and application. So far, we have developed two packages, gMicroMC and gPET, under a joint effort with Dr. Xun Jia’s group in the University of Texas Southwestern Medical Center.

• gMicroMC is a novel microscopic GPU-based MC simulation platform that deals with radiation track and DNA damage calculation. Its first version development enables water radiolysis and DNA damage computation from a single electron. It achieves a 520x speedup with a single GPU card over that executed on a single CPU card. The first version open-source code is available at the Github repository of utaresearch/gMicroMC. New development efforts include support of new physical and chemical modules, various DNA geometries, etc., while its application has been focused on quantitative understanding of radiation phenomena under various radiation scenarios.  
• gPET is a GPU-based MC simulation package dedicated for positron emission tomography (PET) imaging simulation. A speedup factor of 500 was obtained for gPET performed on a single GPU over GATE8.0 on a single CPU. The open source code of gPET is available at the Github repository of utaresearch/gPET. gPET is currently used for high resolution, high sensitivity small animal PET designs with our collaborators in UC Davis.

We are also maintaining another GPU-based MC simulation package of gDPM, initially developed by Dr. Xun Jia’s group. We are currently using gDPM for dose verifications in various radiation conditions.

Microbeam irradiation

To understand the radiation induced biological effect, except for the computational effort, our group also works on designing and constructing microbeam irradiation facilities. So far, two approaches are under investigation:

• Alpha irradiator. This is to build a cost-effective and easy-to-maintain microscopic radiation beamline in our lab. The idea is to focus the alpha particles from a natural radiation source via an electromagnetic lens system, such that the cells can be irradiated at a controlled and constant dose rate manner. The first version of the irradiator is designed based on the permanent magnetic focusing lens (simulation performed with COMSOL), while the entire system is moving from the drawing board to under-fabrication.
• Accelerator based low MeV ion microbeam line. This is to build a more versatile microscopic beamline enabling subcellular matrix irradiation with various ion species, beam currents, dose rates, etc. A novel electrostatic focusing quadruplet lens system was simulated (with GICOSY and SIMION). Another more powerful electrostatic focusing triplet lens system was also simulated and its aberration is under investigation. The ultimate goal of this project is to obtain a compact and high focusing lens system that could be attached to existing low MeV accelerators for radiobiological study.  

Deep learning based auto-treatment planning

Dose-to-volume histogram (DVH)-based optimization engine is widely used in modern treatment planning systems (TPSs). However, it is usually a complex process to perform inverse treatment planning using these TPSs. To obtain a clinically acceptable treatment plan, it is usually a labor intensive and time consuming process for a human planner to tune the relevant parameters.

Reinforcement learning (RL) is known for its vast capability of solving problems where sequential decision making is needed. RL algorithms that incorporate deep neural networks, known as DRL, can beat human experts playing numerous games, as well as the world champions of Go. Its general structures contain:

• Agent: the entity that performs actions in environment and gets rewards based on the action it takes;
• Environment: the world that the agent interacts with;
• State(s): a concurrent and immediate situation returned by environment that the agent can take actions against;
• Action: a response given by the agent facing a given state;
• Reward: the feedback by which we measure the success or failure of an agent’s actions in a given state.

In this project, we are trying to train a DRL neural network, which can mimic the behavior of human planners by automatically operating the DVH-based optimization engine for high-quality treatment plans. 

Molecular Dynamics based Radiobiological Studies

Molecular dynamics (MD) simulations predict how every atom in a protein or other molecular system will move over time, based on a general model of the physics governing interatomic interactions (Karplus and Mc Cammon, 2002).

In our research, we are interested in applying this technique to capture the biomolecular processes of critical cellular matrices, such as DNA and the cellular membrane, in response to ionizing radiation. When the chemical and physical environments of macromolecules are perturbed by ionizing radiation, dynamic bond formation and breakage, as well as conformational changes, occur. By capturing these changes, we aim to develop a fundamental understanding of the biological effects of ionizing radiation and help elucidate clinically relevant mysteries.