Getting Started in Medical AI
A Guide to Online Learning Resources for Clinicians
(Originally published June, 23, 2021 on towardsdatascience.com)
As Artificial Intelligence (AI) transforms medical practice, all clinicians will become users of machine learning technology and will need to learn the basics of AI and machine learning. Clinicians needing to upgrade their AI knowledge may be baffled by the many available educational options.
I recently completed a four-month sabbatical, during which I investigated the growing patchwork of educational courses, videos, and books intended to build AI knowledge and skills. What follows are my thoughts based on that experience.
I have organized my suggestions by career trajectory, starting with basic resources for radiologists and other medical professionals who have minimal experience, and ending with advanced programs for anyone with a strong technical background and more ambitious goals. You can use these suggestions to tailor your own path based on your background, experience, and career aims.
Radiologists and Radiology Trainees
If you are new to AI and machine learning the AI for Everyone course, offered in Andrew Ng’s deeplearning.ai program through Coursera, highlights why AI is the focus of so many industries, not just healthcare. For healthcare-specific material, you can sample the webinars offered by the Radiological Society of North America (small fee), Society for Imaging Informatics in Medicine (SIIM) and American College of Radiology (ACR), as well as the research seminars, journal clubs, panel discussions offered by health AI research centers, including the YouTube channel of the Center for Artificial Intelligence in Medicine and Imaging (AIMI Center).
Whether you are a radiologist-in-training or a board-certified radiologist, the National Imaging Informatics Course (NIIC), sponsored by the RSNA and SIIM, provides an clear introduction to the broader world of informatics – a combination of recorded and live lectures, with a "flipped classroom" that enables small group discussion with informatics luminaries and fellow students. This curriculum is available to any residency program for a nominal fee, but residents will need to persuade their program director to provide a break from the clinical schedule. The course is also available to fully-trained radiologists for a fee.
If you attend the RSNA Annual Meeting, I highly recommend visiting the RSNA’s Deep Learning Lab. During a 90-minute Beginner Class, you will use your own laptop to build and test a deep learning algorithm that classifies medical images. This brief experience is the perfect complement to the didactic material above, and is critical to understanding how AI algorithms are built, including the pitfalls.
Research Collaborators and Practice Leaders
Many of you have ambitions to become more than just an informed AI user. Perhaps you may aspire to become the AI leader for your clinical practice – assisting in decisions about the adoption of AI algorithms, developing plans for the implementation and monitoring of AI technology, and deciding on policies for management of clinical practice data. Or you may seek to become a domain expert who collaborates with a research laboratory, or to works with industry on the development of machine learning tools. These activities require a richer understanding of neural networks and deep learning.
One of the fundamental tricks of machine learning is to re-cast complex computation tasks as a set of matrix operations. The graphical processing units (GPUs) that are the lifeblood of machine learning gain their power through speedy matrix math. If you never learned linear algebra or, like me, have forgotten most of what you learned, you will need to refresh your knowledge of vectors, matrices, tensors, and the related mathematical operations. The visually-oriented YouTube videos from 3Blue1Brown give crystal-clear visual explanations of matrix operations necessary for neural networks. The first 10 videos require about 90 minutes at normal speed.
The AI in Healthcare Specialization can serve as the foundation of anyone’s health AI Education. This course was released near the end of my sabbatical, but has already attracted thousands of learners. The specialization starts with an Introduction to Healthcare led by health economist Laurence Baker, who introduces the essential elements of a health care system, focusing on the U.S. Even if you are already working in the U.S. health system, this course serves as a useful health policy refresher. Many key points are directly relevant to AI, such as a discussion of the disparate data sources that may be available from patients, providers, and intermediaries.
The second course in this specialization, led by Nigam Shah, is all about health data. How can we find the right health care data to answer a specific question? What biases may affect the data because of how they were collected? How should researchers handle temporal data, missing values, and text? These questions are answered using specific examples that illustrate concepts like feature matrices and knowledge graphs. The Ethics module in Week 7 is particularly timely.
