The promise of AI in medical imaging: improving medical practice
Posted 3rd May 2019 by Joshua Sewell
In our lab of Quantitative Imaging and Artificial Intelligence, we’re developing AI applications in a variety of areas, such as radiology and pathology. The goal is to develop applications that meet unmet medical needs, particularly in relation to precision medicine and clinical prediction.
In order to develop these AI methodologies, we need to get our hands on a lot of data. We’re working in areas of deep learning particularly to try and mitigate that challenge. The most exciting thing is that with the advent of much more patient data in electronic form, the possibility of learning from that data is exploding.
Medical practice has been developing slowly over time through accrued knowledge and discovery. But with the induction of electronic data, there is now the opportunity to accelerate discovery because we can do analysis with much greater rapidity.
The prospect of improving medical practice with AI
The prospect of producing applications that will improve medical practice is enormous. The primary goal is to try and elevate all practitioners to the same level of expertise. There’s a lot of variability in clinical practice and an area where these technologies can have a great impact is not only improving efficiency but improving accuracy across the board.
AI methods offer tremendous potential to benefit patients because the current recommendations for treatment are based on the clinical experience of practitioners. Experience varies from practitioner to practitioner and is based on evidence gathered from published literature, which accrues slowly. Now, there is the promise of making increasingly more rapid advances as data science is applied to these large data sets.
The accuracy of practitioners in terms of their ability to detect, diagnose, and make appropriate recommendations is very well studied in radiology and to some extent pathology. As much of my work is in radiology, I’ll use an example from that field.
In mammography practice, positive predictor value of practitioners varies greatly. That variation is based on clinician experience and so the mammography regulations from MQSA require a minimal amount of mammography interpretation experience on an annual basis in order to be credentialed for making those interpretations. This illustrates the variability based on experience.
There are human limitations in the ability to perceive abnormalities, and then to render a correct diagnosis. A certain amount of that can be rectified by just having enough experience, but there is also a gap in terms of clinical performance. Therefore it is perhaps unsurprising that studies show that if an image is looked at by two different people, the interpretation accuracy improves: a clear example of how two heads are better than one.
There is also a limited ability of the human eye to see abnormalities in images. There have been some recent studies looking at pathology image features that can predict survival that is not at all apparent to the human eye. There are examples of AI improving practice not simply based on variation in clinical judgment, but based on the ability of an algorithm to perceive what the human eye simply cannot.
Thus the two ways that AI can improve medical practice when it comes to imaging is to function as a second opinion and to detect features otherwise imperceptible to human vision alone.
The challenges of building AI applications
When it comes to building these AI applications, there are a number of technical challenges:
- How do we use unstructured data types like images and text documents to build AI models?
- How do we integrate multiple different pieces of information to build more robust AI models?
- How do we get computer algorithms to process longitudinal data and detect important changes in patients over time?
Those are all technical challenges, which divide our work into two areas: one is focusing on applications, and the other one is technical development to enable those applications. But, as I will discuss in my next post, when it comes to addressing challenges, we really want to be focusing on the robustness and reliability of the applications themselves.
Daniel Rubin is Professor of Biomedical Data Science and Radiology at Stanford University, USA.
Leave a Reply