Nov 14, 2022
This week, please join authors Qiang Zhang and Matthew Burrage as well as Senior Associate Editor Victoria Delgado as they discuss the article "Artificial Intelligence for Contrast-free MRI: Scar Assessment in Myocardial Infarction Using Deep Learning-Based Virtual Native Enhancement."
Dr. Carolyn Lam:
Welcome to Circulation On the Run, your weekly podcast summary and backstage pass to the journal and its editors. We're your cohosts. I'm Dr. Carolyn Lam, associate editor from the National Heart Center and Duke National University of Singapore.
Dr. Peder Myhre:
And I'm Dr. Peder Myhre from University of Akershus University Hospital in Norway.
Dr. Carolyn Lam:
Peder, today's feature discussion is on AI for contrast-free MRI. Isn't that so cool, using AI to perhaps understand what we could see only with contrast, but now in a contrast-free manner. Now I know that sound a bit confusing, but I hope very, very enticing, because everyone's going to have to wait for a little while before we get to that interesting feature discussion. And for now, let's talk about some of the papers we have in today's issue, shall we?
Dr. Peder Myhre:
Yes, Carolyn, I can't wait for the feature discussion, but we're going to start with some of the other papers in this week's issue, and we're going to start in the world of preclinical science with a paper looking at human cardiac reprogramming, because Carolyn, direct cardiac reprogramming of fibroblasts into cardiomyocytes has emerged as one of the promising strategies to remuscularize the injured myocardium. Yet it is still insufficient to generate functional induced cardiomyocytes from human fibroblasts using conventional reprogramming cocktails and underlying molecular mechanisms are not really well understood.
Transcriptional factors often act in concert and form tightly controlled networks featuring with common targets among different transcriptional factors. Therefore, missing one component during heart development could lead to heart function defects and congenital heart disease. And in this study by corresponding author Yang Zhou from the University of Alabama at Birmingham, the authors perform transcriptomic comparison between human induced cardiomyocytes and functional cardiomyocytes to assess additional factors that govern transcriptional activation of gene programs associated with sarcomere contractility.
Dr. Carolyn Lam:
Wow. Really nicely explained. Thanks, Peder. So what did they find?
Dr. Peder Myhre:
So Carolyn, through these computational analysis of transcriptomic data, the authors identified TBX20 as the most under expressed transcription factor in human induced cardiomyocytes compared to endogenous cardiomyocytes. They also demonstrated that TBX20 enhances human cardiac reprogramming and improves contractility and mitochondrial function in the reprogrammed cardiomyocytes.
Dr. Carolyn Lam:
Nice. Could you summarize the clinical implications, please?
Dr. Peder Myhre:
Yes. So the clinical implications are that enhancing the efficiency and quality of direct cardiac reprogramming for human fibroblast is a critical step in the clinical translation of this technology, and better understanding of this synergistic regulation of key cardiac transcription factors during reprogramming will provide new insights into the genetic basis in normal and diseased hearts. Well, Carolyn, please tell me about your next paper.
Dr. Carolyn Lam:
Thanks, and we're moving now to kidney disease. Now end stage renal disease is associated with a high risk of cardiovascular events, but what about mild to moderate kidney dysfunction? Is it causally related to coronary heart disease and stroke? Well, today's authors give us a clue, and it's from corresponding author Dr. Di Angelantonio from University of Cambridge and colleagues who took a very unique combined approach to answer this question.
They first conducted observational analyses using individual level data from four huge population based data sources, namely the emerging risk factors collaboration, Epic CVD, Jillion Veteran Program and UK Biobank. Can you imagine this comprised almost 650,000 participants with no history of cardiovascular disease or diabetes at baseline, yielding almost 43,000 and 15,700 incident coronary heart disease and stroke events respectively during a 6.8 million person years of follow up.
So huge observational study, which they then followed with a Mendelian randomization analyses using a genetic risk score of 218 variants for GFR and involving participants in Epic CVD Million Veterans Program and the UK Biobank.
Dr. Peder Myhre:
Wow, Carolyn, this is a topic that I think many of us have really been wondering and thinking about. The mild to moderate kidney dysfunction, what does it really mean? And what a beautiful study to answer this. So what did they find?
Dr. Carolyn Lam:
First, there was a U-shaped association of creatinine-based GFR with coronary heart disease and stroke with higher risk in participants with GFR values below 60 or more than 105 mills per minute per 1.73 meters squared. Mendelian randomization analyses for coronary heart disease showed an association among participants with GFR below 60, but not for those with GFR above 105.
