My role is to talk a bit about risk assessment, and if I can work out how to go forward, that would be good. So, I'm going to talk about the factors contributing to breast cancer risk, risk models, and advances in risk modeling. Now, clearly, we all know that population-based screening is really not quite sufficient for purpose, but if we're going to be able to offer risk-adapted screening, it's essential that we can precisely estimate risk in an individual woman using well-calibrated risk models that take into account factors such as ethnicity, for example, and then we can institute appropriate imaging, taking into account factors like density, and that will be the subject of the next two talks. The classical risk factors are shown here. Other than age, the important ones are family history, then there's the life course factors, and the lifestyle factors as well, and let's not forget the impact of benign breast biopsy on personal risk, particularly if you have an atypical proliferation, and notably, lobular carcinoma in situ. Together, these risk factors can be used to define a high-risk group who might benefit from more intensive screening or preventive treatment, and a low-risk group, because where the benefits of screening are full, we have to consider that the harms of screening might be more important. This slide really emphasizes the importance of family history, and from it, you can see that the more affected first-degree relatives you have, the more your risk goes up. Now, if you've got a lot of affected family members, that raises the possibility of a significant dominant genetic mutation, and I'm not going to say too much about that, because these women should be referred to a specialist genetic clinic for assessment, but I'd just like to point out that though the odds ratio with a BRCA1 mutation in particular is extremely high, fortunately, it's not that prevalent in the community. I would like to draw your attention to the fact that it's really only the BRCA1 mutation that is responsible for cancers in women under the age of 40, and it continues to have an effect on risk throughout latter years as well. You can see also that regarding the genes of lesser effect, as you get older, their risk in relation to population risk becomes less significant. Of great interest recently has been the single nucleotide polymorphisms, or SNPs, which represent change in one nucleotide in a gene mostly identified through genome-wide association studies. Initially, there were 18, then 77, 143, and now we're running at 313 SNPs, shown to confer breast cancer risk. Individually, their effect is quite small, but you can add them together to form a polygenic risk score, and notice that if you're in the top centile for PRS313, your lifetime risk is actually very high indeed. It probably accounts for about 20% of familial risk, but importantly, the effect of the polygenic risk score does not seem to be mediated through density. Regarding age at menarche, menopause, first full-term pregnancy, etc., etc., there's not a lot you can do about that. If you have your first child under the age of 20, you have lots of children, and you breastfeed for a year on each occasion, you can reduce your risk quite significantly. But perhaps of more interest are hormone replacement therapy, alcohol intake, and body mass index, because, of course, they can be modified. Body mass index in particular is an interesting one. It looks as if its weight gain as an adult, and particularly after the menopausal transition, that is important in terms of risk. And, of course, at the moment, none of these factors are considered in a population-based screening strategy. Aside from age, gender, and possibly the polygenic risk score, the really important factor is breast density, and this too is modifiable. The very well-known meta-analysis from 2006 indicated that going from the lowest to the highest quintile of density increased the relative risk four to six-fold, depending on what measure of density you were using. And if you take very dense women compared to average-breasted women, the relative risk is increased two-fold. A recent re-analysis of that data, and of data accrued since then, suggests that, if anything, the strength and precision of density as a biomarker has been underestimated, particularly for continuous measures. And the importance of density is just illustrated by this slide. You'll notice that if you have more than 50% density, sure, your relative risk is increased to some extent, but because you find it in up to 40% of the population, it may account for as much as 50% of cases in the population, the population attributable risk proportion. Indeed, these authors found that density was the most prevalent, the strongest risk factor, and they estimated that if they were able to convert BI-RADS C or D women to BI-RADS B, they would be able to avert about 40% of pre-menopausal and nearly a quarter of all post-menopausal cancers. Now, the topic of how we assess density has been very well covered, and it's a whole different lecture. Of course, we can do it visually in the clinic. We can use automated area-based methods or volumetric methods, and I'm not going to dwell on that, but I would like to emphasize that with visual assessment, the problem is the significant inter- and inter-observer variation, particularly in the interquartile range, which is where most women sit. Then there's the impact of post-processing of the digital mammogram. This is the same woman a year apart on two different units, and she hadn't gone through the menopause or stopped taking HRT. There are international differences, and then there's the impact of the BI-RADS fifth edition as opposed to the fourth edition. Finally, with increasing usage of DBT, there is some evidence that density appears less on synthetic 2D. There are many fully automated area-based softwares available now. They mostly utilize the four viewing images, which is an advantage, and they do seem to be associated with risk as you would expect, though some groups have found that the associations with risk are somewhat weaker compared to, say, the visual analog scale. The volumetric methods consider pixel intensity in relation to compressed breast thickness, and they have to use the raw images. On the other hand, they have the