Betting on Monte Carlo: GPUs a ‘Game Changer’ for Nuking Noise in Nuclear Imaging

Andras Wirth is like many early AI researchers: His deep learning ambitions only turned into reality because of a sea change in technology.

A physicist, Wirth wanted to run Monte Carlo algorithms to make leaping advances in nuclear imaging, which was previously computationally impossible without massive supercomputers.

A decade ago, his breakthrough came when his lab began using GPUs and the first CUDA release on the computationally demanding algorithms.

On Thursday at the GPU Technology Conference in Silicon Valley, Wirth, who leads nuclear imaging at Mediso Medical, spoke about his company’s groundbreaking work.

Wirth’s team of CUDA programmers runs Monte Carlo method transport calculations on GPUs to enhance image quality. This helps to eliminate the usual degenerating effects that come from inaccuracies in physical modeling.

Monte Carlo transport methods rely on modeling the physical processes that contribute to acquiring the image of a patient. For maximum precision, the modeling consists of simulating billions of photon tracks. These photon tracks are random by nature, thus the simulation itself has to be random —  just like the games in the city of Monte Carlo.

Besides improving the image quality of scans, the main issue for nuclear medicine is the need to lower the dose of injected radioactive isotopes without impairing the diagnostic value of the acquired images. Neural networks help cope with the increasing noise level while also maintaining the useful information with a performance that is unrivaled by  conventional methods.

The lowered dosages are a boon to patients and the facilities that administer the radioactive substances, and the GPU-accelerated technique behind it holds great promise across the field.

“This is a complete game changer — it can have an effect on every type of nuclear medical procedure,” Wirth said.

Los Alamos to Budapest

The Monte Carlo method dates back to research at the Manhattan Project in the 1940s. But it wasn’t until recently that researchers and engineers applied GPUs to the computationally demanding algorithms.

Wirth’s work with GPUs on Monte Carlo methods have added to the capabilities of Budapest-based Mediso’s software used in its cameras for SPECT scans. SPECT (single-photon emission computerized tomography) scans rely on radioisotopes that are injected into the bloodstream of patients. Clinicians then use specialized cameras to capture 3D images of organs.

Mediso trained its U-Net convolutional neural network architecture on 1,000 images of bone scans. U-nets are used in medical imaging to bolster image segmentation so that different areas of details can be outlined.

It took a lot of computing power to do these types of calculations, Wirth said. “Traditionally, only supercomputers were able to do these type of calculations,” he said. “Until, GPUs appeared for general computing, it didn’t even make sense to try out Monte Carlo particle transport calculations in medical imaging.”

GPUs Lower Dose

Radioisotopes administered in medical imaging are low-level carcinogens for patients, expensive for imaging facilities to obtain and require special handling.

“Nobody likes to have nuclear isotopes in their body. That’s why we want to minimize the dose injected to the body — there are risks,” said Wirth.

However, when you lower a radioisotope dose, those lines are more difficult to decipher and blurring occurs that makes it difficult to spot lesions in bones.

Mediso used its neural network solutions running on GPUs to help to minimize that imaging “noise” while reducing the radioisotope dose administered to patients by one-eighth.

“It’s hard to imagine developing neural network-based products without the help of GPUs nowadays. It doesn’t stop there, however: since processing time is crucial in medical imaging, GPU technology has become a vital element of imaging products,” Wirth said.

The post Betting on Monte Carlo: GPUs a ‘Game Changer’ for Nuking Noise in Nuclear Imaging appeared first on The Official NVIDIA Blog.

by Scott Martin via The Official NVIDIA Blog