MIDL 2018 Panel discussion about radiology and artificial intelligence
The international conference on Medical Imaging with Deep Learning meeting took place at the CASA Hotel in Amsterdam, 4–6th of July, 2018. The full 3-day program can be found on the meeting’s website. The meeting is mainly intended for deep learning researchers, health care companies and clinicians. From attending the meeting I could see that the deep learning researchers were in the vast majority. There were several oral sessions about new developments in deep learning, and many interesting posters were presented.
An interesting panel discussion about the influence of artificial intelligence on the radiology profession now and in the future took place on Friday,July 6th. Both the industry and academy were represented in the panel:
Bram van Ginneken (Radboud University, Nijmegen), Heinrich von Busch (Siemens Healthineers), Tim Salimans (Aidence / Open AI), Graham Taylor (University of Guelph), Hayit Greenspan (Tel Aviv University), Ron Summers (NIH) , Mathias Prokop (Radboud University, Nijmegen)
I made a concise summary of this interesting discussion.
Question 1: the famous statement of Geoffry Hinton about the future radiologists: is there only 5 years to go from now?
Prokop: radiology training asks 10–20 years to become a valuable clinical person. We will not need this in the future. The training will be different. There will be a point where the pattern recognition part is taken over by machines for a large percentage. Workloads are becoming too heavy. Radiologists are pattern recognition experts for large part of the day but they would like to get rid of that. They will have to reinvent themselves as radiologist
Greenspan: 80 % of the radiology job is detection and measurement, only 20% is other type of input in the diagnosis/reporting. Part of the work can be replaced
Taylor: Book recommendation: “Prediction Machines” . The book is a nice analysis on the general question of how AI will impact the nature of work: jobs will change, but there will be less job elimination then predeicted. But nevertheless training of radiologists will have to adapt to the nature of the work.
Salimans: Agrees with all other speakers: automation changes jobs, not always replaces them. Some radiologists are willing to collaborate and engage, others are scared. Dialogue should be openend, but progress is being made in that field
Von Busch: Workload in radiology rises faster than number of radiologists, AI should assist them in dealing with this. AI should assist radiologists in becoming doctors again
Question 2: In clinical practice we can’t see many radiologists using AI tools at this moment. Where are the tools now? Why aren’t they being used yet? Is the AI winter ont its way?
Von Busch: tools are now too limited/narrow. The tools address certain areas, there are still many gaps to fill, it is not reasonable to expect that everything will be there at once. More complex tasks will come, eg for radiotherapy planning
Salimans: AI is already providing advantages to Amazon, Google,… but in medicine there are many regulations and it’s going slower than in industry. It’s less easy to innovate in this field. All systems around it also have to adapt.
Taylor: ML services are becoming better on iPhone, which for most users this is very obvious and is going fast (internet giants are going fast and improving). It’s about time scale. Healthcare has a different time scale than commercial companies. It simply takes more time
Greenspan: we have to use data that we have, and those are still limited. We are generating algorithms but academic methodologies are not changing and research methodology in AI is not adapting to the classical research models in medicine. The companies should ask to the radiologist: “You as radiologist, what do you need?”. Tools that support radiologists need to be provided.
- we (AI industry) need to adapt research to real needs
- we have to talk about apps / packages adapted to clinical use cases
- not only accuracy counts, we also have to think about false positives etc. This asks for different thinking, designing, methodology (academic research vs industry research)
- each algorithm needs to be very strong, translation is needed through packaging them (multiple applications in one package)
Tim Leiner: some observations need to be made
- the companies will never take liability for their products and software, the radiologists will be indispensable in the final decisions that are made
- ethical part: how do we deal with conclusions made by AI algorithm? Will someone be willing to refute it?
- safety aspect needs to be addressed properly
Summers: AI systems can operatie on different scales. Concerns raise about the more advanced types of AI, will they decide about the treatment or diagnosis?
Physicians will always make decisions, AI algorithms will help physicians.
Question from public: we could turn around the question and ask if it’s it unethical to have a diagnosis from humans?
Prokop: How secure it this system in making a prediction? Data out of training set can be biased! If a doctor missed something he shouldn’t automatically be sued. The question needs to be asked if he should have seen it or not.
We need to know what is the best software before using it, and the best datasets need to be made available
Greenspan: Data need to be correct. We have less data than we would like to see, data problem is very big barrier to get going.
Salimans: Regulation agencies do ask to show that datasets work. The combination with human judgement remains necessary . The human needs to remain in the loop. There also is a factor of trust: explainable AI is necessary! How does it work?
Taylor: definition of explainability of software algorithms needed, it does not exist yet
Question 3: the best talents in Academy are bought by industry: is this true? Is this problem for Academy? Europe is not keeping up, vs. US and China Should be protect European research? How?
Taylor: academics can always work together with industry, no real choices need to be made
Greenspan: it is a problem, in the US the industry picks out good academic and highly rankend researchers. How do we do things jointly? Industry needs academics to find new talent, so industry should certainly support the academics (hiring more faculty, training, etc). Joint efforts are needed.
Von Busch: everyone likes to see good products, so industry needs talented people. Collaboration is the key, for both parties. The solution is to get companies more involved in education, to make them part of ecosystem (eg fund system courses), to create more flexible positions in the academic (combination of work). When students and postdocs leave academics there is a significant problem
Erik Ranschaert, radiologist, EuSoMII vice-president
