Radiologists as Knowledge Experts in a World of Artificial Intelligence (Summary of Radiology Residents and Fellows AI journal club).
Recap….
At 1800 EST on 6th Dec 2017, we held our first Artificial Intelligence journal club, a success in itself given over 102 registered attendees with 60 people joining the call. There are many professional organizations around the practice of Radiology. The AI journal club is brain child of Dr. Daniel Ortiz who serves as the vice chair for the resident and fellows section of the ACR (American college of Radiology ) , the organization that deals with primarily the business of Radiology including accreditation of facilities and legislature on issues like the appropriate age when women should get their screening mammograms.
AI in medicine has been trending on the news, and during the annual RSNA 2017 conference in Dec 2017 which is the largest medical conference in the world , machine learning and artificial intelligence sessions were standing room only.
Most of the recently published articles have a familiar theme where computers beat radiologists.
While most speakers on AI in medicine have gradually moved away from the idea of “replacing radiologists”, the papers around deep learning focuses on man versus machine.
The reading assignment for the journal club was to watch a TED talk on AI by Jeremy Howard(past president of kaggle), who teaches the FastAi course with his wife Rachel Thomas available here http://www.fast.ai . The TED talk is available here https://www.npr.org/2017/04/21/524702525/jeremy-howard-will-artificial-intelligence-be-the-last-human-invention . The highights of the TED talk that we reviewed are summarized below.
The big takeaway is the idea that Jeremy introduces the concept of service workers who form a large percentage of the workforce in the world as shown in this map.
A service worker performs the tasks below . A radiologist is in the class of service workers because we look at diagnostic images, speak to dictate our reports and integrate our clinical knowledge to get to a diagnosis.
With this context , Jeremy then references the future of work , and contrasts it with the industrial revolution which plateaued after a while when maximum innovation opportunities were met. He provides multiple examples of how the role of knowledge workers changes with better computers especially in a world where computers can listen , see and learn. With this background , we reviewed the role of radiologists in a world of AI or better yet attempted to answer this tweet from Andrew Ng on the future of radiologists https://twitter.com/andrewyng/status/884810469575344128?lang=en.
Dr Bibb Allen , the past chair of the ACR was our guest speaker and gave a phenomenon presentation of the ACR data science institute(ACRDSI) https://www.acr.org/Advocacy/Informatics/Data-Science-Institute. Dr. Allen reiterated that Radiology has always been changing …in fact he describes how frequent he would perform cerebral angiograms as a trainee, a task now performed only by the Neuro interventional radiologists. He reemphasized that it’s naive to think our practice will not change given this new revolution of informatics tools … but ultimately it will not be “man versus machine” but “radiologist and machine VERSUS radiologist without machine”. Why does Dr Allen think so?
Look at teaching how to drive a car , a skill impacted in a 16 year old in a matter of weeks .. then look at how long it’s taken to train autonomous cars , and still we don’t feel 100% comfortable letting the car drive itself… Or the autopilot example where we both appreciate having the human pilots and the machines working together..
We reviewed the history of Radiology … Radiology 1.0 with use of film , 2.0 in the introduction of digital systems and 3.0 focusing on bringing value back to the patient… 4.0 is surely AI and machine learning changes on medicine. 4.0 focusses on the value chain of providing imaging services.
Dr Allen reminds us of the “Gamut’s of Radiology” a book that has been used to train residents and one I can recommend as you think of what features to explore and train your deep learning algorithms on medical data. Gamuts contains 4600 unique imaging findings , 13,000 unique conditions with an intresection of 57 000 findings and conditions. How do we prioritize useful features for patient care and radiology?
Looking at the scope of radiology where we work in different organ systems, different modalities and different findings (features), then the role of radiologists whether as service workers or knowledge experts is to develop and prioritize areas where true value is achieved to make sure man and machine win.
How is the ACR through the data science institute(DSI) doing this ?
The ACR DSI mission is summarized below
The DSI has selected various areas of expertise to work on to bring AI to improve patient care.
For example, through development of proper use cases, then there can be standardization of use cases where new developers can get started in AI from concept , validation and integration into clinical care, and post implementation monitoring.
