Tag: artificial intelligence

Posted on by

All The Hype About Artificial Intelligence Products Versus What Really Happened!

artificial intelligence products

Flashback to 2018 and 2019, and you can read some of my posts about the RSNA’s latest and greatest artificial intelligence products. So, what percentage of those products has your hospital or imaging center incorporated into their workflow? For us, I can tell you that number is exceedingly low. And, I am willing to bet that most of you have a similar story. So, I figured it would be fun to go through some of the promises that silicon valley has made over the past several years versus what has come to fruition in daily practice. Let’s go over their overpromises and underdelivering. It would be fun to do this every few years to check up on the progression of AI technology in Radiology!

Reading Mammo With Half The Amount of Readers

Ironically, if you have seen any decrease in the number of breast studies coming to your institution, it has far less to do with artificial intelligence. Instead, it is probably related to Covid! Nevertheless, most of the work has returned. And, I don’t know of any institutions that are using artificial intelligence to replace the initial screening reads for mammograms. (as enticing as that sounds!) Most places have the hebeejeebees for ethically, legally, and financially replacing a mammo reader with a computer!

Workflow Will Be Seamlessly Integrated

I was hoping this one would have taken place by now. But to no avail. Yes, we will be getting another EHR/RIS system to replace the one that we have right now. But rumor has it that, although better, it is nowhere near seamless. I am still waiting for the day that I pull up a case, and my computer instantly opens up a case, the appropriate priors, the relevant labs, a brief pertinent history, an internet blurb on the disease entity, and the patient’s most recent surgeries without having to click a button. I believe the day will come. But, I’m not sure it will be there during my working lifetime!

Radiologist Will Have No Role In Reading Bone Ages

This concept makes a lot of sense. AI should read cases with a low likelihood of lawsuits and unlikely dire outcomes. What better study for artificial intelligence to read than a bone age? It certainly meets those criteria. Furthermore, we analyze and match up features of hands with features similar to standard cases. This process should be easy peasy chicken squeezy. (Maybe in my dreams!) Well, I am still waiting for my institution to incorporate this incredible technology!

Dictation Will Be Entirely Automated Into Standard Reports

If I had a dime for every time a company would say, your reports would be so much easier without our technology; I would be a veritable gazillionaire. Of course, they will standardize everything. And, with one button click, the clinician will be able to localize your disease pathology on a film. Where is this technology? Certainly not at our institution. (And, probably not at yours either!)

CAD Artificial Intelligence Products For Mammography Will Work Well Much Better!

Maybe, CAD detection has improved. But, I do not notice it one bit at the institutions I work. For me, it seems like the same old random placement of circles and stars to match supposed masses and calcifications. Rarely (if ever) has it noticeably helped me. And it does not seem to have changed much. Heck, but what do I know?

Artificial Intelligence Products Will Help With Diagnosis On Chest X-Rays

I saw some tremendous potential technologies at the RSNA to help make multiple diagnoses on chest films. It would issue a probability here and there for different disease entities. Well, I have not heard a whisper of this program coming to our institution any time soon. And, I have a sneaking suspicion, you will not see at yours either.

Improved Triage

Finally, I have heard of computer programs that will pre-search for life-altering diagnoses such as intracranial bleeds so that it will draw your attention to these cases first. I would love a program like that, and the technology should not be too advanced. But, I am still waiting and waiting and waiting and waiting…

Still Waiting For These Great Artificial Intelligence Products!

So, where does all of this leave now? I would have to say right back where we started. We have not seen that much yet except for some behind-the-scenes CT and PET-CT image improvement. Let’s do another checkup every once in a while. I have a feeling, though, these products will take a lot longer than anyone initially expected!

Posted on by

2019 RSNA – All About Artificial Intelligence, Phase 2

RSNA

What is it that I love about the RSNA? Well, it’s the only conference out there that I have found that can give you a flavor of the direction that radiology is moving. It’s where you can see the newest trends from vendors, educators, and researchers alike. Everyone gathers in one place, from all over the world. And, therefore, it gives you the Zeitgeist (I love that word!) of the state of radiology. Naturally, the only downside is the size of the conference. There’s so much going on that you can get lost in the shuffle if you don’t make any plans to know what you are attending beforehand. (Which I certainly did before I came!) So, according to what I saw at the 2019 RSNA, let me synthesize what is happening out there!