Results were not materially different after adjustment for traditional cardiovascular risk factors and the Mendelian randomization results for stroke were nonsignificant but broadly similar to those for coronary heart disease. So in summary, in people without manifest cardiovascular disease or diabetes, mild to moderate kidney dysfunction is causally related to the risk of coronary heart disease, highlighting the potential value of preventive approaches that preserve and modulate kidney function.
Dr. Peder Myhre:
Thank you, Carolyn, for such a great summary and an important result from that study. I'm going to now take us back to the world of preclinical science and talk about diabetic cardiomyopathy and exercise. And we both know that patients with diabetes are vulnerable to development of myocardial dysfunction, and that exercise, our favorite thing, for maintaining cardiovascular health, especially in patients with diabetes.
And despite a wealth of evidence supporting that cardiometabolic benefits of exercise, the precise exercise responsive signals that confer the beneficial effects of exercise in cardiomyocytes to remain poorly defined. And previous studies have identified fibroblast growth factor 21, FGF21, a peptide hormone with pleiotropic benefits on cardiometabolic hemostasis as an exercise responsive factor.
And in this study from Aimin Xu from the University of Hong Kong, the authors investigated a six-week exercise intervention program in FGF21 knockout mice and wild-type litter mates that all had diabetic cardiomyopathy induced by high fat diet and injection of streptozotocin.
Dr. Carolyn Lam:
Nice. So what did they find?
Dr. Peder Myhre:
Yeah, the authors found that exercise lowers circulating FGF21 levels, therefore remodeling the heart as an FGF21 sensitive target organ. And the protective effects of exercise against diabetic cardiomyopathy are therefore compromised in mice with deficiency of FGF21. They also identified Sirtuin-3 as an obligor downstream effector on FGF21, preserving mitochondrial integrity and cardiac function. Finally, the authors demonstrated that FGF21 induces Sirtuin-3 expression through AMPK-FOXO3 signaling access.
Dr. Carolyn Lam:
So could you put that together for us better? So what are the clinical implications?
Dr. Peder Myhre:
So the clinical implications from this paper is that circulating FGF21 is a potential biomarker for assessment of exercise efficacy in improving cardiac functions. And exercise is a potent FGF21 sensitizer in cardiomyocyte and has the potential to enhance the therapeutic benefits of FGF21 analogs in diabetic cardiomyopathy, and selective activation of FGF21 signal in cardiomyocytes may serve as exercise mimetics and represent a promising targeted intervention for precise management of diabetic cardiomyopathy.
Dr. Carolyn Lam:
Oh my goodness. That is fascinating. Thank you, Peder. Well let's wrap up with what else there is in today's issue. There's an On My Mind paper by Dr. Weir entitled, “The Emperor's New Clothes: Aren't We Just Treating Grades of Heart Failure with Reduced Ejection Fraction.”
Dr. Peder Myhre:
And there is a Research Letter by Dr. James Martin from Baylor College of Medicine entitled “Gene Therapy Knockdown of Hippo Signaling Resolves Arrhythmic Events in Pigs after Myocardial Infarction.”
Dr. Carolyn Lam:
Very nice. Thanks, Peder. So wow, let's go onto a featured discussion on AI for contrast-free MRI and a virtual native enhancement here coming right up.
Dr. Peder Myhre:
Awesome.
Dr. Carolyn Lam:
Now we all know that myocardial scar is currently assessed non-invasively using cardiac MRI with late gadolinium enhancement as what we would call the imaging gold standard. Wouldn't it be amazing to have a contrast-free approach, which could provide the same information with many advantages such as a faster or cheaper scan, and without contrast associated problems? Well guess what? We're about to discuss that today in a feature publication in today's issue, and I am so pleased to have the co first authors with us today. They are Dr. Qiang Zhang and Dr. Matthew Burridge, both from University of Oxford, and to discuss it as well, our senior associate editor, Dr. Victoria Delgado from Barcelona. So welcome, everyone.
Qiang Zhang, could I start with you and ask you, I understand you're a machine learning expert, which means you're probably smarter than all of us here. Could you maybe explain in simple terms what made you and Dr. Burridge do the study?
Dr. Qiang Zhang:
First? Thank you so much, Carolyn and Victoria, for the invitation. As you have mentioned, late gadolinium enhancement, or LGE, has been the imaging gold standard in clinical practice for myocardial catheterization including scar assessment for patients with myocardial infarction. However, LGE requires the injection for gadolinium contrast, and this is cautioned in some patient groups and increases the scan time and cost.
