 So our second speaker is Dr. Stefan Schmitz-Valkenberg, who is incidentally, he said incidentally, also moved to the Moran Eye Center from Bonn in Germany. In Bonn, he was the assistant medical director of the Department of Ophthalmology, where he specialized and still specializes in clinical and surgical treatment of macular retinal diseases. He also co-founded and directed the reading center there. Which so, and the Bonn Reading Center was one of the major, not the major reading center in Europe. He's an expert in retinal imaging and as director of the reading center, he's developed and implemented many standard methods for people to grade images. He really is at the forefront of AMD research and he's widely recognized as a major player in the field. He currently holds the John A. Huntsman, sorry, John Huntsman's chair, and he moved to Utah to lead U-Read, which will be the Moran's own reading center. Thank you very much, Musa. Good morning, Moran, dear Randy, dear colleagues, dear ladies and gentlemen, before I continue with the topic AMD Revisited, I would like to take the opportunity to thank all of you for the very warm welcome here in the States, in Utah and Salt Lake City, and particularly in the Moran. Monica, the two boys and myself, we are really enjoying our new life right now. It's a wonderful experience and saying this, I also have to mention that the second half of last year was quite tough for us. We were separated and these are some impressions of our two boys separated from their mother, just with their father and he's trying to get them used to the US-American culture. So in October, I brought them over to Salt Lake City in the custody of their mother. I think that a great time, my time was not so good, I was pretty alone for almost three months in Germany, and these are some impressions from our fair wilderness with my team, with my post-graduates and Bon and the team of the reading center. Saying this, today it's a great pleasure for Monica and myself to be here at Grand Rounds and present, and actually we had a great discussion at home, what should we present? And there were lots of topics on the table, and one of them is, which was abandoned very fast, was that I should maybe represent some surgical cases from my collection. And preparing for this talk, I thought I should show at least one surgical video and this is this retracting me of a giant rental tear, which went quite good. And at the end, when I looked into the eye, and I have to say I didn't do this by purpose and this was just as nature gave me the case, I saw this laser at the edge of the tour and it was actually the first letter of the first name of my wife. So these are my disclosures. And when we talk about AMD, I think we have to mention drusen because drusen are the hallmark of the disease because drusen, they lead to atrophy and atrophy is a cause of visual loss in this patient, the most common cause. And today with multimodal retinal imaging, we can better monitor the evolution of the disease from drusen to atrophy. Terms like nascent geographic atrophy have been introduced. Different patterns leading to atrophy have been described. So multimodal retinal imaging, and this applies of course not only to AMD, allows us for the non-invasive mapping of disease evolution at higher resolution. And in the last two or three years, we have learned with artificial intelligence that actually multimodal retinal imaging is a perfect playground for AI. However, at the end of the day, the most important question that remains is actually, what is the impact on function for all these morphological changes? Because this is essential for the affected individuals and patients, and also, and this should be not forgotten, of importance for regulatory bodies, caregivers and payers, particularly when we are trying to introduce a new theoretical intervention. And in this context, I would like to introduce today the term inferred sensitivity, which is the AI-based estimate of local retinal sensitivity from retinal structure as detected by multimodal retinal imaging. Inferred sensitivity has been already applied in a few conditions, including AMD, and it has been also suggested that inferred sensitivity may serve as a crazy functional outcome in clinical trials. And then before I will talk more about the prediction of retinal function based on morphology, I would like to share with you a clinical case. So this is a 51-year-old female, relatively young, diagnosed with AMD. And please imagine this is a typical patient seeing other doctors coming to the university setting for a second opinion, and the patient is really eager to get involved into a clinical study. Reports of slow deterioration of both eyes more on the right and the left, and the general and family history is basically unremarkable. In colors, we see substantial pigmentary changes, vitreiform material, hypo-pigmentation, as well as large and small drusen, including particular drusen. These later changes come much more prominent with the force in the geography, but we see the typical stars in the sky pattern. Of course, we also, we are in the era of OCT, so of course we're also doing OCT, and there's one particular finding which catches our attention, and this is his cleft. It's a separate little pocket there. So what are you going to do with this? Anybody? I don't see any hands, so of course we didn't do this because I showed you the flow scene, there's no leakage. But this is actually very, very some, so should we treat this? And what is the prognosis? Of course, they have to fear that this may collapse, but the question is when will it take month or even years? Will we see atrophy? And thus, I was coming back to my first question, will the patient lose vision? So, we observe this patient, this is the picture one year later, and