Summary
You can listen to the archive of the journal club here : https://drive.google.com/file/d/1BkuT5DaoFs9G18oCs9MHETWKAqDCQuqx/view
We had 30 minutes dedicated to Q and A focussing on how radiologists can better prepare themselves for the future. Here are some of the top questions and comments discussed in the journal club.
- I think radiologists who refuse to learn about data science will be less relevant. A similar analogy is the cardiologists who refused to be trained in radiofrequency ablation in the 80’s. You can survive but just less relevant.
- Do i need to learn programming?
- No need as in future decades one will not need to program
- Health IT is much broader in scope than AI
- Thank you for the wonderful presentation. The TED talk discussed an instance where non-medical programmers started a medical venture successfully. Conversely, do you feel that AI enable non-programmers to play a larger role in developing software since AI is trained rather than programed? I want to play amore active role in radiology related AI but my programming skills not very strong.
- Will radiologists who fail to learn programing and AI basics become the radiologists of the 1990’s who failed to learn how to ready CT and MRI? - Could you please expand on the regulation of AI algorithms and how the FDA and ACR will ensure clinically robust algorithms?
- I’m reminded of the early days of PACS, around 20 years ago. We let the development get away from us, and so the vendors created products that appealed to IT and C-Suite types, and were not optimally designed for us. How do we get involved in AI so this doesn’t happen again?
- When one radiologist can read at the speed of 3 radiologists (because their accuracy is high, and measurement burdon is low) won’t the demand for radiologists go down?
- I think it is not “reading at the speed of 3 radiologists” but “reading at the accuracy of 3 radiologists combined” - Should we include AI basic training into Medical Education, radiology curriculum or CME?
- Despite the common use of that idea that we become the data scientists and information managers of medicine, should we be spening all our intellectual capital on the potentials of AI or should we be training radiologists to be more general informaticians, dealing with the less flashy but more functional IT functions like basic standards and interopeability and EHR function, etc.? It seems like a lot of other specialties are trying to capitalize on that role right now?
- As a medical student going into radiology, how would you prepare yourself to use this technology to boost your career rather than limit it? How would you position yourself during residency to be able to leverage this technology to its full potential? - It would be great to have an open source library of codes/applicable algorithms to get a way to apply it in any hospital, research, or improve the existing ones.
- How will smaller groups/private practices be included in the ACRs efforts in AI development?
- Also it would be great to get contacts in the AI companies if we want to annotate images.
- Money is always an important part of the equation and there are so many companies out there who intend to profit from this technology. How will the smaller radiology groups afford the price tag that this technology will have or will we see more smaller groups and private practice groups taken over by hospital systems that will be able to afford this technology? - What strategies should radiologists use to remain relevant as AI improves?
I am going to summarise my study plan below for getting up to speed in deep learning
My path
Background — Masters in health informatics , python and Java programming
Enrolled for Udacity nanodegree in deep learning (https://www.udacity.com/course/deep-learning-nanodegree-foundation--nd101) — Took a pause since it was too technical and did not fill the bigger picture problem for me . The course costs 600 dollars. For example , when i asked how to set the learning rate , the response i got was its random.
Enrolled as an international fellow in Fast AI course (ongoing — http://www.fast.ai .). There are archives of the course available for free here http://course.fast.ai
The course has a very active community at http://forums.fast.ai and uses the kaggle competitions for “homework” — encouraging you to try out more deep learning challenges using concepts learnt in class. I like this course because its pythonic — using pyTorch library but also explains multiple concepts like how to determine learning rates , dropout , differential learning rates etc.
I am doing the NIPS paper implementation challenge https://nurture.ai/nips-challenge/ to write code implementations of papers submitted at NIPS 2017 conference with mentorship support provided.
A colleague (Alex from Canada ) who took the past Fast AI course sent this to me
Other resources from other participants
1. Udemy. https://www.udemy.com/complete-python-bootcamp/learn/v4/ and https://www.udemy.com/python-for-data-science-and-machine-learning-bootcamp/learn/v4/
2. https://www.coursera.org/learn/machine-learning it uses octave, not python, but I much more prefer that language
3. ABPM — Clinical informatics fellowship programs open to all specialties https://www.amia.org/membership/academic-forum/clinical-informatics-fellowships
Next journal club will be in January 2018. We are looking for a mix of AI scientists to get involved so please connect with us if you are interested.
