Clinical Focus Dialogue: AI in Mammography: Leveling the Playing Field for Global Disparities

Hello, my name is Doctor Rachel Brem. I'm the director of breast imaging and intervention at the George Washington University and the co-founder and the medical director of the Brem Foundation to defeat breast cancer. I'm delighted to be here today to share in the clinical focus dialogue and I have the privilege to share with you today some data and my thoughts about artificial intelligence in mammography and how we can use that robust technology to level the playing field of global disparities. So, mammography is an excellent examination. It's resulted in a significant portion of the 40% reduction in breast cancer mortality of that we've seen in the United States over the past two decades. And we've seen a significant impact of mammography on mortality reduction. But it's an imperfect examination. Only 85% of breast cancers are seen. Mammographic Lee today and in women with dense breast tissue up to 50% of breast cancers can be missed. In addition, there are a lot of issues with mortality reduction, in that it is unequal among different populations, with the mortality reduction in the United States. For Black American women being 40% lower, so there are problems with mammography. One is with access, not every woman has access to mammography, and it is an imperfect examination. Not only do we not see all mammograms at all cancers at the time that we interpret, but studies have shown us that if we look at the. Mammograms of women who are diagnosed with breast cancer. One year, two years and some. Earlier we can find evidence of that cancer that was not actionable and not seen by radiologists. So we have some challenges that we have to deal with. So why are we looking at AI to help us with these challenges? Because we need more help in being more accurate in mammographic interpretation and doing so in less time. We have to harness the enormous amount of genomic information that exists in mammograms to allow for accurate individualized risk based assessment and perhaps screening. Protocols as well and we have to increase the availability of not only mammography but high quality mammography and mammographic interpretation in the United States and globally to further improve the outcomes from breast cancer. And we really need to do that if you look at the world map in the top cancers. In women, you see that this pink that breast cancer overwhelms the map except for a few countries in Africa, one country in South America and Australia. Which has a very high incidence of skin cancer. Breast cancer is the most common cancer in women in the world globally. If you look at the incidence rates you see very interestingly what we've known for some time, which is in the Western countries in Western Europe, and in the developed countries and Canada and the United States and in Australia, the incidence of breast cancer is the highest, but when we look at the mortality map around the map around the world, we see something that's really very disheartening, and that is in the areas in the developed world where breast cancer. Has the highest incidence. It has the lowest mortality rates and in those countries in Africa where we have relatively low rates of breast cancer, we see very high mortality rates and if you look at the two maps side by side, it's really very stark because here you see lower incidence, higher mortality, higher incidence in the in Canada and North America, and lower mortality and so clearly this is emblematic of the disparities. In health care in breast cancer around the globe. So clearly we see this disparity and how can we address these disparities? Well, we need to deal with the most limited resource which is human resources. I've actually spoken to some of the leadership in in healthcare in Africa. They can get machines, but they can't get our people, radiologists, physicians to interpret those images. And how do we leverage this technology to address? This clinical need, how do we optimize the radiologists we have by minimizing the time it takes to accurately and interpret mammograms and decrease the differential in the worldwide training and specialization? And how do we answer the critical radiologist breast imaging shortage? And this isn't only something in the world, it's here in our country, in the United States as well. If you look at the city patterns of screening mammography uptake. It is not universal. In New England you see the highest uptake on the coasts. Both coasts. The uptake is highest, and here in the mountain regions we see the lowest uptake of screening mammography in those regions where there isn't a mammographic center or a breast imaging center on every corner and where the population is much more rural and less dense. And then when we look at ethnic differences, black women have a 36% higher age dust digested breast cancer death rate in the United States than white women and the relative risk of death in black and non Hispanic blacks is 71% higher. So we clearly have disparities that we have to address and consider. Black Race is generally correlated with lower socioeconomic levels. In the United States and what that translates to is reduced access to quality health care and the same thing is true in terms of socioeconomic for you United States, let Latin X Hispanics, and yet Hispanics with breast cancer have a prognosis that's better than even Caucasian Americans. Now we really have to think about mammography not just as mammography, but as high quality mammography. And this is really emblematic by the