DON'T GO INTO RADIOLOGY - AI is Taking Over

An artificial intelligence tool that reads chest X-rays without oversight from a radiologist got regulatory clearance in the European Union last week — a first for a fully autonomous medical imaging AI and the company Oxipit. It’s a big milestone for AI and likely to be contentious, as radiologists have spent the last few years pushing back on efforts to fully automate parts of their job.

However, is this going to help or hurt radiologists? In this video, I talk about this article and then also discuss my thoughts on artificial intelligence taking over Radiology.

03:53 - What is Chestlink
05:56 - What does American College of Radiology think?
06:52 - Questions I have about AI
09:28 - Does this reduce a Radiologist’s workload?
12:32 - My Thoughts on AI in Radiology

Join this channel to get access to perks:

______
CONTACT ME:

📹 TikTok - @DrCellini
-----------
-----------
MY TOOLS & GEAR:

**As an Amazon Associate I earn commission with use of the above links on qualifying purchases**
-----------

OTHER STUFF:

*None of this is meant to be construed as investment advice, it's for entertainment purposes only. Links above include affiliate commission or referrals. I'm part of an affiliate network and I receive compensation from partnering websites. The video is accurate as of the posting date but may not be accurate in the future.

#radiology #artificialintelligence #AI

[naglisramanauskas]

Dear Dr. Cellini,

Naglis, Chief Medical Officer and researcher from Oxipit (also a radiologist) here, found your excellent channel, an honor to be featured!

First of all, thank you for an excellent video and your valid and reasonable points.
Let me provide some feedback:
1) Software made no (clinically significant) errors during the pilot phase. Your comment was correct that it does not list what was considered a non clinically significant error. Let me elaborate briefly about this: clinically significant errors are defined as errors which if unreported have the potential to cause patient harm. This would include everything that would be broadly considered clinically significant and/or actionable (nodules, pneumothoraces, consolidations, subtle minor nodular opacities, which cannot be excluded to be significant and many others, which are too many to be listed). Some findings such as degenerative changes of the thoracic spine, some age correlated sclerotic changes of the aortic arch and other similar ones, can be considered clinically insignificant and might be classified among the normal studies. There will be findings, which might depend on the institution if they are to be considered as significant or not, and this is identified during the initial retrospective and prospective validation stage in each institution. Any steps towards autonomy are only made once the radiologists from a specific institution are confident in the validation and safety of the solution.
2) Who is liable for the AI software? To answer this question shortly, once the software is running in an autonomous mode we aim to take the responsibility for the analysis of the studies, which were classified as normal by the software. You have correctly noted, that if that is not the case, there is marginal value to be gained for the radiologist, as at most this will save only a few seconds of time per study. Before we get to the autonomous stage and the final signature is still on the institution/radiologists side, the liability is the same as it is now - on the institution side. Having said that, there are questions which we cannot yet fully answer, we aim to be transparent about that. Even though we are already certified, there are national and international laws which are still to be worked out. CE certificate is not the end of this process - it is a benchmark which allows the next steps to be taken.
3) Will this actually help radiologists? How much it will help exactly depends highly on the institution and the market. A few examples: A primary care institution which makes 100 CXR studies per day, and lets say on average 80 of those studies are normal. Ultimately the software could potentially automate the reporting of ~35 of the studies, which could result in tangible resource/cost savings and allow the radiologists to spend more time on more interesting and challenging cases (either CXRs or other modalities). Also, there are institutions in some markets who already cannot report a significant fraction of their studies (both in western and the developing world countries), and some patients do not get a report for their chest X-ray either altogether, or during a useful amount of time. We believe automating even a sample of these studies can significantly contribute to improving patient outcomes.
4) How much does the software actually cost? Even though I cannot give a specific price per study (this depends on the market and the cost per CXR in a specific market) this is projected to cost less than the radiologist resources which would be required to report the automated studies in a regular way, with a significant margin.

Finally, in regards to your finishing notes about if the radiologists are at risk to lose a significant amount of work anytime soon: short answer - no. This is the first step to automate a fraction of the normal studies of one radiological modality, volumes of which are already too large for the current radiologist resources to properly manage. It is likely that somewhat soon we will see some similar examples in a couple of other modalities. However, I do not think that it is or will be possible to automate abnormal or borderlilne studies anytime soon, mostly because of factors not related to the AI performance, but because of radiological factors, such as the importance of clinical information/context and interreader subjectivity. For the obviously normal studies the effect of these factors can be minimized, but for abnormals they will remain very substantial. Therefore the large majority of the radiological studies will remain in the radiologist's hands for the foreseeable future, unless we figure out general AI soon, which will take all of our jobs and create an utopia on earth.

Thank you once again for the feature and if you would like to discuss further, we are open to your questions :)

[RiaanZoetmulder]

Hello Dr. Cellini,

As somebody pursuing a PhD in medical image analysis (the branch of AI that applies AI to problems in radiology), I agree with your points. Artificial intelligence is unlikely to replace radiologists any time soon and we could probably use more radiologists rather than fewer.

In my opinion, AI will not render radiologists obsolete for several reasons. First, the performance of the algorithms depends very much on the modality that is being used and the image analysis task that is being considered. For example, if you want to delineate thrombi on NCCT/CTA with an algorithm, this is a very difficult task due to the object being very small, image artefacts making the scan less clear and sometimes no hyperdense artery sign being present. If however, you want to delineate very large tumors in the brain on MR scans this is a very easy task because they are large and easily visible. The latter problem would be a fairly easy to publish research paper and would draw attention from the media, but not the former problem. So this creates the impression that AI is good at everything, even though it is not. Second, there is a problem in AI that is called the domain shift problem. To put this into a medical context, if you train on a dataset that was acquired using machine A and you report your results on data acquired on the same machine it will get a certain performance. If however, you change the machine that you use, it will get a lower performance due to the images looking slightly different. This would render some algorithms that work well on relevant problems useless until they are re-trained on data that includes scans made on this new machine. Third, the data leakage problem. Many early research applying AI (deep learning specifically) included people that did not have much experience working with medical data. These were people coming from computer vision where you work with individual images (most of the time) rather than patient data. As a result, specifically CT and MR scans were not correctly distributed to training, validation and testing sets. The correct way of doing so would be assigning patients to splits. What would happen was all slices of all of the the scans would be shuffled and assigned to splits. This means that highly correlated slices would be trained and tested on, leading to inflated results in some studies. Now to provide some context, I got this information from a colleague working in the same lab as I am and did not read the original paper, but I have seen this problem in a number of studies myself. Fortunately, this is improving with better practices and the field maturing more.

To end on a positive note, there are several areas that, in my opinion, AI will be very helpful for. A great example is research. If an image feature is labor intensive to annotate but could be a possible alternative endpoint AI can definitely be of use here, especially when you are working with large-scale clinical trials or registries. This makes research into the efficacy of medication and treatment much more feasible. Another application would be to allow for quantification of image features that inform treatment or are prognostic, but that are too labor intensive for a radiologist to do. Volume measurements are a good example. A third line of research that I find quite promising is that of image reconstruction. Reconstructing an MRI image from K-space is very time consuming. Thus, in addition to MRI having several contra-indications, acquisition and reconstruction time are limiting factors. There is research that is working on requiring less samples from K-space, which speeds up acquisition. Finally, AI can help make the image data more easily readable for a radiologist. For example by registering images more closely or extracting center-lines to "stretch" out arteries.

Hope you find my response helpful.

Kind regards,


Riaan

[christiancasteel]

EKG is a great counter point. I’m an Emergency Med resident and the machines are notoriously wrong, both with over reading and under reading. Unless it’s stone cold normal the machine read is useless.

[ristogrkovski]

And then in the fine print: "We (the super-mega-smart AI start-up) don't assume any risk or responsibility should misdiagnosis by our AI software lead to a serious health consequences. Every final clinical decision rests on physician's shoulders."

[lalishankar]

Excellent. As a radiologist for the last 40 years- I have been told radiologists will be replaced by other specialities and AI. Still practicing with no real fear. Your assessment of the situation is perfect. We need to be involved in the development of the AI technology for DI and in incorporating it into the radiology diagnosis algorithm. Great explanation given by Dr. Cellini.

[TheBlackMage]

I really wish AI companies would start focusing on making me more efficient as a starting point. For example, have the AI program pull out clinically relevant information from the chart or automatically measure a lesion that I select (which would also improve consistency). These steps would be a great spot to start and could also allow for data collection which would help further AI development in the future.

We use basic AI to diagnose bleeds, PEs, etc. It's definitely good for prioritizing workflow but it makes a lot of mistakes. Would be terrible to rely on it.

[alextgordon]

I'm not a radiologist, but a machine learning practitioner, so I'll give my perspective. The issue with radiology is that that problem is in principle an ideal "shape" to be tackled by machine learning. The job involves a large amount of transforming data from one form (imaging, patient data) into another form (language), and there's plenty of potential training data being collected every day.

In machine learning, progress is often sudden. A problem will go from "almost impossible" to "mostly solved" in a short period of time without much warning. For example, the first computer to win a professional Go match against a human: October 2015. The last human to win a professional Go match against a computer: March 2016. 6 months. A couple of years before, the conventional wisdom was that that this problem was at least a decade away from being solved.

If you are planning a career, not just in medicine but in any discipline, and you want to know if your career is at risk of major changes as a result of machine learning, it would be a mistake to look at how researchers are trying to solve that problem today. Few of the methods we use today will be in use 10 years from now, let alone 20. The primary architecture of modern language models was only introduced 5 years ago, for instance. Much of the language modelling research that was done >5 years ago is obsolete. The fact that a problem has not been solved yet means that the state-of-the-art is wrong. Not that the problem itself is insoluble.

Instead, you have to look at the fundamentals. How easy is it to train a model? How much interaction with the physical world is essential to the job? If much of the job involves doing things at a computer, the risk is increased.

Will human radiologists still exist in the future? Yes of course. Will ML models *one day* gain superiority over human radiologists at a majority of relevant tasks? Certainly. Will this shift occur within the careers of today's radiologists? It seems likely, given the rate of progress in related areas.

[mitchhennen]

I just wanted to say how much respect I have for IR doctors. I do have Crohn’s disease and I became very sick last week. I’ve had a non healing abdominal fistula for a year now. I’ve had 3 drains put in the past year. Last week one of my old drain spots started draining and along with one of my fistula holes on my stomach. So I went to hospital and they found a huge pocket of infected fluid along with stool from the fistula being connected to my small bowel. I also have an ileostomy. IR did place another drain in me but I’m going to need surgery to fix this fistula. But the IR team were amazing. I really appreciate what you guys do.

[jn-lucr.]

I once heard somewhere that AI will never be autonomous because it'd be very difficult to sue when mistakes happen... and we love to sue.

[LJStability]

Also, you are 100% right about the EKG software. Our department chair in radiology told us that studies at our hospital showed that the software is wrong like 50% of the time. In other words, never follow what the program says. Read it yourself.

[chiaroscuro]

As someone applying to Med school next year and looking to Radiology, I agree that I am more excited about AI than worried. Additionally, the whole “AI taking over” really only would apply to a pure Diagnostic Radiologist, as like you mentioned as an Interventional Rad you do much more than reading scans. Great vid on the topic :)

[Roghany]

love to see AI take the core exam before being allowed to practice

[guillaumel]

Hi Michael! It's the same with CPAs! I think AI and advanced technology may have an impact for accounting technicians in the future, loosing their job?? We will see!

[luis.steinhorst]

In the past, when computed tomography was the big new step, they sayed that the radiologist would no longer be necessary, after all, they thought it was possible to see the inside of the patient. What happened was the opposite. The radiologist now has a great new method, greatly increasing his work and relevance.

[S-tcnt]

AI will "replace" Radiologist the way calculators "replaced" mathematicians. Technology amplifies human output.

Ironically, when the x-ray itself was invented, "people" thought it would replace Physicians entirely. Why do we need a Physician when we can just look at these images and know exactly what the problem is? It turns out x-rays and other forms of medical imaging vastly expanded our understanding of how little we know, and the need for well trained Physicians actually expanded due to the need to interpret these ever complicating technologies.

AI will likely be implemented in Physician training as an augmenting instrument in delivering patient care.

[sajjadmalik]

As a pathologist, we are beginning to see the incorporation of AI into daily practice…it’s very limited right now but I’m confident it will explode over the next 5 years. It will revolutionize pathology and will probably make us more efficient. I wouldn’t trivialize how normal chest X-rays can be read by AI algorithms because it’s just the start. My philosophy is we have to work with them instead of against them….our jobs aren’t necessarily at risk but they will evolve. Just my two cents!

[SlimShady]

Could you make a video going through some radiology cases you found interesting or scans that are complex so we can see your thought process on these scans

[sardissozo]

I would make 2 arguments for AI. 1. The double check is awesome. While you suggested it's a doubling of work, it's also a bit of a failsafe. Redundancy in this field is beneficial. 2. AI could be used to prioritize reading. It's like a computer based triage. Get the radiologist looking at likely abnormal images first. Delaying care for 80 no-finding reads seems unnecessarily risky. I work in a Cardiac Cath Lab but my most significant experience was as a patient post-covid... An emergency room Doctor started treating me for blood clots and DVT's before any radiologist saw anything from my medical record because it was so obvious when the images were in... I can imagine in some places that wouldn't have happened.

[boahgeil]

Im a data scientist (the people developing AI software) and play around with the idea to go back to university and study medicine. And the most interesting area is actually radiology to me :)

[BeyondDentistry]

As a dentist studying computer science, I am fascinated by this topic; genuinely hoping to see AI applied to orthodontic treatment planning. There is so much good data now with many orthodontist taking pre and post tx CBCT.