The third course, Fundamentals of Machine Learning for Healthcare, is focused on Machine Learning methods. Instructors Matt Lungren and Serena Yeung define essential terms and demystify jargon. You will learn about loss functions, gradient descent, and other key concepts like over- and under-fitting and concepts important to algorithm evaluation. The lectures on machine learning pitfalls, including discussion of causality, context, trust, and explainability, are invaluable. Anyone planning to begin a large project should pay close attention to the material in Week 6 on assembling interdisciplinary teams.
The fourth course, Evaluations of AI Applications in Healthcare, led by Tina Hernandez-Boussard, discusses how we decide when an AI algorithm is working well, including the concepts of utility and feasibility, and practical considerations of lead-time necessary for action. The importance of iteration during implementation and monitoring after deployment are emphasized. Week 4 provides an extensive catalog of biases that can affect AI evaluation, and discusses fairness, calibration, and transparency. Regulatory considerations and the FDA are included, as is a brief discussion of the global regulatory picture.
A capstone course, led by all the instructors, ties all the material together.
Machine Learning Researchers
Anyone seeking to become a full-fledged data scientist should learn basic software development skills. Excellent interactive courses for python, a popular language for machine learning tasks, can be found online at W3Schools, Codecademy, and many other places.
To learn the variety of deep learning methods for medical imaging, the full series of hands-on courses from RSNA’s Deep Learning Lab covers a wealth of topics including cohort building, data de-identification, training best practices, classification, and segmentation. As you begin building practical deep learning models, you can consult Brad Erickson’s Magician’s Corner, a series of articles in the Radiology: Artificial Intelligence journal that parallel the RSNA Deep Learning Lab courses and provide step-by-step guidance on system setup.
Jeremy Howard’s fast.ai course was my personal favorite learning tool and is the ideal resource for anyone with a strong technical background who knows a little bit of coding. The course takes a top-down approach starting with a few lines of computer code, rather than establishing detailed mathematical notation. The first lesson will guide you through the construction and testing of your own deep learning model. (Mine could distinguish 37 species of dogs and cats with 95 percent accuracy.) Each lesson demystifies deep learning using a hands-on approach, with pre-written code in Jupyter notebooks that you can explore, modify, and re-use. There are no abstract theorems or proofs to slog through, but by the end of the course you will know how to read and understand the Methods section of any deep learning paper.
I didn’t spend much time with the exercises because I am a coding veteran who now primarily manages projects rather than writing code. But that did not detract from my learning. For those who prefer the printed page, Jeremy just co-authored a book that provides the same material in the same order. A more advanced online course is also now available.
Andrew Ng’s groundbreaking and wildly popular deep learning specialization from deeplearning.ai on Coursera is a nice complement to the fast.ai course. Andrew takes a conventional bottom-up approach, starting with the theoretical basis for fundamental concepts in deep learning. He recounts the history and evolution of many seminal AI concepts, introduces standard notation, and highlights key references. Numerous exercises provide practical experience with what you have learned. I skipped the third course, Structuring Machine Learning Projects, because of my experience in that area, but others may find it useful. The final course in the series covers sequence models and will be of particular interest to radiologists and other clinicians working with text data. It reviews current progress in natural language processing methods, including Bidirectional Encoder Representations from Transformers (BERT) and other transformer methods. Like fast.ai, I found the entire course easy to follow without completing the exercises.
Reference Materials
For the truly hard core, several academic texts provide notation-heavy mathematical foundations for machine learning concepts. Deep Learning by Goodfellow, Bengio, and Courville, is available on Amazon, and freely available online. Neural Networks and Deep Learning by Michael Nielsen, is also freely available online. These resources are not for the faint-of-heart, but are useful as reference texts for those wanting a deeper dive in specific areas.
I wish you well as your pursue your career in machine learning!