Metropolitan Chicago Breast Cancer Task Force. This was initiated in 2008, and it was due to marked health care disparities that we're seeing in death rates in Chicago with huge variations in fragmentation of care, particularly in Chicago, Southside and what they did is that they developed a large, patient navigation system where women in the South side of Chicago were sent. Not to mammography centers, but to mammography centers of excellence and to care for their breast cancer. Not at any local area, but NCI designated cancer centers. It was previously called the Metropolitan Chicago Breast Cancer Task Force, now called Equal Hope, and it continues to go on. But the reason it was established is because between 2005 and 2007, black women in Chicago died of breast cancer at a rate 62% higher than white women. And they found that this was due to structural racism, not biological differences, because similar findings were not true in other cities that had large African American populations like New York and San Francisco. And what did they find after they initiated this program? When they looked at the disparity in breast cancer mortality among African Americans in the ten largest cities with African Americans, they found that there was an increase between 1999 and 2013 everywhere except Chicago. Where this navigation system was initiated and they found that it wasn't just availability of care, but we really have to think about availability of quality of care, because that's really what's going to make the difference and can AI elevate the quality of mammographic interpretation consistently, both in the United States and globally. And that's what we're going to talk about today, and we know that in the United States, Black American women have a lower use of 3D mammography. The latest technology. There's a longer time from the time of diagnosis to the initiation of treatment. There are few ACR centers, American College of Radiology Centers of Excellence in underserved communities that are more frequently predominantly black and fewer black women in clinical trials so we can harness their protein. The genomic information to develop targeted therapy. And black women have higher recurrence rate as noted by their Oncotype DX. So what's the opportunity for AI? It has the opportunity in the potential to improve the accuracy and decrease the variability of the interpretations, both geographically and to bring general radiologists up to the level of performance of subspecialized breast imagers. Mammography has a substantial false negative rate, false positive rate, and perhaps a I can impact that as well. Can AI help us to triage cases by their complexity and to serve as a secondary there in the European trials or screening programs that require a second reader? And can I allow us to have the same number of radiologists, but who can do more in the same amount of time by decreasing the interpretation time? And what about standalone AI? Maybe not in the US, but what about in Africa? If most mammograms will never pass humanize. So that we might be able to really screen larger numbers of women. And what about risk assessment with that deep and rich genomic proteomic information in a womans mammogram? And can we actually move risk assessment from population based like we do in the risk assessment programs that we have now to oh specific womans risk assessment by what's in her individual mammogram? So briefly? AI is machine learning its computers to mimic human cognition. Machine learning uses statistical models that are trained on available datasets and deep learning is a process where the computer continues to learn based on what it's seen. It's ongoing learning and the computer continually teaches itself the feedback to improve. So let's ask, can AI help us with radiologist accuracy? So here's the study that was published looking at 242 D mammograms, 100 cancers, 40 false positives and 100 normals. They were interpreted by 14 radiologists, initially with out and then with AI using transpara by screen point and the outcomes were what you would expect. They were looking at the area under the curve. Specificity, sensitivity and reading time. And here you see the performance of the radiologist versus the performance of the radiologist plus AI. The yellow being unaided, the blue with AI support. And there was a statistically significant improvement in the area under the curve. When radiologists interpret mammograms with AI. Now this dashed blue line is the performance of transparent at the time that the study I just shared with you is done. But with time this is continually improving and this solid line is more indicative of the performance of this AI program currently. What about what about AI with 3D with Tomo? With tomosynthesis and again a study looking at 240 digital breast tomosynthesis that were read by without AI and with AI by 18 radiologists. And here you can see that the performance of the radiologists improved with AI so it improved in 2D and it improved in 3D. In another study by another group, you can see the same thing that not only was there improvement in performance or the area under the curve with AI, the sensitivity improved the specificity improved. The recall rate decreased, and the reading time decreased and you can see that there was a in this study, statistically significant increase in sensitivity and a trend towards improving. A positive improvement in specificity as well now being a radiologist, I can't do this talk without at least showing you one case and here it is where you can see a mammogram, where the AI transpara