Developments in vastly scalable IT infrastructure will soon increase the rate at which machine learning systems gain the capacity to transform the field of medical imaging across clinical, operational and business domains. Moreover, if the pace seems to be picking up, that’s because data management on a massive scale has advanced exponentially over just the past several years.
Consider: Current graphics-processing unit (GPU) technology uses several thousand processing cores and delivers more than 100 teraflops of deep learning performance. GPUs are particularly well suited to image processing, which is a perfect first fit within healthcare for convolutional neural networks (CNNs). That’s why CNNs are increasingly being applied to many medical imaging applications, most conspicuously (and least surprisingly) in radiology.
From a global view, it’s clear that the acceleration of performance improvement in medical imaging AI systems—specifically as supported by advances in data storage and other aspects of IT infrastructure—is just one of numerous game-changing scenarios likely to unfold in the weeks, months and years just ahead.
All are worth considering in some detail. Possibly none would be considered especially noteworthy outside of technology circles if not for one salient point: The medical imaging industry is still in a very early phase of its adoption of AI-fueled innovation. Whatever specific applications and new wrinkles emerge next will disrupt all aspects of the profession. That’s why now is the time for healthcare providers to begin constructing a machine learning strategy—and to leave room in it for adjustments to unforeseeable developments as well as expected end points.
Before charting a path to there from here, it will be helpful to consider some of the aforementioned near-future scenarios and what to expect as each unfolds.
- The market for machine learning is surging, fueled by a need to improve accuracy and efficiency. So far, the use of machine learning in routine diagnostic imaging has been limited. That is poised to change dramatically in the coming years as CNNs and other deep learning techniques proliferate. Forces driving the market for AI-based software that analyzes medical images include increasing radiologist workloads, errors and discrepancies in radiologists’ image interpretations, and the increasing use of AI for nonclinical applications such as workflow prioritization, quality assurance and overall practice management.
- Machine learning will improve the care experience for radiology patients. From initial scheduling of the imaging examination to final diagnosis and follow-up appointments—and every step along the way, including image acquisition, findings reporting and treatment planning—AI and machine learning will facilitate smarter, more streamlined processes that help achieve better outcomes while also adding to overall patient satisfaction. For providers this will translate to operational efficiencies and substantial cost savings.
- Research and innovation will yield new and compelling use cases. As of now the radiological subspecialty furthest along with AI is screening and diagnostic breast care. This is so because the field has been using pre-AI computer-assisted detection (CAD) techniques to find cancers for many years, and machine learning has been a natural next step. However, use cases for advanced AI-aided image analysis are well in development for diseases of the lungs, brain, heart, musculoskeletal structures and all internal organs.
- Technology vendors will offer new medical imaging AI products as quickly as academic research centers can test and prove new algorithms. Studies on machine learning have been expanding rapidly in recent years in peer-reviewed medical journals. Accordingly, established technology vendors are collaborating with the researchers at a heightened pace, seeking to remain competitive in an AI-driven healthcare market. To that end, strategic distribution and technology licensing partnerships are increasing and the first online AI marketplaces are already open for business.
- No ‘ology’ will be left behind. Radiology isn’t alone in its readiness to adopt machine learning advances for various use cases. Ophthalmology and oncology are gaining ground, while dermatology, surgery and pathology are all in the game. In fact, pathology is something of a sister ’ology to radiology, since both depend on ultrafine image interpretation of anatomical systems. Over time, quantitative imaging biomarkers extracted from diagnostic imaging (radiomics) will be combined with quantitative biomarkers of pathophysiology and other sources of patient data for better prognostic prediction of disease progression. This will enable improved treatment planning and better patient outcomes.
- Health providers will leverage data hubs to unlock the value of their data. The push for more integrated care across U.S. healthcare is combining with the rise of machine learning to demand optimized data interoperability. This is best achieved via centralized platforms on which healthcare provider organizations can build a base for all diagnostic medicine and care management. Forward-thinking providers are moving to enterprise data hubs, which will increasingly facilitate access and interoperability of both structured and unstructured data across multiple clinical applications. With such a broad base of data to draw from and train on, machine learning algorithms will be able to produce ever richer and deeper insights.
Advances in AI-ready IT infrastructure represents the cornerstones on which modern healthcare will build a constantly learning system of continuous quality improvement across clinical, operational and business domains. Along with the aforementioned GPU servers, all-flash storage is gaining acceptance over legacy hard disk drives in health IT infrastructure due to the lower read/write latency, higher throughput performance and maximized scalability. As these kinds of infrastructure improvements continue to strengthen the architectures undergirding AI-aided systems in healthcare, watch for machine learning to continue improving care quality and patient experience even as it increases efficiencies and drives down healthcare costs.
Adapted from a white paper by Simon Harris, principal analyst with Signify Research. To download the full document, “What’s New for Machine Learning in Medical Imaging: Predictions for 2019 and Beyond,” click here.
View more features from this issue: