Share this post:

Spectral MD deep learningThe speed and precision with which doctors diagnose burns can have a significant impact on recovery times. However, misdiagnosis rates are as high as 30 percent, even among specialists.

At Spectral MD, we developed a deep-learning tool called DeepView Wound Imaging System to revolutionize burn analysis. As we looked to bring our solution to market, we harnessed IBM Cloud bare metal servers with state-of-the-art NVIDIA Tesla graphics processing units (GPUs) to accelerate training of the solution.

Empowering doctors with AI

Until now, doctors have largely relied on the evidence of their own eyes to determine the best treatment for burns. When they misdiagnose, patients can be subjected to unnecessary surgery, or conversely miss out on treatment that could speed up their recovery.

To solve this challenge, we created the DeepView Wound Imaging System. Using multispectral imaging, it distinguishes between damaged and healthy human tissue far more accurately than is possible for the human eye.

Training the deep-learning algorithms that underpin our solution requires significant computing power. As we entered the clinical-trial phase, we knew we needed more performance to create a robust and fast model for accurately diagnosing and triaging burns.

Unleashing huge computational power

As a startup, purchasing and managing our own infrastructure didn’t appeal to us. Instead, we looked at cloud services providers. IBM Cloud immediately stood out for its flexibility and customization options. Based on the percentage of time we needed access to a GPU instance, we concluded that IBM would be less costly than other cloud providers.

IBM devised the ideal IT architecture to take our DeepView solution to the next level. Consisting of two IBM Cloud bare metal servers featuring leading-edge NVIDIA Tesla GPUs, it provides the heavyweight performance we need.

To create our models, we use thousands of multispectral images of accurately diagnosed burns to train our deep-learning algorithms. Next, we use a combination of gradient descent and cross-validation techniques to improve their accuracy.

We quickly saw the impact of IBM Cloud solutions on training times for the system. Before, we needed around three hours for the initial training run for just one algorithm. Switching to IBM Cloud solutions enabled us to reduce the required time to less than an hour.

The training process enables our algorithms to analyze the optical signatures of wound images at high speed, deducing how photons interact with tissues beneath the surface of a patient’s skin. Using this information, our solution can determine to a high degree of accuracy whether a wound is significant enough to require surgery.

Transforming recovery from burns

With IBM looking after our IT infrastructure, our engineering and scientific teams have more time to concentrate on bringing DeepView to the market quickly and cost effectively. What’s more, the IBM Cloud team keeps track of the latest technology, so we benefit from ongoing performance enhancements without needing to monitor the market.

The enormous processing power of the IBM Cloud bare metal servers with GPUs is dramatically reducing the time we spend training our deep-learning algorithms. We’re completing tasks in 90 percent less time than would have been possible otherwise.

We’re still in the clinical-trial phase, but once DeepView is in hospitals around the US, we anticipate that it will decrease the percentage of misdiagnosed burns from 30 percent to just 5 percent. Beyond that, it will give doctors the insight they need to administer more effective care, enabling faster recovery. With help from IBM Cloud, Spectral MD is transforming the diagnosis, triage and treatment of burns.

Read the case study and watch the video for more details.


via Cloud computing news

November 19, 2018 at 03:09PM