advantage of being reliable and repeatable. Now when it comes to risk prediction, density is significantly associated not only with the risk of having a screen-detected cancer, but also interval cancers, no matter what measure of density you use. In this study from the Dutch group, you can see that Volpara density grade D women, they had a higher incidence of screen-detected cancer, but they had a much higher incidence of interval cancers. On the other hand, Volpara and clinical BI-RADS had similar discriminatory capacity when it came to risk. Now this study is from the Manchester UK PROCAS study predicting risk of cancer at screening. The group here looked at five measures of density, and what they found was that both for screen-detected cancer at entry and subsequent development of cancer during the follow-up period, the visual analog scale was most strongly associated with risk, closely followed by Volpara. And interestingly, for this population, Contra didn't seem to be predictive at all. No matter what method you use, it's important to recognize that there is very poor agreement between methods. So the same tool should be used if you're doing longitudinal studies on an individual woman, perhaps to monitor the effect of an intervention, or if you're studying a particular population. Then there's the question of whether you should use absolute or percentage measures. Intuitively, you would think that an absolute measure was more predictive, and indeed at the higher end of the density spectrum that appears to be the case. Other groups have found that percent density seems to be more predictive, but I think that's probably a reflection of the confounding effect of body mass index and age. And since most women fall into BI-RADS 2 or 3, or BNC categories, it makes sense to use a continuous measure of density, which is likely to be more discriminatory. Now of course there are many other mammographic measures of risk, in particular texture and parenchymal patterns, and it's many years now since John Wolfe first described this. This is an area of intense research, but at the moment these textural measures are not ready for routine clinical usage in the clinic. There are many case control studies, they all show a clear association with risk that appears independent from the risk with density, and indeed the effect of texture and density together appears to be additive. But really we need more external validation, and this was really well shown by Wang et al., who developed a NICE measure that seemed to work very well when applied to the UK PROCAS cohort, but when it was applied to a group of US women, it had no predictive effect whatsoever. Looking at the risk models, I draw your attention to a NICE review here, which reviews them in detail. They've either been developed by and large through regression models from case control studies, or through genetic segregation studies, if you're suspective of the presence of a high-risk genetic mutation. It's important to bear in mind that the risk factors and the weightings differ between the various regression models, so you need to choose your model carefully, bearing in mind the population that you want to look at. For example, the Gale model was developed looking at white US women, the Tyracusic model through white European women. Nonetheless, despite all those considerations, these models can identify women who either are eligible for supplemental MRI or who need chemo prevention. Now, when it comes to risk models, there are a few important statistics. The first is decalibration. So that is the ratio of expected to observed cancers, and a ratio of one means the model is perfect. So here, for example, you can see that at lower levels of risk, the model is performing pretty well, but at high risk, it is overestimating risk. At an individual level, the important statistic is the C-statistic, or area under the curve. By and large, an area under the curve of 0.7 would be regarded as okay, and 0.8, good. But it's noteworthy that for most of the risk models in use, you're very lucky if you get a C-statistic greater than about, say, 0.6. And finally, there's the net reclassification improvement, and this is the percentage of women who are reclassified correctly with a newer model into higher or lower risk categories. So what happens if we add density to a model? This was first done in 2005, and the early studies did show better discrimination, but with only very modest effect. And in fact, if you find a study where the effect is extremely large, you will tend to find that it's a small study without external validation. An important consideration is that density is extremely strongly correlated with age and body mass index, so you need to adjust for that with a density residual. Again, the population you're looking at is highly important, and also the risk period, because many of the models perform quite well at predicting one- to five-year risk, but are not so good at 10-year risk or lifetime risk. At the moment, these two models, incorporate density, and they both have the advantage of being web-based, and density is also now incorporated into Bodicea, too. It may be that density alone is insufficiently discriminatory, and in this case control study, when they looked at BIRADS-D women, they found that only those with dense volume in the third and fourth quartile appeared to be at significantly increased risk, as you see here from the Breast Cancer Surveillance Consortium five-year risk estimates. In this U.S. study, either BIRADS or volumetric density was added to tyrocusic, and you can see the impact of that is to better spread out the risk, and in this particular study, actually, BIRADS density performed the best. In a sub-cohort group from the PROCAS study, the authors took tyrocusic version six and added to it a density residual using the visual analog scale and a polygenic risk score, and you can see the progressive improvement in the area under the curve as you do that. Importantly, in the 20% or so of women at more than five percent 10-year risk, they accounted for 30% of all breast cancers and 35% of interval cancers, and at the other end of the spectrum, the 33% at low risk only accounted for 18% of all breast cancers and 17% of intervals. In further work from this group, they then looked at the