you see there's still separate fluid, and the time cost goes on, actually it goes over seven years, and indeed at year five you see the collapse, but this looks pretty normal to me. This is so who is serious again, you see the pigmented between attachment to collapse and almost normal retina. At high resolution, we can actually appreciate that at the time of the collapse, the outer retina looks pretty intact, the external limiting membrane is continuous, and they only settle iterations below at the outer retina. What was the visit at this point? How well did the patient see at this point? You were, this is my next slide. I was saying, okay, I'll shut up. The title of my talk is Drusen and Function. So I have to talk about this. All right, all right, all right. I stand correct. But this is for the people, not for us, the people who don't believe us. This is a dense OCT scan showing you again that there's indeed a collapse and etc. So what about visual functions? So the next question, this is best corrected visual acuity, and it's actually getting better. You know there's some noise in this assessment, so I show you now the visual acuity as assessed in our clinical trials unit, it was a natural history study, and there's not so much noise, because also reflecting if you do this EDTRS assessment, it's much more accurate. We also did low luminance visual acuity. I'm not sure if everybody is aware of it. It has become recently more and more popular because it picks up changes in AMD patients, which have typically troubles, low light conditions, contrast sensitivity. We started this the last two years, but in this case, it doesn't give you much signal. Basically I have to say low visual, low luminance visual acuity assessment is based on BCVA assessment, the only differences is that you put a neutral density filter in front of the eye of the patient. But we're also a little deeper into this, and now we are coming to fund this controlled perimetry, or also called microperimetry, which allows us to project stimuli at certain metal locations in real time to assess retinal sensitivity. So this is one year after call-up, second year and third year, done with a NIDIC MP1 device. And to a very high level, if you have green colors, this is good retinal function. If it's red, it's bad, it's a loss of function, and yellowish is somewhere in between. And actually again, it's getting more greenish, so it's actually getting better over time. Probably also reflecting learning effects of this test. When we look closer to this, and this is now getting a little bit more into detail, I highlighted here the points at the side of the lesion, which were by random at this very location of the PED. And the red colors are showing you the points which show a loss compared to normals over five decibels. And if you compare this to the overall sensitivity for all 56 test points, there's no difference. So on a rough estimate, there's no difference between this very specific lesion on the retinal and the overall sensitivity of this eye. We also did scotopic testing. And here, there's a larger difference. The area involved has more points with a reduction, severe reduction, as compared to the overall sensitivity assessment of this eye. So showing us that there might be a signal in scotopic sensitivity, it might be more accurate to assess functional deficits in these kind of patients. Saying this, I would like to talk more about fundus control perimetry. And I have to say as retinal specialists in the community, this has a very bad reputation. We are not used like the glaucoma specialists to do visual field testing. It's complicated. And we are spoiled by multimodal retinal imaging. It's so easy. It's so fast. And it's so accurate. In fact, our fast resolution of fundus control perimetry is about two orders lower as multimodal retinal imaging. It takes more time. It's not seconds, it's minutes. And also, it's a subjective test. So patient cooperation is highly relevant. We have to take into account patient-specific factors like test time, wrong pressure events, and other behavioral factors. And basically, there's a trade-off between test time, spatial resolution, the number of stimuli you can project in a patient during one session and a point-wise test accuracy like the chess strategy. So again, I'm not sure if there are any glaucoma specialists here, but this is the far you can go with retinal specialists on this topic. So this is a case of geographic atrophy. Let's have a look again. This is a typical multimodal imaging, even OCT, eye resolution OCT. And now let's have a look from the patient perspective because we're talking about function today. So we have to turn around the lesion because the patient sees outside down, so to speak. And if he looks at the chart, this is a typical finding. I'm sure you've seen that before. You have the four ways of sparing a patient can pick up single letters on the chart while they're at the same time reporting of difficulties while reading. So if you apply retinal, the fundamental perimetry in this eye, and this is a typical standardized grid, these points are not really picking up the problem, the lesion. The individual distribution of atrophy spots is not recognized. The result is that we have a limited and variable number of tests similarly in relation to atrophic areas. And particularly the assessment of function changes over time is challenging. So to overcome these limitations, we have introduced what we call patient-tailored perimetry grids in geographic atrophy. So basically, we're taking the imaging data and then we are generating individualized grids. So we design eyes or holes around the atrophic lesions like an onion-shaped pattern. And in order to maximize the number of test points in disease-relevant regions, so in the junctional zone