To give a brief answer, may I repeat the following phrase: artificial intelligence, artificial intelligence, artificial intelligence. To prove that point, for the first time, the 2019 RSNA dedicated an entire tech floor to these businesses (although it was a little off the beaten path of the main floor!) And, in the interim, radiologists, researchers, engineers, and large companies hosted numerous conferences and speaking events.

The Real Zeitgeist of 2019 RSNA

So what has changed from last year to this one? Well, first and foremost, the speakers were no longer trying to convince us that radiology is going to replace our jobs. That approach was so last year! Instead, it seemed that everyone already knows that artificial intelligence will become more like assistance devices for the radiologist. Whether it be data integration, automated detection, triage, or report formation, the nuts and bolts of artificial intelligence now assume a much more benign path that will ingratiate the radiologist’s whims. No more terminator bots to destroy radiology!

Confirming this notion, interestingly enough, for all the hype and bluster, few applications are ready for prime time. And even more, most applications are not even close to FDA approval. But, I will talk about some of the apps that will eventually become day-to-day tools that have the potential to become ubiquitous and readily available to radiologists. Moreover, I will discuss some others that just got my attention (for better or for worse!) Here were some of my favorite discussions during the conference.

Artificial Intelligence Technologies

Watson- All About Integration

Now, if IBM could swing it, Watson has the potential to be the best of all technologies coming down the pike. From my perspective, they have one of the most useful approaches to artificial intelligence for the radiologist. So, what will Watson eventually do? Well, it’s attempting to satisfy the dream of all us. It will take all the patient history, labs, progress notes, priors, and other tidbits of information that become useful, even data about the patient’s primary disease entity itself. And then, Watson will integrate all the relevant data buried in the digital world on any imaging case and display it in a readable format for the radiologist.

If successful, this technology can be a game-changer. But, it depends on the ability to sift through immense amounts of information in RIS, PACS, and EHR systems, among other individual databases in any given hospital. I am most excited about this technology because it will render our interpretations so much more useful. I am sick of the irrelevant one-word histories that we often receive!

Mammo Dreams

Mammography also is a primary target on the radar in radiology. Loads of lecturers were coming up with ways to incorporate some of the technologies. Out of the ones that I heard, one of the applications would screen all the mammograms and officially read about a quarter or third of the mammograms that were stone cold (Steven Austin) normal. According to the radiology research, AI could achieve 100 percent specificity for a negative study in this percentage of cases without the input of the radiologist.

Now, I loved the idea of decreasing a radiologist’s mammography workload. But, they were looking at cases numbering in the thousands. Let’s say you have a million cases. Would you also have 100 percent specificity? That remains to be seen. And, I don’t know if any company will be able to take on that liability in our litigious environment. Scary, to say the least. These companies may want to think twice about that ramification!

Low Liability Products

Lower liability AI products will be in the cards for the more immediate future for radiology. Whether it be bone age, triage, improvement of image quality, reconstruction assistance, or improved CAD, these foci are the targeted products that we will see first. Although most products are under the radar or not in current use in radiology departments throughout the country, I think we will see them incorporated over the next five years. And I am looking forward to seeing their results!

What Artificial Intelligence Products Will Fail In The Short Term?

As I roamed through the AI floor, I realized that lots of products offered detection with probabilities of diagnosis. For instance, I saw a chest x-ray diagnosis booth. And, their artificial intelligence product showed the abnormality along with tons of percentages for the likelihood of diseases. At least, in the United States, I don’t see much of a role in this technology. In those places with a lack of a radiology workforce (third world countries), it may take on a different relevance. But, lots of these technologies have limited applicability to the current status of the field. And, I don’t think they are anywhere near prime time.

My Final Take On This All From The 2019 RSNA!

Slowly, under the radar, we are beginning to see some of the fruition of the promises that artificial intelligence has made. And some companies are beginning to incorporate these more focused technologies into the hardware and software that imaging centers are buying. But, we are a bit farther away from seeing the explosive changes that AI potentially can offer. Whether it be true integration, mammography reads, and more, unfortunately, we are not quite there yet. Let’s continue to keep a watch and revisit the changes. Until next year at the RSNA!