impact of going from a polygenic risk score with 18 to 143 SNPs, and if you look at the purple bars here, you can see how an increasing number of SNPs and a density residual better define for the first time a low-risk group and a high-risk group, and importantly, it seems that the effect of density and a polygenic risk score are independent, so they can be used together. And finally, and I'm very grateful to Professor Gareth Evans for this, what you see here is the impact of the enhanced tyrocusic model, and at the lower end, you can see about 19% of women developed 7% of all breast cancers, but very few of them were advanced, whereas at the 5% or plus 10-year risk, these women developed 43% of the breast cancers and nearly half were advanced stage. Now, this sort of approach has been used for other models and recently validated in other countries, but importantly, we need to find a polygenic risk score that works better for black and ethnic minority women, because it seems that the one in common usage doesn't. So just finally, I cannot leave this talk without saying something about machine learning, and in particular, deep learning. This has been used to look at density assessment, risk assessment, and even brackiness from a mammogram. This group were able to predict visual analog scales simply from their CNN model. Lima Natal validated a model, which they tested first amongst their academic radiologists, and subsequently, Donchus Natal took it out into the community, and what they found was that for binary classification into dense, non-dense, there was very good acceptance, though I accept that acceptance is not the same as agreement. And finally, Mohammed Natal used huge number of mammograms in order to train their model to differentiate between BI-RADS C and D mammograms with very significant effect. And there are lots of groups who are looking at deep learning for risk prediction. The reason I want to mention this particular work is for the following. The MGH group in the publication by Yala Ratel found that their model was able to differentiate between women at risk regardless of breast density, and much better than Tyra Cusick, as you see from the gray bars. They've subsequently refined that model and recently tested it in seven other sites in the U.S. and in five other countries, and what's really exciting is that this appears to discriminate equally well for risk in non-white and other ethnic groups at all ages and densities. And just to finish up with, you can see here in the MGH cohort defined as high risk by using Tyra Cusick more than 20 percent lifetime risk. You can see from the operating point here that at that operating point, their model had better sensitivity and specificity. So in conclusion, in this lightning whistle-stop tour, I hope I've shown you that density and a polygenic risk score improve our ability to identify risk in a woman, but we're not quite there yet. I do believe that quantitative measures with radiomics and AI are going to help us big time in the future, but we've got a lot to do in terms of defining the best metrics, exploring optimal incorporation, standardizing our methodology and reporting of results, and critically undertaking robust external validation. Thank you very much indeed. So Sarah gave an outstanding presentation about breast density as well as SNPs and other risk factors. One of the things that I just want to start with is that dense tissue can mask detection of cancers. Mammographic sensitivity is only on the order of 50 percent or even lower in the densest breasts, and also these women are at higher risk for having an interval cancer detected because of symptoms after a normal screening mammogram. And so for those reasons, we consider supplemental screening. I'm going to be talking about ultrasound, which the focus has really been on women with dense breasts in general. So when should we consider screening ultrasound? I'll get into some of the details. What are the benefits? What's the cancer detection rate? What are the types of cancers? Do we see reduced interval cancers? And what are the cons? False alarms and implementation issues are challenging. In general, one of the virtues of ultrasound, of course, is that it does not require any contrast injection, unlike contrast-enhanced mammography and MRI. There's minimal discomfort to the patient, no ionizing radiation, so we're not causing even theoretical exposure to radiation. It's easy to guide biopsy of anything that we find on ultrasound that's suspicious. It has been shown to reduce interval cancer rates in studies from Japan. Even the ACRIN trial, we saw a low interval cancer rate, and in studies from Italy. And the cancers that are found are nearly all invasive cancers, so these are biologically more important. Overall cancer detection rate, I'll go into more detail, but average is about two to three per thousand. And it's a widely available technology that's relatively inexpensive. On the converse side, there is a relatively high rate of induced biopsy due to findings on ultrasound. And as I'll go into in more detail, most of those findings are of low suspicion. They're either BI-RADS 4A or even sometimes BI-RADS 3 lesions that prompt a lot of unnecessary follow-up, if not biopsy. And then finally, there remains a huge issue with quality and variability in practice. I'm not sure why the ACR has not really invested in training for breast ultrasound screening, for example, or specific criteria. We did have very explicit training for the ACRIN trial. I would be happy to share the teaching files with anyone who's interested. Here's just an example from our recent study. We're actually comparing ultrasound with tomosynthesis, and the studies are being performed by technologists. In our study, the women have dense breasts and not necessarily any other risk factors. And in retrospect, on her mammogram, she does have a little portion of a node seen in the axilla, but this was called negative and extremely dense. On screening ultrasound, she had an obvious irregular hypoechoic mass, 1.9 centimeter grade 2 invasive ductal cancer that was node positive, as you saw. One thing I do want to consider, you know, we generally talk about ultrasound as a supplement to mammography, and that is the most appropriate use. But we did, we were able to analyze the ACRIN 666 study with respect to what if only an ultrasound was performed for screening to compare it to mammography. Because obviously, if you take away all the cancers seen by mammography, ultrasound is not going to have the chance to look quite as good. And it was interesting because of the 111 cancer events, 58 were detected by ultrasound and 59 were detected by mammography. And that's when they performed alone. And the cancers seen by ultrasound were much more likely to be invasive cancers, 91% compared to only 69% with mammography. And you can see the node positivity was actually greater with those seen on mammography. So if you were to choose a priori with no existing screening program, which test to implement, you might actually start with screening ultrasound based on our results. So who generally gets this right now? Well, in our own practice, one of the not uncommon scenarios is a patient who actually can't tolerate a mammogram at all. We could argue about this, but our disability clinic sends these patients then for screening ultrasound. They don't always schedule it appropriately, and sometimes it really bolloxes up our day, but this is what happens. We also have seen multiple national guidelines recommending that for women who are appropriately considered for screening with MRI, but it can't be performed, and that could be for a variety of reasons, including pregnancy, access issues, pacemaker, other metallic implant, or even cost, that screening ultrasound or contrast-enhanced mammography should be considered. And of course, those are generally increased-risk patients, and I'm sure Dr. Mann will discuss that in more detail. The bigger category where it's evolving is the intermediate-risk women, where MRI has been suggested by some national guidelines, including women with personal history of breast cancer, prior atypical biopsy, or LCIS, and women with extremely dense breasts. And I'm aware that USOBE guideline is currently being revised to include potentially MRI screening every four years in those women. There is this broader group of women where the lifetime risk of breast cancer is intermediate. It's not quite the 20 percent threshold that was the ACS standard in 2007, but they have a higher risk than the average patients, 15 to 20 percent, based on family history. And we don't fully, you know, we have these risk models, but as Sarah showed you, they're not that perfect at the individual level. The interplay of these factors is challenging, and so there are many women who have heterogeneously dense breasts and maybe no family history who are somewhere in this middle category where cancer can be hidden, there's more interval cancers, but there's not clear guidance at this time. One thing I do want to mention, and it was very interesting to me, the results were so similar in ACRIN study as well as the Netherlands Dense Trial, that many women who are offered an MRI screen, even for free, will not do it. It's not that attractive of an exam for a variety of reasons, including claustrophobia, injection with an IV, all these various factors, but in the ACRIN study, we saw that only 58 percent of women chose to have an MRI, even at no cost, and interestingly, 59 percent in the Dutch study chose to have an MRI. So I think that's pretty telling that there's a large group of women where they're not going to necessarily have MRI even when they meet the guidelines. Screening ultrasound, so generally should be considered complementary to mammography. We did see 82 percent of all breast cancers when we looked at the combination of the two tests, but importantly, and I'm sure you'll hear more about this in the next talk, even with the combination of ultrasound and mammography, another 15 per thousand cancers were detected when we added screening MRI in the last round of the ACRIN trial, and that's very similar to many other studies, including Christian Kuhl's work and Sardinelli's studies and others. So if you can do the contrast, it's clearly a superior test in terms of cancer detection. Here's just an example from the ACRIN trial of women with dense breasts and a personal history of cancer, and she had a high-grade invasive ductal cancer seen only on ultrasound at the 24-month screen. So the most recent American College of Radiology appropriateness criteria in discussing these women with dense breasts where they're at intermediate risk say that screening ultrasound may be appropriate, but there's disagreement, but women who do not have dense breasts, it's usually inappropriate, and I do want to touch on that last issue a little bit. Average-risk women with dense breasts, they say it may be appropriate, but again, disagreement, and average-risk women with non-dense breasts, it's usually inappropriate. There was a nice analysis of the Japan JSTART study recently published that looked at a subset of women in one of their prefectures where they were able to link the data to the tumor registry, and you'll recall these women were all randomized to have screening ultrasound in addition to mammography or not. And all of the people performing the ultrasound were trained in a standardized two-day course. It could be either a physician or technologist doing the screening ultrasound. And in the sub-study, they found that the cancer detection rate in the women with dense breasts increased to 7.1 per thousand compared to 4.3 per thousand in those who didn't have the screening ultrasound. The interval cancer rate was significantly decreased to .5 per thousand versus 1.8 per thousand. What was surprising was that even in non-dense breasts, they also found a very similar increased cancer detection rate. You can see 6.9 per thousand versus 3.6, and also a significant decrease in interval cancers. So they basically had the same added cancer yield of 2.8 to 3.3 per thousand regardless of breast density, and also, importantly, a reduced interval cancer rate, and this was a two-year interval. I'm not advocating that we do screening ultrasound in women with fatty breasts, but I just think it's important to be aware that there are data that actually support it. So we have several methods to perform ultrasound, handheld, physician or technologist performed using automated arm or automated whole breast ultrasound, and there are other