The Medical Information Specialists, Medical Imaging and AI
Eric J. Topol, a world renowned cardiologist, scientist and author of three instructive books about the future of medicine, called some medical specialists “information specialists” specializing in medical imaging, in his JAMA opinion paper with Dr. Saurabh Jha, and more recently in his latest book — Deep Medicine.
This has nothing to do with the importance of the specialists or their specialty. On the contrary, these specialists — Pathologists and Radiologists — provide some of the most significant sources of diagnostic data for healthcare providers across the board.
The point here is that both of their jobs essentially entail interpreting patterns on images. Dr. Jha and Dr. Topol put it like this in their paper: “The primary purpose of radiologists is the provision of medical information; the image is only a means to information. Radiologists are more aptly considered “information specialists” specializing in medical imaging. This is similar to pathologists, who are also information specialists. Pathologists and radiologists are fundamentally similar because both extract medical information from images.”
We know that image data are quantifiable and that means that Artificial Intelligence (AI) tools are highly capable of processing them and assisting with their interpretation in a timely, and sometimes, more accurate fashion. In fact, many of the major advances in AI have focused on image analysis, and medical imaging is one of the top and most promising use cases for AI in healthcare. Advanced analytics that can drill down to the pixel level on large digital images may allow specialists to identify nuances that the human eye may miss, and AI can screen through images or slides quickly to direct specialists to the right thing to look at and assess, increasing efficiency and the value of the time spent on each case.
Apart from the value of the diagnostic data from these specialists, the shortfall of Radiologists and Pathologists in many parts of the world, and the quite significant rates of false-positives and false-negatives, especially in screening programs, are some reasons that reinforce the importance of this use case.
If we agree that this is an area in healthcare where AI can have a potent impact on the quality and availability of the healthcare we deliver, and I think most of us do, there are some things we need to do/improve to aid adoption and application of AI to medical imaging.
Tackle Fundamental Challenges
It is widely known that one of the first and most prevalent challenges to the development and use of analytics tools in healthcare is access to data, and then quality of that data. This lack of access to large, quality datasets largely impedes the validation of a good and growing number of AI algorithms, especially causal discovery algorithms, and their implementation in medical practice.
There is need to develop more infrastructure to channel, in this case, imaging data from multiple sources into secure data lakes and warehouses to which researchers and innovators may have access within provided regulatory frameworks, in order to enhance development, validation, transparency, reproducibility and generalizability of these algorithms. Some commendable efforts in this direction exist already, like the Joint Imaging Platform of the German Cancer Consortium and the Pittsburgh Health Data Alliance.
Nonetheless, the challenges are not over once algorithms are successfully developed and deployed. These AI tools need to be constantly monitored to ensure that they remain effective and relevant in the way they assist clinicians. For example, the efficacy of an AI algorithm could possibly deteriorate due to changes and updates in software or hardware like slide scanners, MRI machines or other imaging equipment.
As the trend is towards AI tools and models that require minimal or no human participation, the big challenge is how to monitor and analyze these changes in performance of the AI tools. Radiologists and Pathologists may not be able to readily identify systemic errors and biases, and so hospitals and systems that use these tools will need to come up with automated or semi-automated monitoring systems that can evaluate performance over time.
Cooperate to Develop Better Solutions
Yes, AI algorithms will be able to do a lot (probably most) of what Pathologists and Radiologists currently do. Our feelings about this do not play much of a role in this development. In their paper, Jha and Topol aptly remarked, “To avoid being replaced by computers, radiologists must allow themselves to be displaced by computers.”
They continue, “While some radiographic analyses can be automated, others cannot. Radiologists should identify cognitively simple tasks that could be addressed by artificial intelligence, such as screening for lung cancer on CT… Some tasks once performed manually by pathologists have been automated, such as cell counts…Artificial intelligence can perform [even] the more complex tasks of pathologists and, in some instances, with superior accuracy…Even though such studies need larger-scale validation with more diverse tissue types, it is clear in both radiology and pathology that many tasks can be handled by artificial intelligence.”
Currently, AI solutions are often developed with little or no clinician input, and this leads to labyrinthine technologies that are “distasteful” and rejected by clinicians. Collaboration between clinicians and developers needs to be actively fostered to ease the design of clinically-friendly AI. Radiology and Pathology residents should also be introduced to these technologies during their training to enable them interact effectively with the “other side” , better understand, and play active roles in the development of these solutions.
Multidisciplinary teams of data scientists, clinicians, computer scientists and medical ethicists are required to develop robust AI tools in healthcare. If these clinicians must use these tools in practice and harness their full potential, they need to have a fair understanding of how they work and where/how they can go wrong or fail.
Looking Forward
Again, there are a lot of potential roles for AI to play in all aspects of medical imaging. The list is long and alive. Many relevant insights are believed to lay deep in ever more complex medical images and AI can help us derive them.
There are use cases, such as during this pandemic, where it could be used for rapid screening. To protect healthcare providers and other patients, it could be used to rapidly screen (in seconds) patients admitted for reasons other than COVID-19, which could facilitate early detection of the virus and inform triage, isolation and testing for confirmation (results take hours/days).
Some AI technologies are focusing on the beginning of the imaging pipeline. They aim to enhance the quality of the images we get by reducing noise and artifacts and enhancing contrast, thereby providing physicians a clearer view of the patient’s pathology. When the technology is in wide use, it would help us avoid trading off image quality for reduced scan times and radiation dose, and potentially minimize or eliminate the use of contrast agents.
No doubt, moving forward there is a need to rethink the current training and workflows of the “information specialists” to accommodate and exploit these useful technologies for the benefit of patients, clinicians and health systems. Other specialties must learn from them also, as AI technologies will affect and sometimes disrupt the way we all currently practise medicine…for good.
“We are not going to get to the point where all medical diagnoses are not requiring backup,” Topol said last year at the GPU Technology Conference in San Jose. “But we may get to a point [where] certain things like a sore throat or ear infections or a skin rash can be done completely algorithmically — both the diagnosis and the recommendations for treatment.”
It is therefore imperative to consciously shape the future in which AI is part of the healthcare workforce. AI tools could provide clinicians more time and energy to concentrate on providing the human touch. This is an opportunity and not a threat as EIT Health and McKinsey noted in their report, “The impact [of AI] on the workforce will be much more than jobs lost or gained — the work itself will change. At the heart of any change is the opportunity to refocus on and improve patient care. AI can help remove or minimize time spent on routine, administrative tasks, which can take up to 70 percent of a healthcare practitioner’s time.”
