Physicians and technologists are finding new ways to use AI and deep learning to lessen the workload on doctors and to assist in diagnosis of diseases. But barriers to deploying AI models outside of the data center have been difficult to overcome.
Storage requirements have been particularly problematic in applications like medical imaging, where AI models must contend with vast quantities of data. For example, the transverse and cross-sectional 3D images of the eye recorded by techniques such as Optical Coherence Tomography (OCT) can have a resolution as small as 2 to 5 microns, leading to massive images.
Data centers can readily handle these workloads, but getting images into the data center is also problematic due to network bandwidth constraints and patient privacy concerns. Thus, developers are looking for ways to deploy deep learning at the point of data collection.
Here the challenge becomes one of compute requirements. To handle the workloads associated with imaging, developers have turned to heterogeneous mixes of CPUs, GPUs, FPGAs, and other processors. But these complex architectures can be difficult to configure and program.
As an alternative, developers can leverage off-the-shelf technologies. A new generation of AI hardware and software designed specifically for edge applications can greatly simplify development. Developers should consider using these APIs and frameworks to maximize the portability, flexibility, and scalability of their deep learning efforts.
From Concept to Clinic
The IEI solution, shown in Figure 1, is part of a quick-to-deploy computer vision system that includes a workstation, deep learning software development kit, and powerful network-attached storage (NAS) sourced from the company’s QNAP storage division.
We spoke with Don Yu, Director of Product Management at IEI, to understand how an AI-embedded deep learning edge solution improves the accuracy of diagnosing disease and how the technology works.
Yu explained, “As a primary use case we chose OCT imagery, which is a noninvasive imaging test that can be used to diagnose many eye diseases, including glaucoma, age-related macular degeneration (AMD), and diabetic eye disease.”
According to Yu, specialists must undergo extensive training to be able to read OCT images. And reviewing patient test images also takes a significant amount of time. “This kind of application is ideal for AI and deep learning,” said Yu, “and we’ve developed an edge system that handles both training and inferencing, dramatically reducing the amount of time needed to diagnose disease.” (See Figure 2.)
Advanced Technology Leads to Early Diagnosis
Diagnosing AMD quickly delivers real benefits to patients. AMD, like many diseases, does not present obvious symptoms in the early stages, and often goes unnoticed until a patient’s sight is failing. By that time, the disease has reached the middle or late stages, making the disease less operable and treatment less effective.
But if the disease is discovered early, treatment can begin. And the earlier treatment starts, the more likely a patient will retain his or her sight. The same is true for many other diseases, where early detection results in significantly better outcomes.
AMD is a major cause of central vision loss in the developed world, affecting 10% of people 65 years and older, and more than 25% of people older than 75. In the US approximately 2 million people have advanced AMD and more than 8 million have an intermediate form of the disease. These numbers are projected to rise by 50% by 2020, making early detection critical.
Traditional approaches to interpreting medical images, while highly accurate, take a considerable time to complete. They require multiple viewings and discussion by more than one specialist. In areas without such specialists, medical images must be sent out for diagnosis. And wait times often take several weeks before a diagnosis can be made and for treatment to begin.
By using an AI-embedded deep learning edge solution, doctors in a central location can make accurate diagnoses in less time. Therefore, patients can start their treatment earlier, often leading to better outcomes.
Putting the Pieces Together
In our discussion, Yu stated that deep learning success hinges on three factors: data, computing power, and algorithms. On the first factor, he said, “QNAP’s NAS is ideal for medical data storage. We also launched the QPACS app to function as a Digital Imaging and Communications in Medicine (DICOM) server application.”
But there is more to the story. The solution’s NAS supports add-in AI cards, providing additional computational power needed to run tasks more efficiently. For example, the Mustang-F100-A10 acceleration card leverages the Intel FPGA Deep Learning Acceleration Suite to meet extreme compute demands (Figure 3). The company also offers AI cards based on Intel® Core™ processors and Intel® Movidius™ processors, giving developers flexible options.
For algorithms, the system supports frameworks and libraries such as Caffe, MXNet, TensorFlow, and CNTK. Existing containerized solutions can easily be migrated to the platform, and new ones can quickly be started. It also includes a container station, which supports Docker and other container technologies. Through containers, users can select from a wide range of AI frameworks and libraries for easier development.
“IEI also provides bare-metal AI solutions for users to combine with QuAI libraries to meet wide-ranging AI model training needs,” Yu explained.
AI, Deep Learning, and Improved Outcomes
To deliver value to physicians and patients, an AI-embedded deep learning edge solution must:
- Provide the enormous amount of compute power, storage, and network connectivity that medical professionals must possess to manage a massive amount of medical images, in addition to genomic and patient data.
- Combine a range of highly technical, cross-disciplinary skill sets not held by most medical professionals and researchers to build a deep learning platform and data management system for healthcare.
- Set up drivers, containers, data backup/transfer, and network configurations that are beyond the expertise of most data scientists.
- Offer scalability to work in hospitals and clinics of all sizes.
- Meet compliance for HIPAA and other regulations.
IEI’s innovative solution illustrates how developers can meet these goals with off-the-shelf solutions. That’s excellent news for doctors and patients alike. The ability for AI and deep learning to help diagnose disease in less time and increase accuracy makes a significant improvement in people’s lives. In the very near future, AI-embedded deep learning edge solutions will become common in medical facilities.
About the AuthorMore Content by Robert Moss