of atrophy, this is showing you the GA, and then you have its predefined distances to the border, tests to be placed up to about 800 microns away. And at the same time, we have an overall number of test points is limited. And this is showing you some data. So this is done with a Maya device. It's a different device, it's a nighting device. It has three different types of testing, so mesopic testing under room light conditions, and two DAC-adaptive tests, sine and rat. This allows us to better discriminate rod-from-cone function. So DAC-adaptive sine testing particularly shows us rod function. And this is a GA border, and we're going towards the periphery in all three types of testing. And as you can see, the sensitivity is, in this test, the lowest. So indicating that rod function exceeds cone dysfunction. Point number one, I was a little bit too fast, shows you that all these three curves are going up, so there's an increased sensitivity away from the GA border. And even this is only cross-sectional. This also suggests, because the curve is much flatter for DAC-adaptive sine testing, that rod dysfunction precedes cone dysfunction in this disease. So now we're taking this data and applying it for its sensitivity. So we would like to investigate the AI-based prediction of retinal sensitivity based on retinal microstructure. And particularly, I would like to examine the importance of various imaging features and also the potential influence of behavioral factors, as well as looking more into these potential differences between rod and cone function. So this is data I presented at the Macular Society. This is a study where we looked at 41 GAIs with a mean area of 6 square millimeters. So multi-modal imaging was carried out in addition to fundus-controlled perimetry with patient telecrits using this Maya device, involving mesopic testing, DAC-adaptation, and both DAC-adaptive red and cyan testing. So lots of battery of tests to get a good overview of morphology and function in these patients. We also included data for the healthy eyes for standardization to patient data. This shows you an overview, high level of the analysis strategy. So with all this data acquired, first we generated thickness and intensity maps of different retinal layers and imaging modalities. So for OCT, for each slab thickness and intensity maps were generated. Then this data was registered to FCP results. So the Goldman III circle area, for each circle area we have morphology gator. And this was used to employ patient-wise one-out cross-validation with random force prediction models in order to assess the absolute error for prediction accuracy of retinal function. So in other words, we generated functional maps of the retina and then could compare at each location the predicted function to the actually psychophysically measured values. And this slide shows you now the prediction accuracy in different scenarios. Or again, for the three types of testing, mesopic, DAC-adaptive side, DAC-adaptive red testing. This is a mean absolute error and we looked at six different scenarios for each of the three types of testing. And we'll go into details in the next slide, but you see that the accuracy varies between roughly five decibel to about 2.6 decibel. So going into more details, and this is only shown from the topic testing, we looked at different scenarios, as I said. So in scenario 1A, we have an unknown patient, so we haven't done any fundamental perimetry in this patient. And then we apply our model based only on imaging features. So there are 26 markers like thickness of the O and L, the reflectivity of the outer plexiform layer, et cetera. And if you apply the model, we can predict the function with a mean error of 4.6 decibel. If we include in the model B and C behavioral factors like wrong pressure events or even fixation, there's basically no further improvement in scenario 1. In scenario 2, we assume that we have done a short funders-controlled exam, 50% of test points to be precise. And then we can improve our error to 3.1 decibel and even can further improve it with behavioral factors to about 2.9 decibel. So in other words, a certain retal location, you can predict the function in this disease with a mean error of 2.9 decibel in the best model. This is actually substantial difference compared to the first model. So what is the most important feature for prediction? The most important morphological feature and this is the outer nuclear layer thickness. This applies for all three types of testing. And I think, I believe, actually, this is highly biological plausible as well because either the photoreceptors fall out second and third are full retinal thickness and either the inner retinal thickness or the RBEDC thickness. Going a little bit more on the data, this is more complicated. If you look at the structure function correlation of the outer nuclear layers for the most important predictor of morphology, again, for all three types of testing, this is normal O and L thickness. And if the thickness is increased, the curve is flat. So an increased O and L has similar sensitivity as compared to a normal O and L. If the O and L gets reduced, all curves go up. So this function goes down and this steepest curve is seen for the dark end of the scientist, suggesting that the strongest degree of functional loss is seen from the wrong function, again. Fitting in this model on this current understanding that we have earlier and more pronounced broad versus curve function in the disease. So this is the case to better illustrate you this model. This is a typical color funders photograph of a GA lesion. And here below we see the sensitivity map, so the generated maps of function as calculated by the model. And there's a reddish and bluish color at the location of the large patch of photography. So