CUDA Spotlight: Supratik Moulik

Supratik Moulik

GPUs and Medical Imaging & Visualization

This week's Spotlight is on Dr. Supratik Moulik, founder of Triradiate Industries, a software development company based in Sugar Land, Texas.

Triradiate Industries is focused on automated and semi-automated systems for segmentation of CT (computer tomography) scans using GPU-accelerated hardware and software solutions.

Prior to founding Triradiate Industries, Supratik was a Fellow at the University of Pennsylvania specializing in cardiovascular imaging. This interview is part of the CUDA Spotlight Series.

Q & A with Dr. Supratik Moulik

NVIDIA: Supratik, how does GPU computing improve the field of medical imaging?
Supratik: GPUs provide the computational power that allows tasks like vessel and bone segmentation to occur in real time. By accelerating these tasks, more accurate results can be obtained to rapidly provide the data necessary for diagnosis and treatment decision making.

Triradiate Industries Vessel Tracking

Image of a virtual angiogram, an animated display of blood vessels from a CT scan.
The measurements can alert physicians to potentially life-threatening vascular problems.

Computational image analysis in medical imaging is still in its infancy. The role of computers in general and GPUs specifically is constantly evolving. Most radiologists would agree that quantitative analysis is the future of medical imaging.

NVIDIA: What has changed the most in your field in recent years?
Supratik: The scanners used by radiologists have been continuously improving in terms of efficiency and resolution. A consequence of these advances is that even a routine scan of the abdomen can produce up to a thousand images with an inter-slice thickness of less than a millimeter.

Triradiate Industries Liver Mass Viewer

The Triradiate alpha-blend is an image viewer that combines an OpenGL front-end with
a CUDA-accelerated neural network back-end for visualizing and characterizing tumors and masses.

The increase in scan size and data volume has produced an important challenge: more information has to be processed by the radiologist. The way to meet this challenge is to use computers to aid in the diagnostic process in addition to displaying the images.

GPUs in particular provide a way for complex visualization and analysis tasks (which previously required expensive custom hardware) to be performed quickly and with easily-attainable hardware. This translates into faster and more readily-available diagnostic tools which allow doctors to spend more time on patient care.

NVIDIA: How did you get involved with GPU computing?
Supratik: As an undergraduate at Carnegie Mellon University, I was a cross-disciplinary student in the engineering and physics departments. I had the opportunity to apply my practical engineering skills in various research projects.

One project that I worked on with Dr. James Russ during my senior year involved using FPGAs as co-processors used for retrieving data from particle detectors. As early as the late 1990s when I was doing that work with FPGAs, my eyes were opened to the computing capabilities of non-CPU processors.

Almost a decade later in 2007, I was working on image analysis software for segmenting CT scans and I thought back to my research experience at CMU. I decided to look into co-processor options for accelerating my code. It just so happens that around that time, CUDA v1.0 was being released. I began to explore CUDA, and made an early commitment to GPU computing which has paid off tremendously in terms of computation resources available to my algorithms.

NVIDIA: What is the major advantage of parallel programming with GPUs?
Supratik: In school, students learn how to program for CPUs and it is a great way to grasp the fundamental concepts of computer science. In CPU code, there is generally only one or a small number of computations occurring at a given moment in time, which provides a simple way to conceptualize the program flow.

The problem is that when you are approaching a problem with the CPU tool belt on, it forces you to make certain choices and do things in a specific and often restricted way.

GPU programming is the evolutionary next step which allows programmers to move past some of the barriers that restrict problem solving. For example, when running optimization routines for organ segmentation, the GPU computing model allows for a more thorough search of the variable space in a shorter period of time -- which increases confidence in the results.

For these reasons, I prefer to avoid CPU code for computational portions of a program because now, rather than trying to fit the problem to the processor, I am able to approach the problem more directly with the GPU.

NVIDIA: You mentioned that something happened at GTC 2010 that was an eye opener…
Supratik: The first time I attended GTC was in 2010. During the introductory sessions I accidentally walked into a talk regarding OptiX, NVIDIA's ray tracing software. Since I didn't want to be rude to the presenter, I stayed and listened to the whole lecture.

About half way through the talk, it occurred to me that the approach that they were using to create acceleration structures and cast rays was mathematically equivalent to the problem of vessel tracking that I was working on.

Triradiate Industries Kidneys

3D rendering of segmented vessels (in white) and bones (in gray) overlaid on CT data. Vessel tracking is a core component of the image segmentation algorithms Triradiate Industries is developing and perfectly suited for CUDA acceleration.

After the conference, I went back and completely rewrote my vessel tracking algorithms to reflect this new perspective on a problem that people have been working on for over three decades. It turned out that by trying to extend CPU-based thought processes to GPU algorithms, I had been limiting the ability of my algorithms to fully utilize the parallel nature of GPUs.

Triradiate 3D Viewer Demo on YouTube

NVIDIA: Did you have the chance to attend GTC 2012?
Supratik: Yes. In particular, I found the keynote by Iain Couzin (of Princeton University) regarding collective dynamics to be eye opening with regards to the potential for intelligent distributed decision-making by simple systems.

I hope to find a way to incorporate these simple interactions into the framework of my algorithms in order to allow for more robust and adaptive image segmentation.

Connecting at GTC with others working with GPUs and CUDA has undoubtedly changed the direction of my work and I look forward to learning new and exciting things at GTCs in the years to come.

Bio for Dr. Supratik Moulik

Born and raised in Houston, Texas, Dr. Moulik received a B.S. degree in chemical engineering and physics from Carnegie Mellon University, followed by medical school at the University of Texas Southwestern Medical Center, Dallas. His residency was in diagnostic radiology at Ohio State University. Most recently, he did a cardiovascular imaging Fellowship at the University of Pennsylvania. Upon completion of his training, Dr. Moulik began clinical practice while founding Triradiate Industries, a software company specializing in medical image analysis and computer-aided diagnosis software.

Relevant Links

Triradiate Industries
Paper on GPU Computing Medical Image Analysis:
GTC 2011 Presentation:

Contact Info

Supratik (at) triradiate (dot) com