The winner of the medical imaging category in the AIMed 2017 Abstract Competition has taken first steps towards creating AI tools that go beyond a research project and can actually be used by clinicians in the hospital setting.
Peter D. Chang, MD a Physician Software Engineer and the Director of the Center for AI in Diagnostic Medicine (CAIDM) at the University of California Irvine Healthcare, told AIMed Magazine how researchers can progress from the abstract to the real thing.
AIMed: What types of applications based on deep learning technology can physicians expect to see in the next several years?
Peter Chang: A significant proportion of work performed by physicians on a day-to-day basis involves repetitive, tedious tasks that require a relatively low degree of skill. In medical imaging, this includes tasks such as measuring the size of a tumor or counting the number of abnormal findings on an image.
Not only would physicians prefer to do other, more stimulating activities, it turns out computers are often much more accurate and objective than humans in such tasks. In general, these types of simple applications can be easily solved today with state-of-the-art AI technology.
After all, compared to self-driving cars or digital facial recognition, the extent of technology required to solve these simple medical use-cases is relatively trivial.
AIMed: If these types of problems can already be solved with AI technology today, why have these tools not been translated into clinical practice?
PC: Despite the much-anticipated hype of AI in medicine, it turns out there are significant practical bottlenecks that limit the potential of most deep learning applications today.
Ultimately, many of these shortcomings exist because few if any physicians are actively engaged in developing the next generation of technology, AI or otherwise. It is interesting to note the vast majority of medical startup companies are founded with limited if any physician involvement or oversight.
Without experts that deeply understand both the medical and technical aspects of the problem, there is currently a significant gap in translating cutting-edge AI technology to healthcare.
For example, the vast majority of all research groups apply generic AI algorithms trained on everyday images (e.g. cats, dogs) directly to radiographic exams without any customization.
Furthermore, the large datasets needed to train the current AI algorithms are extraordinarily rare to find in healthcare, and when present are even more difficult to curate properly without the direct input of physicians who understand the strengths and limitations of deep learning technology.
Finally, despite the explosion of medical AI research in the literature over the past several years, few groups have invested in building the necessary infrastructure to prospectively validate these tools directly with physicians in a realistic clinical setting.
AIMed: Given these limitations, what changes do large academic research institutions and/or commercial entities need to adopt to reach the forefront of innovation in this field?
PC: To overcome the challenges of conducting large-scale, meaningful medical AI research, I believe partners in academia and industry will need to work together and consolidate resources. Even in a single academic institution oftentimes many individual groups may be working to solve similar problems in AI with minimal if any collaboration.
As director of the new Center for Artificial Intelligence in Diagnostic Medicine (CAIDM) at University of California Irvine, we are seeking to establish a new paradigm by providing resources (data scientists, software engineers, computational infrastructure) that can be shared across the entire University community by any research scientist or physician.
By maintaining a single robust pipeline for dataset discovery, algorithm development and application deployment, we hope to provide a mechanism for rapid translation of an idea into tools that physicians can use.
Additionally, a critical component of our new initiative is the active engagement of various industry partners to ensure any valuable tool created at our Center can be broadly disseminated into clinical practice. After all, beyond manuscripts or research grant awards, we believe the widespread adoption of new AI-enabled applications is the most important benchmark for success in this field.
AIMed: In your experience at the CAIDM, how do you choose a clinical problem to focus on solving using deep learning technology?
PC: Given the virtually unlimited potential number of problems in medicine to solve and the relatively nascent state of deep learning technology, it is critical to identify the right applications that are most amenable to AI today.
Interestingly, this means some high-profile, intellectually challenging problems in medicine such as cancer diagnosis are not a good fit because there is no clear way to integrate such a tool into the existing treatment paradigm.
For example, based on current standard-of-care any abnormal finding that could lead to even a remote possibility of cancer (e.g. new asymmetry on breast mammogram) will be sampled by a surgical biopsy to exclude the possibility of malignancy with near 100% guaranteed accuracy.
Even if an AI algorithm could predict the biopsy result based on noninvasive imaging alone with 99% accuracy, 1 out of every 100 patients would have a missed diagnosis of cancer, an unacceptable error rate for most medical institutions (e.g. result in litigation).
By contrast, there are many relatively simple tasks an AI tool can perform today with near human-like accuracy. As an example, in patients with stroke, therapeutic decision making is based on the timely detection of blood in the brain using CT scans interpreted by human radiologists.
Recently I have developed a deep learning AI tool capable of detecting even very subtle volumes of blood (<0.01 mL) with overall 97-98% accuracy after being trained on over 10,000 CT exams. Through a platform created at the CAIDM, this AI tool now automatically helps interpret all head CT scans obtained in our emergency room, alerting our radiologists to emergent findings requiring expedited review.
AIMed: After developing a new tool, what are the most significant barriers to widespread adoption into clinical practice?
PC: The Food and Drug Administration (FDA) and European regulatory agencies have very limited experience or precedent with the newer wave of deep learning technology, resulting in several potentially unforeseen obstacles.
For one, compared to traditional, more static machine learning techniques, deep learning tools are capable of self-improvement and evolution over time through continued learning on new data, a behavior that is unaccounted for by current regulatory guidelines.
Furthermore, given the flexibility of deep learning technology and relatively short development timeframe (resulting in an unprecedented large number of potential applications), significant changes to the current protracted approval process will likely need to occur to accommodate the explosion of new AI tools.
From a cultural perspective, there is a general negative bias in the medical community against AI-enabled applications. In part, this relates to the historic limitations of earlier generations of AI technology (e.g. CAD systems for radiology).
There is also a baseline inertia to change which, for some unclear reason, seems to weigh more heavily in medicine.
Finally, and perhaps crucially, it is not uncommon for physicians—certainly among the most highly-educated professionals in our society—to believe the insights they generate can fundamentally never be replaced by an intelligent machine system.
In all these cases, I believe education and open discussion will provide an important role in helping guide the medical community beyond these subconscious preconceived biases.
Dr. Peter Chang, MD is the recently recruited director of the Center for AI in Diagnostic Medicine, a new multi-specialty initiative to develop and integrate artificial intelligence technology across the UC Irvine healthcare system. He is also co-founder of multiple startups including most recently Avicenna.ai, a company focused on deep learning for medical imaging. Dr. Chang’s unique perspective arises from experience both as a radiologist physician and software engineer with expertise in developing deep learning algorithms. His work has led to dozens of abstracts, manuscripts, five best conference paper awards, two provisional patents/copyrights, and top finishes in various competitions including the international 2016 MICCAI challenge. Dr. Chang is regularly invited to speak at various national and international conferences, providing insight into the strengths and weakness of current AI technology as it relates to clinical problems in medicine.