On the other hand, pre-contrast CMR such as Sydney T1-T2 mapping, a gadolinium-free alternative for myocardial catheterization. But their clinical use has been hindered by confounding factors and a lack of clear interpretation. So with our cross deceptor team at Oxford, we developed an artificial intelligence, virtual native enhancement technique VNE.
It can produce a sort of a virtual LGE image but without the need for gadolinium contrast. And we have previously tested it in patients with hypertrophic cardiomyopathy as published in this journal last year. And in this new study together with Matt here, we tested in patients with history of chronic or prior myocardial infarction.
Dr. Carolyn Lam:
Oh wow. Cool. So audience, you heard it. Instead of LGE, we now have VNE, virtual native enhancement. That's super cool. Thank you. Matt, could I bring you in here? So tell us a little bit more about the population you studied and what you both found.
Dr. Matthew Burrage:
Yeah, absolutely. And thank you so much for the invitation as well. So as Chang has said, this was a single sensor study that we performed at the University of Oxford and specifically targeting assessing myocardial scar in patients with a history of chronic or prior MI. So we had two sources for our population data. Well, first we used our real world clinical service data from our institution.
So we screened 11 years worth of patient data for presence of MI. So patients were included. There was a evidence of a previous MI based on an ischemic pattern of LGE, but we specifically excluded patients who had an acute presentation, or if there were features of acute MI on the CMR scan such as presence of myocardial edema or microvascular obstruction. The reason for this is we wanted to keep this as a clean population to avoid the potential confounding effects of myocardial edema or MVO on native T1 values. And so we also excluded other myocardial pathologies such as underlying cardiomyopathies and infiltrative diseases.
A second population dataset came from the OX Army study, which is a single center prospective study of patients presenting with acute MI. And for these patients we used their six month follow up scan to again avoid the confounding effects of edema and pathology. So overall we had a total of 912 patients who have contributed over 4,000 image data sets. The patient characteristics, 81% were male, they had a mean age of 64 years and there were cardiovascular risk factors such as diabetes melitis, hypertension, hypercholesterolemia in 20 to 40% of patients, while just over half had a history of previous revascularization.
We also separately applied the VNE technology to a pig model of myocardial infarction, which was thanks to our collaborator, Rohan Domakuma in the US. And so those were scans performed eight to nine weeks after an induced MI in the LAD territory in a series of pigs. And so this gave us the ability to provide a direct comparison between LGE, VNE, and histopathology in this model.
Dr. Carolyn Lam:
Wow. And results?
Dr. Matthew Burrage:
So what we found and the key results were firstly that VNE provided significantly better image quality than LGE, and this was on blinded analysis by five independent operators from our test data sets. Secondly, the VNE correlated strongly with LGE in terms of quantifying infarct size and the degree of transmurality, so the extent of the MIs in our test data set. We had pretty good overall accuracy of 84% for VNE in detecting scar compared to LGE with no false positive VNE cases.
And finally there was also excellent visuospatial agreement with the histopathology in the pig model of myocardial infarction. So really this, we think, is a technology that provides clinicians with images in a format that firstly they're familiar with, which looks like LGE, provides essentially the same information as LGE, but it can be achieved without the need for any gadolinium contrast agents and can be acquired in a fraction of the time.
So it takes less than one second to generate the VNE image. So as we've said before, we feel there's a lot of potential here for this technology to potentially eliminate the need for gadolinium contrast in a significant proportion of CMR scans, reduced scan times and costs, increased clinical throughput and hopefully improve the accessibility of CMR for patients in the near future.
Dr. Carolyn Lam:
Oh wow. That is tremendous. So first of all, congratulations to both of you. Before I ask Victoria for some thoughts, could I also just check with Qiang Zhang, because all AI algorithms need to be externally validated or surely there's some catch to it, or so-called limitations, or something else you may study. Could you maybe round up by saying is there anything that clinicians should not be applying it to or be aware of some limitations or?
Dr. Qiang Zhang:
Thank you, Carolyn. So a limitation of this study is that the dataset that is used for developing the models, the majority of them are patients around six month after the acute infarction. So where the myocardial infarction is still evolving, which may include residual edema and microvascular obstruction, and that is difficult to assess using the current VNE model.
And also we found it challenging to assess small sub endocardial infarction and actually to address those limitations, we are working on improving the VNE models, training it on even larger data sets and training it on LGE to detect small sub endocardial function. And we will further develop it to detect, for example, acute edema and a microvascular obstruction, and in the meantime develop quality control driven AI models to inform the clinical users of and unreliable results.