found a cancer that was missed by the majority of very highly trained breast imagers. So yes, and I can't improve the accuracy of interpretation of mammography, both 2D and 3D, and even experienced breast imagers had an improvement in sensitivity, but the improvement was greater with general radiologists. What about reading time and hearing a study using an AI using transparent? You can see that there was a marked reduction in reading time in six out of 10 breast imagers. And it was interesting that for the less complex cases, the improvement in time of interpretation was greater. So it turned out that on average it took 35 seconds to read the screening 3D mammogram with this AI. And other AI's have shown similar things in a substantial reduction in interpretation time, with AI of 52% by Doctor Conant, who reported her data. So yes, there is a significant reduction in time for interpretation of both 2D and 3D mammograms. The decrease in time for interpretation with AI is greatest for the least experienced radiologists, but even experienced radiologists have a reduction, and it's great that reduction in time is greatest with lower complexity cases, which is most of mammography which will allow to allow us to go through screening mammography even quicker and more availability of radiologist to interpret mammography. What about disparity in interpretation? So not everybody interprets the same. We know, and increasingly, learning that there is a lot of variability in interpretation. Most mammograms they netted States and globally are interpreted by general radiologists. This year there was a very interesting study that was published by Doctor Moy and her colleagues looking at the radiologist characteristics associated with performance of screening mammography using a large national database, and they found large variability in mammographic interpretation in the Midwest and West. Radiologists perform better than in the northeast. Women perform better than men. Academic radiologists performed better than non academic radiologists. Breast imagers performed substantially better than general radiologists. And those in practice, more time were performed better than those that were in practice less time. So we see a lot of image variability and then that it States and you can only imagine with the various global difference in. In a in interpretation and in training that that would even be more so. Now look at this. This is really quite extraordinary if you look at the acceptable cancer detection rate and the acceptable PV. Two recommendations for biopsy. You can see that only 77% of radiologists in this study were in the acceptable range for cancer detection rate and only a little more than half were in the acceptable range for recommendation of biopsy. So we have a lot of work to do, but if you look at this in terms of performance, of of mammographic interpretations by breast subspecialists in general radiologists, you can see a difference and you can see that general radiologists have increased performance with AI. But what's really interesting is that general radiologists with AI perform better than breast subspecialists without AI, and with AI there really is not a significant difference. Between the sensitivity of cancer detection for breast subspecialists and general radiologists so. Other studies have also shown this. This is a little busy, but if you look at the dark blue line, that's general radiologist performance unaided and the red line is. Breast subspecialist and the green line is performance of AI. So here in two different tests AI performed not only better alone, not only better than general radiologists, but better than breast subspecialists. And here you can see the with radiologists, who were in practice for less than ten years, or those more than ten years and AI helped improve their decision support and their performance, both with less experienced and more experienced radiologists. So can AI level the playing field looks like it really can. It can improve the performance of general radiologists to breast imagers, it can improve the performance of less experienced radiologists to more experienced. And it can improve the performance of all radiologists, regardless of how experienced or how long they've been in practice with mammographic interpretation. Now let's ask the question about AI in interval cancers and we have to find these cancers. These are the evil players. They have a higher stage at the time of diagnosis and a poor prognosis. And here in this study that again was published this year and they looked at the preceding mammogram of 429, screen detected cancers using AI, using two experienced radiologists who went back and looked at the prior mammogram and knew where to look and said Nope, there was nothing there. It's negative. There's a minimal sign of cancer, but it would not have been. Actionable, or it was a false negative and look what they found. They found that a I scored one in three of these cancers three years earlier and there were 67% had no or minimal signs of cancers, 19% of the interval cancers were visible as false negatives or with minimal signs and imaging, and they were all classified as ten as high risk. So if we use AI, we can in fact decrease the number of interval cancers. By detecting them air earlier and it seems and it should follow that a result of finding interval cancers earlier will result in improved patient survival. Can AI function is a double reader? Well, all European screening programs are most are double red, not so in the United States. But can the second reader be