methods in development. For handheld ultrasound, it's important to use standard documentation, and I think most centers have adopted what we did in the ACRIN trial, which is survey scanning in the transverse and sagittal plane, documenting a single image of the largest cyst per quadrant or a negative image for each quadrant, as well as a negative image behind the nipple. It is important for any non-simple cyst lesion to document orthogonal views with and without calipers, and if you can, also with and without Doppler if it's not a simple cyst. Elastography is optional, but can be very helpful for some of the low-suspicion lesions. I'm not going to talk about that further this morning. One of the things that we're learning in the current study where the technologists have been doing the ultrasound is that it's often the case that the very periphery of the breast has not been fully scanned. I've been surprised at how many cancers we actually find after the fact based on tomosynthesis finding where it was missed by the technologist, and so it's always good to look at the breast, see where the gel is, make sure they've really covered the entire periphery of the breast because, in principle, that's actually where ultrasound can improve on the mammogram. You're not limited by positioning, so of course you want to emphasize the edges and particularly the upper-outer quadrant because that's where the money is. It's elective to include the axilla, and really the data don't support that. In the ACRIN trial, we had one patient, about a third of patients had screening ultrasound including the axilla. Only one patient out of 2,500 had a metastatic node identified with an ipsilateral cancer. There were two others where it potentially could have helped but wasn't done, but in two other studies, they basically, this study out of almost 13,000 women, they had only false positives by looking at the axilla, and in this study, they found two non-primary breast cancers in the axilla metastatic nodes from other sources. So overall, regardless of how the ultrasound is performed with either physician, technologist, or automated technique, you get the same cancer detection rates of about 2 to 3 per 1,000. This was maintained in incident screens in the ACRIN trial as well as the prevalence screen. And I've mentioned this issue of the JSTART trial. I think it's very interesting to note that if you compare technologists performing the test to physicians, at least in their training, they found that the technologists had better sensitivity for cancers. When they looked at videos of ultrasound, they had better sensitivity looking at still images than the physicians did. So it is certainly the case that physicians aren't necessarily better at doing this than technologists that are well-trained. There are fewer studies looking at ultrasound after tomosynthesis, including a nice paper Dr. Destounis had presented, but again, the cancer yield seems to hold up at about 2.7 per 1,000. Of course, there are some cancers seen only on tomosynthesis, not seen on ultrasound, so we would not abandon the tomosynthesis, of course. This is from our current study. This patient has dense breasts. The tomosynthesis was called negative. In retrospect, I'm showing a close-up of maybe one or two slices, 1-millimeter slices, that we could see the finding only on the CC view in retrospect. On ultrasound, she actually had two separate areas that were concerning. And this is a double-reader study. So both readers called this suspicious at 9 o'clock in her right breast, a slightly different area than what I just showed you. And a much more subtle finding was called suspicious only by one reader in the 12 o'clock position, not even measured by the technologist. But this, in fact, was an invasive lobular carcinoma, and the initial area was PASH, false positive. So cancers can often be subtle. We do know from the literature that about 15 to 20 percent of cancers seen only on ultrasound are invasive lobular carcinoma. They are overrepresented among cancers seen only sonographically. A nice paper from Min-Sung Bae and colleagues showed that cancers that are seen on ultrasound are more likely smaller and more likely luminal A tumors rather than HER2-positive tumors that were more commonly seen on mammography, and of course, HER2-positive disease often has calcifications associated with it. Here's another patient from our recent study, dense breasts, nothing in retrospect. And on screening ultrasound, and I think this one could have been mistaken for a cyst. It's kind of looks like a complicated cyst, maybe ruptured cyst, and it was recalled for additional evaluation. And on the targeted ultrasound, it's more clearly at least indistinct, if not micro-lobulated mass, and this was a grade two invasive ductal carcinoma. Automated ultrasound generally is performed with these wide field of view transducers, 15 centimeter footprint, typically takes about 15 minutes to perform. So there's actually no savings of time compared to handheld ultrasound, which we found average was actually 13 minutes in year three, 19 minutes in year one in the Akron study. And then you end up with thousands of images to review, and sometimes you need additional acquisitions for larger breasts. And of course, there's also the approach using the automated arm that's adapted to a standard transducer and equipment. Here is a patient, Dr. Vortis' series where she had, the patient has dense breasts, multiple circumscribed masses, could be cysts. In this automated ultrasound, you can see the circled lesion here in the upper outer left breast that is, has enhancement, and that is a cyst, but there was also a subtle irregular mass seen nearby, and that indeed was a grade one invasive ductal carcinoma. In general, across the literature, again, automated ultrasound results, very similar to handheld ultrasound results of cancer detection rate of about 2.3 per thousand, added recall rate also substantial in the literature to date, about 10.6% of patients. There's only one paper of which I'm aware, looking at ultrasound, automated ultrasound after tomosynthesis, which was my colleague Denise Cho's paper, where again, similar cancer yield of 2 per thousand, added recall rate was also substantial. And here is one of those cancers, nothing even in