the model picks up that there is dysfunction. And the model can even do more that two red or two blue spots above here. And these are tiny spots of photography. You can hardly see them on the colors. You actually need the OCT to see this Coroidal Hyper Transmission and the loss of the outer layer. So the model can also pick up the functional loss of these two small foci. And furthermore, there's a rot free area in the very fovea, which also is of course biological plausible. So inferred sensitivity, and this is now looking broad on the topic, can be used potentially in the future at the quality functional surrogate endpoint in future clinical trials. And I think a meaningful application would be that we have test data sets. So a small amount of patients where we did, as we did in this study, extensive psychophysical testing and volumetric multimodal retinal imaging. Then we apply AI to apply this data, this models to larger cohorts. This was allow us to predict retinal dysfunction based on non-invasive retinal imaging. Why the FCP testing would only include a limited number of test stimuli or is even completely waived. So in conclusions, retinal sensitivity may be inferred from structural retinal imaging using AI algorithm, such as random forest regression and patients with geographic atrophy secondary to AMD. It may be also applied to other diseases, but of course we think that this is all disease specific. In contrast to burdensome FCP, this approach provides sensitivity maps, which surface the limits of psychophysical testing in terms of the area covered and the spatial region. And finally inferred sensitivity may serve as a crazy functional surrogate outcome in clinical trials, especially in the consideration of retinal regions beyond areas of geographic atrophy. And I would like to close with a final case to close this morning session today. This is again a patient with geographic atrophy. Monica has shown this patient before and this is not just pure geographic atrophy. There is this little area here which is by essence CNV lesion and the phobia. So this is the part you really would like to rescue in your patient. And if you now apply microparimetry, you can actually see that retinal function is just preserved in this area where we have the QCNV location. Thank you very much for your attention. So, Stefan, again, have you looked at some of these differences? That's part of the beauty of obviously being here because we have so many of the genotype between homozygous risk at one and homozygous risk at 10. Is that starting to correlate with some of these other differences that you're seeing yet or is it just too soon? We're just beginning to dig into that. I think we are at the very beginning. This is not the end. This is maybe the beginning at the end of the beginning. I think in the chromosome one cases, you see more drusen and it's probably more likely that you have this QCNV cases in this group. In terms of function, maybe the markers are different. So in the chromosome 10 patients, I think we see more severe autorental thinning. And we've certainly seen homozygous 10s that have no obvious drusen at all and have got a geographic atrophy for sub-retinal new medicines. They have severe coroidal thinning and they have severe autonuclear layer thinning and they're actually struggling. If you ask them, they are struggling with vision and low light conditions. So particularly here being in Utah, I'm impressed about all the stars at night. I'm sure if I will ask patients, they may say the first time I noticed I have a problem was that I couldn't see the stars anymore. Fascinating. So have you approached the regulatory authorities either in Europe or in the U.S. with this idea? So far with quasi-functional yet or not? Do you think they'll be receptive to it or not for approval of drugs or anything? I think that there will be. Monica can maybe say a little bit more about it. I have been discussing not with regulatory authorities but several pharmaceutical companies and they're very interested in this. And again, the biggest struggle at the moment is to transfer this application from a uni-center setting to a multi-center trial. And this has been, I think this is the next step that we perform a multi-center trial to demonstrate that this method can really be applied. So I'm sure Monica, if you would like to add something. Yeah, so Paul, this is of course one of the main questions. So we are actually at the beginning with bringing micro-parametry or funders-controlled parametry as an accepted primary functional outcome measure into trials in AMD. And actually, we got a signal by the FDA that if we show some more data, this could be accepted as a functional endpoint. So we saw central visual acuity does not work, not in geographic entropy. It does not work in early or intermediate AMD because the lesions are not always in the very center, in the fovea. And so visual acuity, best corrected central visual acuity will not guide us or will not help to prove a therapeutic effect because it does not pick up the disease progression. This is why actually we, but also other people worldwide, are working now on bringing micro-parametry forward that this gets accepted by the FDA and the functional outcome measure. But as Stefan mentioned, micro-parametry does not have the best reputation right now because by using other devices, there was extremely high variability. And the test did not take into account the lesion area, the lesion size. And so all the results were not really comprehensive. And this is why we worked and we are working now on patient-tailored grids or lesion-tailored grids. And we generate data on reproducibility, retest data. And so we hope that this will help the FDA accept this as a functional outcome measure. Thank you very much. Have a great day.