 

 

 

Posted on by

Does Artificial Intelligence Spell The End For Radiology?

artificial intelligence

Question About Artificial Intelligence:

Greetings, could you elaborate on these words ”(Artificial intelligence) will profoundly affect all of our careers, for better or FOR WORSE” and ”have more to gain and MORE TO LOSE”.

I am asking because, in the above text, you’ve written only about the good things about AI, while with these words, you’re also implying bad things about it, but I, as a reader, don’t know about them as you haven’t listed them.

I am a doctor from Europe whose first specialty choice is radiology, but this artificial intelligence surge is making me think twice about it. Everything I read, including your piece, is a 2-way street ala. ”AI is great, but you must adapt to it.”. The end. Could somebody please tell me HOW I will have to adapt and what the BAD things about AI in the radiology field are? It’s freaking me out! Radiology, as it is now, is a fantastic specialty, but I don’t want to be jobless and incompetent 10,15,20 years from now. It’s a life’s decision, and I have exactly ten days to decide!!

Thanks,

Worried Applicant

 ————————————————————–

Answer:

You are not alone in worrying about the future of radiology and AI. However, after attending the RSNA meeting and talking to colleagues, AI will not take over a radiologist’s job entirely for a long time (if ever). That aside, AI technology may allow fewer radiologists to do the same amount of work that we do right now. Improving triage, artifacts, and integration will make the radiologist’s job easier.

AI Will Not Take Over The World!

Why do I say this and not worry about AI taking over the world? First, the ability of an algorithm to detect something is only as good as the programmer, the number of data points, and the quality of the data. However, programmers have not optimized the algorithms. The data points are too few. And the quality of the data is not uniform. So, I don’t believe that will happen for many, many years from now.

Moreover, deep learning algorithms still have difficulty distinguishing simple solitary findings on a plain film, such as pneumothorax (often mistaken for chest tubes), let alone all the findings on a chest film. Therefore, I don’t believe the interpreting programs can independently function.

More importantly, companies will not want to accept the consequences of the liability of missing findings on films that go unchecked by a radiologist. So, I see AI as more of a team effort instead of a radical upheaval of all radiologist’s jobs. Let’s spread the liability risk!

What Is The Real Downside Of AI?

With the advent of any new technology, we will see our fair share of crashes, bugs, and technical problems. So, I believe that these would be the main downside. But I think the downside is reasonably limited overall. My advice- if you like radiology, you should go for it. If I were deciding on a profession today, I would not let my fears of AI dissuade me from choosing the radiological field.

My two cents,

Barry Julius, MD

Posted on by

RSNA 2018 Meeting: What Residents Should Expect From Artificial Intelligence?

By far, the RSNA is the largest and most publicized radiology meeting of the year. But, I usually attend other meetings instead because so much goes on at once at the RSNA that it is next to impossible to concentrate on one area. Moreover, to get from one side of the Mccormick conference center to the other takes almost 10-15 minutes!

In any case, for the first time in eight years, I bit the bullet and decided to attend the RSNA meeting this year. Partly, I needed additional CME credits, but also I was not able to participate in the SNMMI meeting due to timing. But, I am delighted that I did. Why? It allowed me to grasp the most current themes in radiology that exist today. And, for residents, in particular, I thought it was critical to share with all of you what may be coming down the pike.

To start with, if I had to give one overarching theme from the conference, I would have to say that the central idea was artificial intelligence. Some of these revelations about artificial intelligence were not all that critical. But, others will play an enormous role in your careers down the road. So, I will try to emphasize those items from the conference that will undoubtedly influence your career. And, I will briefly talk about a few issues that the AI companies and academic sorts may overhype.

Strong AI Career Influences

Integration

When you pull up your PACS system to read cases ten or twenty years down the road, no longer will you have to pull up your history, labs, pathology, priors, EHR, and films separately. Instead, all hospitals and outpatient offices will have software and systems that will allow you to sort through all the information at once. Right now, some institutions are more integrated than others. For most of you, lack of integration this will become a relic of the past. Walking through the technical exhibits, you could see many solutions today that will allow the radiologists of the future to read films with all the clinical information at your fingertips.

Triage

Imagine having a helper sort through films to determine which ones you should look at first and others that can wait a bit. Well, now they have multiple software packages that use deep learning to create work lists that make sense. And other programs try to detect STAT findings such as brain bleeds to make sure that radiologists read these studies first. Finally, other software programs can make sure that the correct radiologists are reading the appropriate studies. Right now, most practices do not have the staff to scrutinize cases before dictation. So, all these AI solutions, will allow more efficient and appropriate reading of STAT and essential studies.

Reducing noise

Having stopped at numerous vendors, I noticed that most of the big ones were touting deep learning algorithms to increase the quality of images. What do I mean by that? Many had sophisticated programs that mitigated artifacts and increased conspicuity of lesions and vessels. Some allowed you to image patients with significantly lower contrast dosage to prevent acute renal failure. Motion artifact on a CT scan or PET-CT scan may become a rarity. The future in this arena is now!

Increasing Reading Efficiency And Quality

Right now, some companies have created Computer-Aided Detection (CAD) packages that assist the radiologist in reading images. At the meeting, these solutions seemed to emphasize lung nodules and mammography.  I would expect some improvement over the coming years in these imaging modalities. And, I think we will begin to see other imaging modalities that utilize CAD. CAD will continue to reduce the time and effort that goes into reading studies.

One of the new types of CAD that I thought would be of help to the average radiologist was a bone age reader. It’s the perfect place for AI to begin because medical liability is a bit lower.

Additionally, new software packages can integrate CAD functions into the current dictation and PACS systems. We will see a lot more integration to improve radiologist reading efficiency.

Weaker AI Career Influences

Radiology 3.0

As much as the RSNA academics liked to state that we will no longer be image-centric and instead become patient-centric, I don’t see many powerful economic and political factors to drive the current radiology business in that direction. Currently, I am a bit skeptical about the rate of progress toward that goal. I have a feeling we will still have considerable time pressures to get tons of cases out rapidly.  Until fee for service no longer becomes relevant, radiologists will not have the time to see each patient after reading their chest film. It’s just not realistic. However, we will have more information at our fingertips about our patients’ care to make better reports and decisions. But seeing a patient after reading each film is a pipe dream.

Driving Direct Patient Care

In one of the plenary sessions, a computer scientist gave a whole lecture on improving metrics such as hand washing and patient falls with artificial intelligence. She discussed placing sensors all around the hospital to create a virtual environment that can sense these events to improve patient morbidity and mortality. While I agree that we should try to improve these issues since they cause harm to patients, the lecturer did not convince me that hospitals and institutions are ready to spend the money and time to accomplish these goals. For the foreseeable future, I see too many financial and legal hurdles to extrapolate these ideas to a larger scale.

Artificial Intelligence And The RSNA- Final Take Home Messages

Artificial intelligence will have a profound effect upon all of our careers, for better or for worse. But, the younger generations of radiologists have more to gain and more to lose. Therefore, for residents, especially, it is critical to follow the developments within the field. And, the RSNA meeting is just the right place to get a sense of AI and your future. If you have an opportunity to attend a meeting like the RSNA, it is well worth it. Take advantage of the event and learn about how the main themes will affect your career!

Posted on by

Artificial Intelligence And Radiology Voice Recognition Technology: What Can We Expect?

Do you get this irony? We hear so much lately about artificial intelligence and how it can potentially affect radiology. But, for all this talk about the application of artificial intelligence, I have heard barely a squeak on anything tangible about applying artificial intelligence to real-world voice recognition technology. Why do I find this so strange? Startup companies espouse artificial intelligence for so many applications, some with questionable benefit. Yet, sitting right in front of everyone’s face is the most obvious work efficiency improvement, the application of artificial intelligence to enhance voice recognition. It is an area that desperately needs attention!

To me, it makes no sense that companies do not pursue this avenue. Unlike other health applications, applying artificial intelligence to voice recognition technology will unlikely result in lawsuits or untoward health effects (unless the AI switches rights with lefts or unwittingly adds a lot of nos to our dictations!) And, voice recognition is exactly the type of technology that fits the paradigm of why developers construct artificial intelligence. Everyone’s voice is different and we all choose different words to express ourselves. So, a technology like artificial intelligence that learns the subtleties of each of our voices and vocabulary should really make a difference in daily work life. So, why don’t we hear about breakthroughs on the voice recognition front? Let’s take a look at what’s out there already…

My Internet Literature Search

Since so much potential exists for the intersection of AI and voice recognition, I started a simple internet search on this topic. And, guess what? This is the first article I found. Microsoft announced a milestone. The company’s most accurate artificial intelligence enhanced software reached an error rate for transcription of conversational speech measuring 5.1%. (1)

Next, I found another article from Inc. that talks about the world’s most accurate voice recognition technologies. The top three are as follows: Baidu, Hound, and Siri. For those of you that do not know these enterprises well, I will briefly discuss each of them.

First of all, Baidu… Baidu is a Chinese company similar to Google but made for China. Why is this needed the most? Well, think about typing in Mandarin and how long it takes to type. In Mandarin, it is much shorter to speak than to write. So, that makes sense. Second, Hound… Honestly, I had never heard of this enterprise prior to writing this article. Apparently, it was a first comer in the voice recognition personal assistant realm and is a fairly accurate digital assistant. And lastly, of course, is Siri by Apple… To say the least from my experience, if this technology is considered to the be one of the world’s most accurate, artificial intelligence voice recognition does not even come close to where it should be. I can’t tell you how many times Siri interprets my language incorrectly! (2)

What’s In Store For Radiology Voice Recognition?

Now, call me crazy… But, none of these technologies sound so great to me. If a speech recognition system gets approximately 1 out of every 20 words wrong as in each of these technologies, that could be a recipe for disaster in the world of radiology reporting. And, this is the best that artificial intelligence offers for voice recognition?

In addition to these “seminal” articles, I did find an interesting merger between the ACR and Nuance Communications to set up a collaborative effort to improve radiology reporting. (3) But, nothing tangible has yet been created to significantly improve voice recognition technology. It’s all in the initial phase. This leads me to believe there is a long way to go.

Final Thoughts

Sorry to break the news but… I don’t see any significant improvement in the quality of our radiology dictation software technology for a long time. So, until artificial intelligence software developers take voice recognition technology seriously and apply their talents to this area, change will not be around the corner. Therefore, continue to check your work many times over and dictate cautiously!

(1) https://techcrunch.com/2017/08/20/microsofts-speech-recognition-system-hits-a-new-accuracy-milestone/

(2) https://www.inc.com/kevin-j-ryan/internet-trends-7-most-accurate-word-recognition-platforms.html

(3) https://www.nuance.com/about-us/newsroom/press-releases/philips-and-nuance-bring-ai-into-radiology-reporting.html

Posted on by

Should Artificial Intelligence Be Feared Or Welcomed?

artificial intelligence

Question:

Hello!

My name is Yasmin Amer, and I’m a producer for WBUR in Boston. I’m working on a segment about machine learning and medicine, and, of course, radiology is part of that discussion. I spoke to a local doctor and machine learning specialist who says artificial intelligence will make the field more exciting. Is this the attitude of many med students and residents interested in radiology? Are they primarily excited about tech in radiology, or is there any nervousness there? I’m happy I came across this blog – I would love your input.

thank you,

Yasmin Ameren

—————————————————————————

Answer To The Artificial Intelligence Question:

Yasmin,

Speaking to my residents about the topic, several of them fear the onset of artificial intelligence and its effect on radiology. Therefore, some residents have decided to go into “hands-on” fields like interventional radiology and breast imaging.

However, most others have responded they don’t see how a machine can synthesize the context of a case, the images, and all the patient-related factors to arrive at a final impression that tailors itself directly to a patient. Let me give you an example in the next paragraph.

Sometimes, two similar ultrasound findings can lead to entirely different management scenarios on breast ultrasound. An MRI may be the most appropriate for a noncompliant patient with multiple slightly complex cysts instead of serial follow-up ultrasounds. On the other hand, in a low-risk patient with the same cysts, the most appropriate conclusion may be to follow them every six months. These are slightly different patients with the same images. How would the artificial intelligence judge who is noncompliant? So, it takes more than just pattern recognition to process the information and arrive at a viable conclusion for an individual patient. I don’t think we are quite there yet.

Then, legal barriers prevent easy entry into the independent practice of radiology. Are large companies going to take responsibility if the machines make mistakes? Billions of dollars of losses are potentially at stake.

It is also interesting that applications to the radiology field have dramatically increased over the past few years. Improvement of the job market right now likely contributes to the increasing desirability of radiology. But that cannot be all. If applicants thought artificial intelligence would rob residents of their future 25-30-year radiology careers, we would not receive so many applications for radiological residency programs.

Long story short. Some fears of the unknown consequences of artificial intelligence exist. Overwhelmingly, however, I believe most resident concerns of artificial intelligence encroaching upon the radiologist’s work are less than the expected barriers to independent widespread implementation without supervision by a radiologist.

I hope that helps,

Barry Julius, MD

Posted on by

Has Technology Ruined Your Chance Of Employment In Radiology?

Has Technology Ruined Your Chances of Employment in Radiology?

Among the many benefits of living in the Computer Age are the rapid technological advancements that continue to bring convenience and joy to our daily lives. From handheld devices with 24/7 internet access to cars that drive themselves, the future many hoped for (and dreamed of) is happening right now. But while the positive aspects of new technologies mostly outweigh the negatives, disruptive change naturally creates both winners and losers, particularly on the employment front. The medical field is not immune to this phenomenon.

In the recent past, victims of technological encroachment tended to be lower skilled workers whose roles could be easily automated. Today however, potential job automation targets include professionals in high-skill fields ranging from law to engineering to medicine. In short, automation is now “blind to the color of your collar”, according to Jerry Kaplan, author of “Humans Need Not Apply”, (https://www.amazon.com/Humans-Need-Not-Apply-Intelligence/dp/0300213557) a sobering book that sheds light on the uncertain future facing modern workforces.

All of this is a roundabout way of asking a very uncomfortable question: Are robots coming for your radiology job?

The short answer is no…but don’t let your guard down. Here’s why.

Today the poster child of artificial intelligence (AI), IBM’s “Watson”, can already find clots in pulmonary arteries. And unlike a busy radiologist who might read 20,000 or so studies per year, Watson is on target to review 30 billion medical images (http://www.medscape.com/viewarticle/863127) It goes without saying that Watson’s only going to get better.

What’s more, a number of Silicon Valley startups are currently applying new technologies to automate and improve the delivery of medicine. One firm in particular, Enlitic, is even developing a deep-learning system that uses AI to analyze X-ray and CT scans. According to an article in the Economist, (http://www.economist.com/news/special-report/21700758-will-smarter-machines-cause-mass-unemployment-automation-and-anxiety) Enlitic’s system has performed 50% better in tests than a group of three expert radiologists at classifying malignant tumors. When used to examine X-rays, their deep-learning system also significantly outperformed human experts. Of course, this emerging technology leaves much to be desired in the bedside manner department, but that’s what robot doctors (http://www.techtimes.com/articles/131870/20160209/will-robots-in-healthcare-make-doctors-obsolete.htm) are for.

Now before you go and trade your radiology degree for a barista outfit, consider the fact that according to most experts, including the CEO of Elitic himself (Igor Barani, MD, a radiation oncologist), artificial intelligence and radiologists aren’t diametrically opposed. In fact, they’re largely symbiotic. By design, AI will increasingly free radiologists from mundane tasks that can be automated, like reviewing CT scans for lung nodules. As Barani puts it, “tasks that can be automated should be given to the machine—not as surrender but secession.” This outlook portends a future in which radiologists are increasingly empowered to deliver better patient care, not supplanted by robotic overlords.

Regardless of what technology naysayers say, there will always be radiology careers for talented individuals (http://scpmgphysiciancareers.com/) to pursue. That being said, the role of radiologists will almost certainly narrow in the coming years and decades to one of inference, not detection — and that’s an important takeaway. With little doubt, the medical field will require fewer radiologists per capita because of deep learning technologies that simply do a better job of identifying anomalies. The successful radiologists of tomorrow will be the ones who can reduce AI-generated data into useful information that helps patients get better, faster. That’s not a future to be scared of; it’s one all current and prospective radiologists should eagerly anticipate.