Dr. Carolyn Lam:
Wow, thank you. So Victoria, now I'm dying to hear your thoughts. How do you think this fits in the landscape of all AI imaging now?
Dr. Victoria Delgado:
I think that it's an excellent development and I congratulate the others for the article and the proof of concept that we can move away from the late enhancement and the use of gadolinium enhancement. I think that this is a major step forward because as Matt said, they are going to decrease very much the time of scanning and the post processing because is automatically done as far as I understand. So even if you can interpret yourself the amount of so-called virtual enhancement, the system gives you a value for that extension of the virtual in non-gadolinium enhancement. So that reduces very much the variability that can be in each observer if that is done automatically.
But my question to them is also if that can be influenced by the type of scanner that you use, for example on echocardiography, that's much more my field of interest, it depends very much sometimes how the images are processed of which are the vendors that we have used to acquire the images. Is this a limitation for your software? Can you foresee there some variability or is completely independent?
Dr. Qiang Zhang:
Thank you, Victoria. So we are aware of actually the difference of the data produced by different scan of vendors and the advantage of AI-driven methods is that it is data driven. So we plan to incorporate dataset from other vendors so that the trend that VNE models can work with like multiple scanner vendors. This actually will be done alongside the ongoing standardization program of T1 mapping in our group, which is the underpinned technology for VNE. And this is led by Professor Stephan Pitchnik and Vanessa Farrera. And we actually hope the VNE technology as AI driven methods could contribute to a solution to the CMO standardization between the scanner vendor.
Dr. Victoria Delgado:
And another question, if I may follow in this CMR, it has been proposed as a very valuable imaging technique to assess infarct size and to see the efficacy of some therapies to reduce the myocardial infarction size. How do you think that this new methods will impact in future trials and the way we have been interpreting the previous trials, like for example, the one that you use for the validation?
Dr. Matthew Burrage:
Yeah, thanks Victoria. It's a really, really excellent question. I think there's a lot of potential for the new VNE technology to also become a clinical endpoint in some of these trials in terms of reduction in infarct size, because the information that we get is more or less the same as we get from the LGE. So there's lots of potential that we can, again, use this as a biomarker in trials for looking at reduction in infarct size and reperfusion therapies. But it has the benefit that it can be done quicker and without gadolinium contrast.
Dr. Victoria Delgado:
This is amazing guideline and really I would have a lot of questions for them as well. And knowing the literature, for example, in the Scenic center in Madrid that they have been scanning the evolution of myocardial infarction from 0.02 weeks to see how this would translate with your technique. That will be amazing to understand how this can be done.
Dr. Carolyn Lam:
Oh wow, there you go. New research idea right there. Well how about if we end with a very quick question for each of the first authors. So maybe Matt, you could start, I mean is this ready for primetime and clinical use? And if it's not, what needs to be done to get there? In other words, where are you headed as the next step?
Dr. Matthew Burrage:
So again, thank you, Carolyn, that's a really excellent question and I think the next step before this becomes ready for primetime clinical use is validating this technology really across the spectrum of other myocardial pathologies. So the next work that we are developing this on is in patients with acute myocardial infarction, and then extending this to sort of acute inflammatory conditions like myocarditis, other non-ischemic cardiomyopathies, things like amyloidosis as well.
So this will be the next step into rollout and we are looking to track things like VNE burden and how that relates to clinical outcomes, similar to the previous LGE papers have done across different myocardial pathologies, but then ultimately aiming towards clinical rollout within the next few years.
Dr. Qiang Zhang:
Yeah, I think pretty much what Matt has said, we're going to develop the deep learning methods and test it further on pretty much the whole spectrum of commonly encountered diseases, and then more complex pathologies such as acute pathologies like edema, microvascular obstruction, and then we test on large population study like UK Biobank and other prospective clinical trials. And of course the most importantly is to roll out for real world clinical use. And as Matt said, we are aiming to do this within the next two to five years.
Dr. Carolyn Lam:
Wow, this is amazing. Both Victoria and I said thank you, congratulations on this landmark piece of work. Thank you for publishing it in circulation. Audience, thank you for joining us today from Greg, Peder, myself. You've been listening to Circulation on the Run, and don't forget to tune in again next week.
Dr. Greg Hundley:
This program is copyright of the American Heart Association 2022. The opinions expressed by speakers in this podcast are their own and not necessarily those of the editors or of the American Heart Association. For more, please visit ahajournals.org.