Naked-eye stereoscopic imaging system for real-time medical imaging

"We evaluated a variety of development environments for parallel computing," said Herlambang, the researcher responsible for the CUDA implementation. "We ultimately chose CUDA because it allows us to use familiar C language syntax for development challenges.

One very interesting area of ​​imaging technology is naked-eye stereo imaging, which can display three-dimensional images without special glasses. This interesting technology has potential applications not only in entertainment, but also as a practical technology for a variety of professional applications. Professor Takeyoshi Dohi of the Department of Mechanical Information at the Graduate School of Information Science and Technology at the University of Tokyo and his colleagues studied NVIDIA's CUDA™ parallel computing platform and believed that medical imaging is one of the very promising application areas of this platform.

Since 2000, the research team at this university has developed a system in which live cross-sectional images obtained in real time through CT or MRI scanning are viewed as volume textures, which can be reproduced not only as 3D images through volume rendering but also as stereoscopic video displays for use in IV systems.

This system revolutionized real-time, volumetric, in-vivo imaging. However, it is extremely computationally intensive, with the volume rendering alone creating a high processing workload, and then there is the further processing required to achieve volumetric imaging. For each image frame, there are many angles that need to be displayed simultaneously. Multiply this by the number of frames in a video, and you begin to see an astounding number of calculations that must be done with great accuracy in a very short time.

In the 2001 study, a Pentium III 800 MHz PC was used to process some 512 x 512 resolution images, and real-time volume rendering and stereoscopic reconstruction took more than 10 seconds to generate one frame. To speed up the process, the research team tried using an UltraSPARC III 900 MHz machine with 60 CPUs, which was the most powerful computer at the time. But the best result they could get was only five frames per second. From a practical point of view, this speed was not fast enough.

Solution

Volume rendering and subsequent conversion to IV format require data-parallel vector computation. For this purpose, the best computing paradigm is the GPU. Accordingly, Liao and Herlambang set out to investigate GPU implementation using CUDA, a general-purpose C-language GPU development environment from NVIDIA.

First, the researchers developed a prototype system using the latest generation of GPUs, the GeForce® 8800 GTX. When running the data set used in the 2001 study on the GPU using CUDA, performance improved to 13 to 14 frames per second. The researchers were surprised that the GPU delivered nearly three times the performance, even though the UltraSPARC system cost tens of millions of yen and was hundreds of times more expensive than the GPU. Moreover, according to the team's research, NVIDIA's GPU is at least 70 times faster than the latest multi-core CPUs. In addition, the tests showed that the GPU's performance was even better for larger-scale volume texture data.

Currently, the research team is using NVIDIA's latest desktop supercomputer, the Tesla™ D870, to optimize the current IV system for Tesla using CUDA, a move that is expected to result in even greater performance improvements.

Effect

In addition, we can take advantage of newer, faster GPUs without having to modify existing systems. If an environment makes it possible to debug large CUDA programs, CUDA will become a more powerful parallel computing development environment, and we hope it will be more widely used in the field of medical imaging processing.”

If images from CT and MRI can be viewed in real time in a three-dimensional manner, doctors can examine the state of patient tissue and make diagnoses without the need for biopsies and surgical procedures. In addition, some doctors can view such images simultaneously and communicate with each other. This enables some doctors to perform arthroscopy and other minimally invasive surgical techniques simultaneously, with each surgeon observing the procedure in real time.

It is very difficult to introduce large parallel computing arrays into clinical devices, but the powerful computing power of GPUs and Teslas makes it possible to provide compact parallel computing modules.

Integral Videography (IV) principle diagram

Image and text: calculation of IV element image calculation of IV pixel image Voxel data virtual lens array

Flat display

Space image formation: Space image formation

Micro convex lens array: observer: observer

Solution

Example of viewing IV images from a distance. In front of the display

At a distance of two meters, a very realistic, three-dimensional image of the yellow bamboo pole is formed, which looks like it is held in the hand. Even when the observer moves, the image remains visible as if it is "held in the hand." To create a high-resolution, three-dimensional image, processing methods such as volume rendering are used, but they require extremely high computing power.

Effect

IV system using CUDA