Medical Imaging Drives GPU Accelerated Deep Learning Developments
November 27, 2017 Nicole Hemsoth
Although most recognize GE as a leading name in energy, the company has steadily built a healthcare empire over the course of decades, beginning in the 1950s in particular with its leadership in medical X-ray machines and later CT systems in the 1970s and today, with devices that touch a broad range of uses.
Much of GE Healthcare’s current medical device business is rooted in imaging hardware and software systems, including CT imaging machines and other diagnostic equipment. The company has also invested significantly in the drug discovery and production arena in recent years—something the new CEO of GE, John Flannery (who previously led the healthcare division at GE), identified as one of three main focal points for GE’s financial future.
According to Flannery, the company’s healthcare unit has one million scanners in service globally, which generate 50,000 scans every few moments. As one might imagine, this kind of volume will increasingly require more processing and analysis capabilities cooked in—something the company is seeking to get ahead with in today’s partnership with Nvidia. More specifically, the two companies want to make deep learning a core part of the image processing and analytical workflow, which takes both hardware and software expertise.
Image recognition and analysis is at the heart of much of the revival of deep learning in the last few years, so it comes as no surprise that medical image analysis has been a target for deep learning research and development. GPU maker, Nvidia, has been at the center of much of the recent progress in AI hardware and software co-design since its accelerators have (so far) no peer for the training, and in some cases, also inferencing of neural networks. At the last few annual GPU Technology Conference (GTC) events, there has been keen emphasis on deep learning, but few areas had as many presentations in this realm as image recognition and analysis and the associated host of medical applications.
For GE’s diagnostic equipment business, the Nvidia partnership will extend to 50,000 imaging devices and will, according to GE Healthcare’s CEO, Kieran Murphy, “accelerate the speed at which healthcare data can be processed” so the company can use the “GPU accelerated deep learning solutions to design more sophisticated neural networks for healthcare and medical applications—from real-time condition assessment to point-of-care interventions to predictive analytics for clinical decision-making.” The goal is to cut down on further or unnecessary imaging scans and to get far faster results from higher resolution imaging from ever-smaller devices.
“The average hospital generates 50 petabytes of data annually, through medical images, clinical charts and sensors, as well as operational and financial sources. Yet, less than 3 percent of that data is actionable, tagged or analyzed. GE Healthcare and NVIDIA will harness more of this data by combining powerful applications built customers, best-in-class medical devices and the fast processing speeds of GPUs – all to make AI a reality in healthcare.”
For now, the partnership extends to key CT scanning machines and 4D ultrasound gear with additional work on bolstering the software ecosystem to support these devices. GE Healthcare will be among the first large-scale users of the GPU Cloud containerized approach, described in more detail here. This will be used as the analytics platform for both development and deployment of GE Healthcare’s device models.
There is little technical information about which GPU devices are deployed for this analysis for now, but a conversation with someone on the deep learning device side at GE will talk to us soon about what is required for training, whether or not these are embedded GPUs that do much of the work on the device with only post-processing done on the GPU cloud platform, and how data is managed. Until then, the important point is that the first use cases at scale of GPU accelerated deep learning in medicine will begin to emerge, highlighting how medical diagnostics could change in the future.
According to a recent report from the National Institutes of Health in the U.S., the opportunities for deep learning in healthcare will continue to mount—and extend far beyond medical image analysis. “Deep learning is well suited to medical big data, and can be used to extract useful knowledge from it. This new AI technology has a potential to perform automatic lesion detection, suggest differential diagnoses, and compose preliminary radiology reports,” in imaging, but has applications along the broader scope of precision medicine applications. The report points out potential complications for the field that case studies at GE Healthcare and elsewhere can begin to address, however. These include the building accurate training models (avoiding overfitting in diverse datasets), the black box problem, and the possible legal and ethical issues (diagnostics that do not involve human analysts).
Ultimately, the NIH concludes, “At present, radiologists experience an increasing number of complex imaging tests. This makes it difficult to finish reading in time and provide accurate reports. However, the new technology of deep learning is expected to help radiologists provide a more accurate diagnosis, by delivering quantitative analysis of suspicious lesions, and may also enable a shorter time for reading due to automatic report generation and voice recognition, both of which are benefits that AI can provide in the clinical workflow.”