Fighting Tuberculosis with GPUs and Deep Learning

For those in developing countries with tuberculosis, the difference between life and death often comes down to having a physician with the expertise to properly read chest X-rays. And the numbers show that many are dying of the disease unnecessarily.

TB has passed HIV/AIDS as the world’s top infectious killer, with the World Health Organization estimating that 1.8 million people died from the disease in 2015. Some 95 percent of those deaths occurred in low- and middle-income countries where access to radiological expertise is often minimal.

A pair of researchers at Philadelphia’s Thomas Jefferson University aim to change that. By combining their passions for chest X-rays and deep learning, Paras Lakhani, an assistant professor of radiology, and Baskaran Sundaram, a professor of radiology, may have opened the door to stemming the incidence of TB.

“A lot of developing countries just don’t have the resources to deal with these challenges,” said Lakhani.

Deep Dive Into AI

Fortunately, Lakhani’s decision two years ago to dive into deep learning may lead to a solution. He said he got “really obsessed” with the method, and read hundreds of papers before acquiring a GPU and building his own machine.

He then obtained public TB datasets from the National Institutes of Health, the Belarus Tuberculosis Portal, and Thomas Jefferson University Hospital, so he could start testing models.

Armed with an NVIDIA TITAN X GPU, and supported by the Caffe deep learning framework, CUDA and cuDNN, as well as NVIDIA DIGITS deep learning GPU training system, Lakhani and Sundaram trained a model using the more than 1,000 public TB images they had vetted.

Lakhani said the model ran 40 times faster than the benchmark tests the pair had run on CPUs.

The work could lead to healthcare providers in developing countries being able to upload chest X-rays, compare them against Lakhani and Sundaram’s model, and then accurately diagnose any anomalies.

Before that happens, however, there are additional challenges to overcome. For instance, chest X-rays are typically huge files of 2,500 by 3,000 pixels. And while GPUs can handle that level of resolution, the enormous amount of data tests the deep learning models.

Lakhani said he and Sundaram are constantly experimenting with ways to get around this — by uploading only portions of images with abnormalities, or by creating deeper networks, for instance.

“I’ve learned how to tweak the hyper-parameters much better, and basically build better models,” said Lakhani.

More Challenges on Horizon

The pair will also have to tackle an even more complicated challenge: The subtleties that are apparent in the largest X-ray files aren’t as easy to spot as the resolution shrinks.

As a result, Lakhani and Sundaram are working on achieving the right balance of resolution and data so that their models can detect the subtler indicators of TB infection.

That said, Lakhani understands there are limits to what the models can achieve.

“I’ve seen so many subtle examples of TB that I have my doubts that even models like this can catch it all,” he said.

Lakhani and Sundaram haven’t yet decided on how exactly they’ll put their model to work once it’s done, but they aren’t ruling out a commercial approach.

“I really don’t know what the best is to help the world,” said Lakhani. “Sometimes commercialization is best because you can reach out to people and work with nonprofits.”

What Lakhani is pretty certain of is that his work with Sundaram won’t stop with TB. He intends on taking what he’s learned and applying it to similar chest X-ray-related challenges such as chest fractures, pneumonia, lung infections, heart abnormalities and aorta issues, to name a few.

Feature image credit: Yale Rosen. Licensed via Creative Commons 2.0.

Similar Stories