Up to 30% of burns are misdiagnosed. Spectral MD had developed deep-learning algorithms to improve diagnosis by analyzing images of wounds, but was struggling to scale its IT resources to meet demand.
Deployed IBM Cloud bare metal servers with state-of-the-art NVIDIA Tesla GPUs to train its DeepView Wound Imaging System faster, ultimately increasing the speed and accuracy of burn diagnoses.
Empowersdoctors to administer timely and effective treatment
90% fasterprocessing of wound data, accelerating diagnosis
HelpsSpectral MD reduce risk of misdiagnosis from 30% to 5%
Business challenge story
Helping doctors heal wounds
Diagnosing burns quickly and accurately is crucial for giving patients the best chance of a full and speedy recovery. However, doctors rely primarily on the visual analysis of wounds, which results in misdiagnosis rates of up to 30 percent even among experts. This in turn means that some patients undergo unnecessary surgery, while others miss out on surgery that would improve their recovery.
Aiming to enhance the accuracy of diagnosis, Spectral MD uses machine-learning techniques to amplify the power of multispectral imaging in analyzing the severity of burns.
Jeffrey Thatcher, Chief Scientist at Spectral MD, explains the science behind his company’s DeepView Wound Imaging System: “Chemicals and tissue structures can be characterized very accurately based on the light they reflect. By using multispectral imaging – including non-visible wavelengths of light – we can differentiate pathologies from healthy skin in a way that is impossible for the human eye.”
As Spectral MD brought the DeepView Wound Imaging System to clinical trial, the company found that its existing compute resources could not provide the processing power required to train the solution’s neural networks in building a robust model of wound types.
“We needed more computing power for the deep-learning algorithms at the heart of DeepView,” Thatcher says. “As a start-up with constrained budgets, we didn’t want to commit capital and operational expenditure to buying and running our own on-premises hardware.”
Boosting computational power
Eager to help Spectral MD develop its innovative solution for the benefit of patients and doctors, IBM invited the company to pilot IBM Cloud technology for size, performance and scalability. Based on the percentage of time the company needed access to graphics processing units (GPUs), Spectral MD concluded that IBM Cloud would be less costly than competing offerings.
“We met with IBM consultants and explained what we needed to take DeepView to the next level,” Thatcher says. “The IBM team took the time to understand our algorithms and designed an IT landscape that suited our needs perfectly.”
Spectral MD deployed two IBM Cloud bare metal servers equipped with state-of-the-art NVIDIA Tesla GPUs processors for ultra-fast computation. To build accurate models of different types of burns, Spectral MD feeds thousands of multispectral images of accurately diagnosed burns into its machine-learning algorithms. Training the system to accurately recognize and classify new images is a complex process that starts with a technique called gradient descent, then uses repeated cross validation to verify the accuracy of models.
Previously, it took approximately three hours for Spectral MD to complete the initial training run for a single algorithm. In a typical scenario, Spectral MD subjected four different algorithms to 20 iterations of cross validation each in order to find the best approach. “If you want to compare four algorithms, that’s over a week’s worth of time if you’re not using a GPU,” Thatcher says. “With the support of GPUs on our IBM Cloud bare metal servers, we can reduce the training time from three hours to less than one, and we can compare four algorithms in less than a day.”
Once trained, Spectral MD’s algorithms can rapidly analyze the optical signatures of wound images to infer information about the interactions between photons and the tissues beneath the skin's surface. Based on their deep-learning experience, the algorithms can determine with high accuracy whether a wound is deep or serious enough to require surgery.
Using resources on the IBM Cloud rather than an on-premises solution gives Spectral MD significant advantages in terms of the speed and flexibility of adopting new technologies.
“One of the things I love about our partnership with IBM is that they are in frequent communication with our team and always share the most recent developments in GPU technology with us,” continues Thatcher. “This is a fast-moving space, with engineers and leading technology companies regularly releasing new solutions. Our engineers are often too caught up in developing our algorithms to be able to follow the next big product release in the GPU world, so the fact that IBM keeps us up to date is a big plus.”
Spectral MD’s focus is on developing convolutional neural network (CNN) architectures for image segmentation. Each CNN may have over 100 million parameters, and twice as many computations are performed during each of the thousands of iterations of gradient descent carried out during training. In addition, the team experiments with numerous variations of CNNs, which involves training and re-training algorithms many times. The parallel processing architectures provided by GPUs offer far superior performance for the types of computations that occur in a CNN, making them ideally suited for the development of these models.
Aiding patient recovery
With IBM Cloud bare metal servers powering its DeepView Wound Imaging System, Spectral MD has the high performance, scalability and processing power necessary to develop its solution in the clinical-trial phase.
“Knowing that IBM is taking care of our IT infrastructure needs takes pressure off our engineering and scientific teams, giving them more time to focus on advancing the frontiers of medical science,” Thatcher says. “Developing our solution in the IBM Cloud also means we can utilize cutting-edge GPU processor technology and scale to handle increased data volumes, without incurring the cost of purchasing or maintaining hardware ourselves.”
Already, Spectral MD has seen significant reductions in the time required to train machine learning algorithms in the DeepView Wound Imaging System.
“Thanks to the power of our GPUs, we can complete in less than one day machine learning tasks that previously would have taken us up to 10 days,” Thatcher notes.
Once Spectral MD brings the DeepView Wound Imaging System to hospitals around the country, the company expects to reduce the percentage of misdiagnosed burns from 30 percent to just five percent.
“The DeepView Wound Imaging System has the potential to truly revolutionize the diagnosis, triage and treatment of burns,” concludes Thatcher. “With more confidence in their assessment of a patient’s wounds, doctors will be able to administer the most appropriate care in a targeted and timely manner. What’s more, with accurate diagnosis and personalized treatment, we hope to enable more patients to make a full recovery from their injuries. IBM Cloud technology helps us make all of this possible.”
About Spectral MD
Based in Dallas, Texas, Spectral MD is a medical device start-up company at the clinical-research stage of development. The company has received FDA 510(k) clearance for the first generation of its DeepView solution, which uses advanced multispectral digital imaging, augmented by machine learning, to give clinicians an accurate view of subcutaneous tissues. The solution achieves this in a non-invasive way and without emitting any harmful radiation.