retrospect, very subtle on the automated ultrasound images, and that was a grade two invasive ductal carcinoma. So the main issue that we really have to finish with is reducing false positives. If we're going to consider implementing screening ultrasound, we really have to address this head on. Across the literature, the added recall rate, anywhere between 7 and 10%. It is true, of course, like with any technique, that you have the highest recall rate in the first year. It's going to reduce when you have comparison to prior exams. And this has been, of course, one of the sources of controversy. If it's so many false positives, is it even worth doing? And there are ways to reduce the false positives. If we focus on these, mostly BI-RADS 3 lesions, we can certainly reduce unnecessary follow-up. In the Akron trial, we defined these as circumscribed oval mass, clustered microcysts, which now we would consider BI-RADS 2 findings, probable fat necrosis or scars with uncertainty. Here's a typical baseline finding on mammograms, circumscribed mass, hypoechoic on ultrasound. This is another. That was a fibroadenoma. This is typical fat necrosis, which has a very short interval follow-up recommendation. And here you can see it decreased at six weeks. Across the literature, again, about 6% of patients had BI-RADS 3 recommendations from screening ultrasound in technologist-performed studies, up to 20% in some of the physician-performed ultrasound studies. We need these orthogonal views. This is a good example of why this was one that looks very circumscribed in one view, has a little nose off of it in the second view, and this is papillary DCIS with microinvasion. And I recently reviewed the literature on the BI-RADS 3 findings on screening ultrasound and found that across 15 series, at six-month follow-up, only eight cancers were seen of nearly 4,000 lesions. That's 0.2% malignancy rate. And even at two years of follow-up, the malignancy rate was only nearly 0.4%. And those rates are very similar to BI-RADS 1 or 2 assessment. So it's very reasonable, and this was actually suggested first in the ACRIN trial. Richard Barr was the first author. Twelve-month follow-up is a reasonable alternative to six-month follow-up for BI-RADS 3 findings on screening ultrasound. And, of course, you can give a BI-RADS 3 assessment straight off the bat for handheld ultrasound because we do have those orthogonal views and we do have Doppler images. And just to close out with one last point, this is a patient with typical clustered microcysts. This is a benign finding at this point based on literature, but this patient also looks like clustered microcysts, but she has an ipsilateral cancer elsewhere in her breast, and this was grade 2 DCIS and enhanced on MRI. And I just want to highlight this one very nice paper because they looked at BI-RADS 3 lesions synchronous to a current cancer. Overall malignancy rate was 11 percent. It was higher rate of malignancy in the same quadrant as the primary at 21 percent, 10 percent in a different quadrant of the same breast, and still 4 percent malignant in the contralateral breast. There have been several other abstracts presented with similar results, but this is the only paper I'm aware of. And then one final point is that perhaps BI-RADS 3 can also be used directly on automated ultrasound. Richard Barr recently published similar results where they did this in about 400 women with good results. So to summarize, with screening ultrasound, we have comparable performance with all methods using handheld physician or technologist performed technique with automated ultrasound and apparently even after tomosynthesis, but it does require training of the staff, including interpreting radiologists, and we don't really have standardized training at this point. Most importantly, it's important to reduce false positives and audit your own experience in this regard, and as I said, BI-RADS 3 lesions generally can be seen at one year at the time of the routine screening, at least in the United States. Thank you. I'm going to talk about supplemental breast MR in this personalized screening era, which obviously comes on top of what we were just discussing. And if you think of early cancer detection, this is basically what we do with screening, right? Here, at this point, would cancer become palpable? In a setting where you are aware of palpable mass as an indication of going to the doctor, this is the moment that women walk in, right? In some other places, they might actually wait until the mass becomes really big and then it's here, and you might actually gain something by telling them, come earlier. But in this case, if they come when they're palpable, next thing would be to find them when they're not symptomatic, really not symptomatic. That's what we call screening. And obviously, screening gives a gain in survival, but it's not dependent on how big the cancer is. It's basically dependent on how big the fraction of metastasized cancers is. So what we try to do is reduce that fraction of metastasized cancers, and basically, if we find them earlier, that gain becomes bigger, just as simple as that. So the whole point of doing screening and doing supplemental screening is that we want to find cancers earlier. So what we aim for, in the end, is a state shift. That's actually the only thing that really matters, and it's also what you should get. You find more cancers, but you find them when they're small. So what does breast MR add? Well, nice images to start with that are really easy to evaluate. I mean, it's not so hard to tell in which breast disc here the cancer is. Honestly, in my opinion, this is much easier than mammography reading, really much easier. And you would put that on a screening population. There's many cancers in that screening population, and you would find these with mammography. Then you do MR, usually as a second-line screening examination, and you find a lot more. The added cancer detection rate is always in the order of about 15 per thousand. And if you look at the bigger studies that are out there, especially the later ones here, they all screened up to about 20,000 screening examinations. This is real-life screening reports, right? This is no longer studies. These are retrospective evaluations of what is capable in current