AI instead of a radiologist? And can we double or nearly double the availability of radiologist to interpret mammography? And the answer is yes. Here you can see a radiologist without AI, radiologist with AI, and you can see the improved performance with 19% fewer miss cancers on average in this study. And it you can see that one radiologist with 3D mammography versus 2 that there is a significant increase in performance when AI is used as the double reader, both with 3D and with two D. And when that happens, there's up to 18% fewer miss cancers. So yes, hey I can function as a double reader, and it can probably function as a single reader, which brings us to the next point about standalone AI. Now human resource is the most limited resources in Africa they can buy in many places mammography units, but they don't have anybody to interpret those mammograms. So could we set up a spoken wheel system where most mammograms would never be seen by a human because they have a transparent score? 1234567 and maybe. Only send those with eight, nine and 10 to a centralized mammography facility, and we could markedly expand the availability of MA F AI. So here you can see 101 radiologists, the AI alone, and the performance is better even with AI. And if you look at the sensitivity of a I-495, radiologists AI was higher. In 55, the radiologists were higher in 36, but it still shows that AI. Can function alone with a with a performance at least similar to radiologists alone. And again the performance of of AI is continuously improving as a result of this continual learning. And what about not reading other cancers? So here you can see normal cancers, the distribution and cancers that are case based. 789 or 10? What about it? Fix anything and any case that had a transparent score of 1 to six or seven or eight or even nine never passed human eyes. You could up to 80% of exams with cancer are transparent scores 8010 and So what if in Africa we could loosen our parameters and and really screen and find 87% of cancers that would really be incredible and in the screening population the group of. Cases that had as an AI score of 1 to 7, which was 70% of the screening volume, had a negative predictive value of 99.97%. That sounds pretty good to me. We would be able to screen so many women without that human resource. And maybe even in the United States, we can use this to help us with the decrease volume and maybe some cases don't need to be seen by human eyes at all. Now classifications are a real difficult one, but and we we biopsy too many benign calcifications, can we decrease the number of benign classifications using AI? The answer is probably yes. We could probably decrease by almost a quarter the number of biopsies we do for benign microcalcifications using AI. So in other studies a I had can also be used as a standalone without a radiologist with the performance equal to or better than a radiologist. Standalone AI, yes, it's better than a general radiologist. It might even be better than a breast imager, but it can certainly address the limited radiologist availability and maybe in a spoken wheel system in emerging countries to begin to screen larger populations of women. What about AI? In risk assessment we didn't even begin to address the enormous genomic information in a womans mammogram. There's so much information. What is that? Womans risk? Not what is it? Based on a risk assessment model and maybe we could stratify the interval of screening based on risk or. Use more limited resources, more wisely, smarter in order to give those women at increased risk the ability to get screamed. So the answer that I hope that I showed you and I wanted you to notice that somebody of the studies were published this year and last year the amount of information about AI is exponentially increasing. But even with the small amount of data that I shared with you today, yes a I can't. Increase the performance of general radiologists to that of a breast imager. It it can increase the performance of less experienced radiologists to more experienced radiologists. It can increase the performance of a breast imager to a higher level of performance. That is true in both 2D and 3D mammography. We can increase the availability of the human resource by less interpretation time and having the same radiologists being able to interpret more cases. It's highly effective as the second reader and in some countries in Europe is in fact used as a second reader. It decreases the number of interval cancers and the cost implications of finding cancers three years earlier at an earlier stage is very substantial and it's highly effective as a standalone technology for interpretation of most mammograms within a screening population. A negative predictive value of 99.7%. So maybe. Seventy 8090% of mammograms, particularly in countries that don't have any screening currently, can be screened without ever being seen by human eyes, and this really will have a marked impact on global health care disparities with breast cancer as well as in the United States where we can increase the performance of all radiologists to a level that will improve the care of women with breast cancer. And so the answer is, can artificial intelligence in mammography level the playing field? I think the answer is a resounding yes. We have to learn how to use it smartly, effectively, but I think it is really the key to impacting healthcare disparities in breast cancer in a very meaningful way. So with that I thank you. This is my email if I can answer any questions, please don't hesitate to reach out to me.