practice. And what you see there is that we get a sensitivity of MR alone that ranges from 80 to about 90 percent. So we do miss some cancers with MR still, right? Some of those actually become interval cancers up to about 10 percent. However, there are also some cancers that are still found with mammography. So the question is then, indeed, should we do both? Is MR indeed the supplemental technique on top of mammography? Have a look here. It depends a little bit on the age of the women, I would say. In younger women, under 40, you need to make an enormous amount of mammograms to detect one additional cancer. Honestly, I would say that is no point. So under the age of 40, I really would not advocate performing mammography on top of MR. Going up higher might actually become more useful. At 50 to 60 years, you might find quite some substantial DCAS, because we actually looked at the types of cancer that we find only with mammography. You can see the list here. This is in our own cohort. It's about 8,000 screens. These are cancers that were only detected by mammography, and 80 percent of those is DCAS. So if we look at breast cancer growth, this is also a very well-known image, right? Some cancers grow really fast. Some cancers actually grow very slow and might never actually get to this symptomatic threshold. And if you look at sensitivity profiles of mammography and MR, they look like this. Very famous study from Jenner-Zunk. And what you get to see there is that the sensitivity profile of MR basically means that the more aggressive a cancer is, the more likely that you find it, whereas mammography is basically the other way around. The only concerning thing here is that I think that we also find a bit more low-grade DCAS with MR, right? I mean, it would be nicer if this actually dropped all the way through there. So we might also increase overdiagnosis a little bit if you would go to MR. You need to be aware of that. But still, if you would plot it here, in essence, it looks like this, right? MR finds you these cancers, mammography finds you those, and there might be an area where actually mammography finds cancers, in essence, earlier than MR. So if we talk about the incremental value of mammography, there is indeed a sensitivity increase in a range of zero to 9 percent that averages around 5 percent. There's no official mid-analysis as far as I know, but it will come somewhere in that order. And there's a small penalty to that in the sense of a reduction in specificity of about 0.5 to 2 percent, and that doesn't sound big, but actually there's a lot of false positive biopsies, right? So in essence, if you do MR and mammography, mammography is a supplemental screening technique, not the other way around. Then we go to breast MR, who to offer that. Basically, we are facing a sparsity model. You have always to fight to get access to the MR scanner, right? We are competing with neuroradiologists and MSK radiologists who really want the time that you dedicate to breast for their own studies. So we need to select our patients, and we tended to do that by selecting patients that relatively frequently had cancer and where mammography is really poor. So we'll start with, obviously, the relative risk factors that are out there, and then all the genetic ones are more or less fine, and the old Suslow paper basically also included the patients with a family history. So this is the basics, and it's more or less well accepted. That's also what's in all these retrospective evaluation studies. If you look a little bit more in the personal risk factors, actually things like personal history or LCIS are even more predictive than the family history. So you obviously get to see that we might do something useful there as well. This is just an example case, a patient with a prior history of breast cancer operated here. I honestly don't see anything here. There's no calcifications whatsoever, so there's a bunch of fibroglanular tissue. The MR images are really pategnomonic of a big DCIS there, right? You will not miss this. So there are quite a few studies out there that evaluate the sensitivity of MR in that setting, either after a negative mammogram or in combination with mammography, and you see that even the best performance of mammography doesn't get a sensitivity of more than 45 percent of recurrent cancers, whereas with MR you can get to 85 to 100. What you should realize here is that most of these studies are actually retrospective, which means basically that the patient population is selected. So they are on average relatively young. They have dense breasts. They have aggressive cancers or cancers that were not seen on the original screening mammogram and actually popped up as an interval cancer. These women are logically, I would say, not at ease with mammographic control, so they get the MR controls. Moving on from this to family history, because as I said, there's also a bit of an issue, not because it was not in the original guideline, but at least in Europe it was not really adopted. There were lots of issues about cost-effectiveness. So the gene mutation theory is yes, but anything below, well, maybe. We actually run a randomized controlled trial on that in the Netherlands. It's a small one, but it's really nice, had women with a familiar risk and a lifetime risk of 20 to 50 percent, and what you see there is that there were about 680 women in both groups. You find way more cancers with MR. Not unexpected. Remember, that's what we want. The cancer detection rate is about 14.2 per 1,000, with a supplemental detection rate of 9.3 per 1,000. You should realize that some of these women were actually prescreened in the previous era with MR, which explains why this number is a bit lower. The important things here are basically, on this slide, I would say, with MR, you detect more cancer, but also a lot smaller cancer, average size 12 millimeters versus 18 at the lower stage, and I think really important, 83 percent of these cancers are node negative, whereas in the non-MR arm, only 38 percent of these cancers are node negative at that point in time. So that's where you really make the difference. And then we did a cost-effectiveness analysis on that. This is the efficiency frontier, as it's called, and we basically said, okay, if you screen those women, the most cost-effect point at about 21,000 euros per quality-adjusted life year is one MR only, no mammography, every one and a half year. The alternative next to it, it's about here, is one year MR, next year mammography, one year MR, next year mammography, and so on. If you have an annual evaluation of these patients, that might be preferable, but actually one MR every one and a half year has a little bit better cost-effectiveness point. And you see here that the price is — oh, sorry, here — the price is basically dominated by the unit price of MR screening, which means that we really should work to actually get a price down. It's well feasible to do that at about 350 euros, but I think that we should be capable of bringing it further down to about 200 to make this really working. Next thing is obviously density. You have heard a lot about it, and this is not something that's new to anyone. If breasts become denser, the risk of breast cancer definitely increases, the risk of missing cancers increases, and this is the sensitivity of mammography if you look in population screening, right, with reports of interval cancers. If you do MR in these women, the sensitivity is not nearly in the order of 40 percent. It's more 20 or so. So a quite famous dense trial would say now — basically investigated this. It's actually only the extremely dense population and randomizing them to MR screening, as I know. And I walked you through this graph because it basically shows the entire study. 40,000 women in that study included, but they were randomized at that point in time, and only the ones randomized to MR were subsequently invited to the study. So the other ones were in the study, but they didn't know. I think that's quite important because that partly explains why only 60 percent picked up the invitation for MR screening. The others just didn't want to participate in a research project because that was what it was. We couldn't tell them whether it would work, right? So about 5,000 of those underwent MR, and in those we found 80 cancers, which comes down to 16.5 per thousand additional cancers of detection, very similar to what you saw in the previous studies. Overall, that had — in the women that actually had an MR, the official study is on invitation, but on the women that had an MR, the interval cancer rate dropped to 0.8 per thousand from 5 per thousand, which is in both the invited but not participating group and the control group. So we have a massive reduction here in interval cancers. So that is no over-diagnosis, right? We only find them earlier, and you also see that in a subsequent round where we find only 5.8 per thousand additional screenings, this still is a 300 percent increase of the detection rate of mammography, because with mammography we detected before this MR another 2 per thousand cancers. So the total amount actually drops after this first huge peak to more or less the numbers that we expect from Dutch population screening. False positive rate actually goes down really nice to 26.3 per thousand, basically 3 percent recall rate, quite similar to what we see everywhere around the globe for mammography screening. And I think this is really important. In the incidence round, so the follow-up round, all the cancers were not negative, and they were all early stage. So actually in this population, the screened population with MR, we basically have the possibility to eradicate breast cancer-related mortality, right? That's really massive. We have not ever been able to do something like this before. Then what does it do on over-diagnosis? As I said, we find a little bit more low-grade DCIS, so the over-diagnosis definitely does go up a little, but the number of life-safes goes from mammography screening once every two years to MR screening every four years. So we reduce the frequency about 60 or 40 percent up. And then if you look at the ratios, life-safes versus over-diagnosis, we definitely do better with MR screening than with mammography screening. If you look at the cost, basically the four-year MR screening has a cost-effectiveness that's around 15,000 euros per quality-adjusted life year. Actually, if you move it a bit further, then you would save, in essence, with this strategy, 7.6 per thousand lives at a cost of about 75,000 euros per life, or 11,500 euros per quality-adjusted life year. And actually, we think that is really, really cheap. It's a bargain there. So then who to screen with MR? We have the high-adverse risk group, basically the ones with very dense breasts, at least once every four years, intermediate risk, family risk, LCIS, and so on, at least once every one and a half years, and obviously the very high-risk group, BRCA mutation carriers, and so basically every year for BRCA1, maybe even more. Are there further expanding indications? Well, I guess yes, because Christiano had this very nice paper already a couple of years ago, where she actually screened all women with average risk, and what you see there is similar cancer detection rates, 22 per thousand in the first round, dropping to 7 per thousand in follow-up rounds, but hardly related to density. So actually, it's prevalent in all density categories. So we actually think that you should stratify these patients a bit further still. This is one potential option to do that. This is actually a graph that shows the proportion of interval cancers that can be detected by selecting patients based on, for example, density. If you would select 10 percent of patients, you will catch about 22 percent of interval cancers here, right? If you stratify patients on density, 10 percent, so basically extremely dense. If you would combine that with an algorithm that looks at the risk for interval cancers, that has in itself sort of a 35 percent possibility to find them. We can find in 10 percent of patients more than half of the interval cancers. So by doing this, we don't expand the amount of MRs that we make. We are still facing the shortage of MR capacity, but we can select them even better, so we prevent these kind of things from happening. That's it. Thanks for your attention. I look forward to the discussion.

risk assessment

breast cancer screening

genetics

breast density

personalized risk models

mammography limitations

ultrasound

MRI

machine learning