Can Al improve the accuracy of Radiology Radiologv Radiologvsts Radiologvst Radiologysts radiologist ORIGINAL RESEARCH • Cancer Control Cancer Cross Mark Estimated age-standardized incidence rates (World) in 2020, all cancers, both sexes, all ages Top cancer per country, estimated age-standardized incidence rates (World) in 2020, females, all ages Estimated age-standardized mortality rates (World) in 2020, breast, all ages What impact did Al have on 2D reading time City Patterns of Screening Mammography Uptake Example of a 3D mammogram where most radiologists only Radiologyst Radiologists 0 101010' 01010 • 1010' DOI (2021) 313940-5947 Cancer detection rate Transpara algorithm improvement over time Radiologists perform better with Transpara@ PPV2 Opportunities for Al in Mammography Artificial Intelligence On average 35 Reduction in Reading Time for Tomosynthesis Results: ROC Curves with detected the invasive cancer when using Transpara@ DBT Al Accuracy of Interpretations: Radiologist + Al Radiologist Characteristics Associated with Interpretive and Disparity across the United States Artificial Intelligence in Mammography: ORIGINAL PAPER •Acceptable region per 1000 •Acceptable region 2.50 per 1000 Impact of Al on interpretation of Breast Specialist and General Radiologist Radiologists perform better with Transpara mammographic interpretation in 3D •Acceptable region: 2040% seconds to read based on Transpara Score Accuracy of Interpretations: Change in Interpretation time with Al: DBT Health Disparities in the US: Breast Cancer Breast Cancer Health Disparaties Metropolitan Chicago Breast Cancer Task Global Health Disparities: Breast Cancer Impact on healthcare disparities: Time Why Al in Screening for Breast Cancer What is Al and how do we use it with Can Al "level the playing field" with Radiologist + Al Radiologists 2019 in 3D What about Stand alone Al Al and Interval Cancers Change in Sensitivity (DBT): Breast Exam-based Transpara Score Need to "level the playing field" Can Al decrease the number of Interval Al and Risk Assessment Al as a Double Reader Standalone Al s pport• s pport Performance of Screening Mammography: A National Artificial Intelligence in ARTICLE s ppo 2019 31. Transpara helps improve accuracy8 9 demonstrated with FDA reader study data from 2018 (2D) and 2019 (Tomo) Calcifications that were biopsied with a screening 3D and without Al CAD Slice 33 33 Slice 33 Slice 34 Up to Stand-Alone Artificial Intelligence for Breast Cancer Detection Up to 87% of with Transpara Al Can it level the playing field in Breast Cancer Can artificial intelligence reduce the interval cancer rate Mammography Database (NMD) Study Improving and Efficiency With Concurrent Changes in the racial disparity in breast cancer mortality Standalone performance also as good as radiologists in 3D exam using in mammography screening? in Mammography: Comparison With 101 Radiologists Improving accuracy and decrease variability of interpretations: Thank you Of Artificial Intelligence for Digital exams with cancer fewer missed disparility in interpretation mammographic interpretation with Mammography Mammography for Interpretation mammography in Mammography Force Specialists and General Radiologists Better than general radiologist improvement with General Radiologists Breast Cancers .1Racliologv 1k )logv Radiologvst Racli( v Radiologv )logv ORIGINAL RESEARCH BREAST IMAGING in the ten US cities with the largest African American populations MD.* ScD. PhD. PhD. Transpara Score Transpara Transpara• Transpara@ M SCO. P h O. PhD. ScD. PhD. PhD. Can artificial intelligence reduce the interval cancer rate transparæ Transpara Score Mammography: Alejandro Rodriguez-Ruiz, Kristina Ling. Albert Gubern•Merida, Mireille Broeders, Alejandro Rodriguez-Ruiz, Kristina Lång. Albert Gubern•Merida, Mireille Broeders, Alejandro Rodriguez-Ruiz, Kristina Lång. Albert Gubern-Merida, Mireille Broeders, t transpara• are Transpara with Transpara With Transpara transpara• T ranspara@ Transoara 2D Transpara Score Disparities WA cancers MASC. MD. and MO. MSEE Kristiru • Solveig Hofvind'•a • Alejandro Rodriguez-Ruiz s • Ingvar Andersson Al (test 1) (test 2) 23% of benign Variability in interpretation accuracy with different years of experience, Gisella Gennaro, Paola Clauser, Thomas H. Helbich, Margarita Chevalier, Tao Tan, Gisella Gennaro, Paola Clauser, Thomas H _ Helbich, Margarita Chevalier, Tao Tan, from 1999 to 2013: The reduction in breast cancer mortality Transpara Scores 9 or below were mostly 1-0 in mammography screening? Score 10 Lower use of 3D mammography in Black Women in the US Siemens Healthineers presents Thomas Mertelmeier. Matthew G. Wallis, Ingvar Andersson, Sophia Zackrisson, Detection of Breast Cancer with Mammography: Unaided With Transpara geographic differences, gender, sub-specialization biopsies of 1.0 120 with Transpara disparity in Chicago Ritse M. Mann, Ioannis Sechopoulos Increase in accuracy OR OK Reduced reading time for low suspicious mammography exams Can increase performance of general radiologist to that of a breast imager Effect of an Artificial Intelligence Support System MN Readers with DBT Al-CAD: calcifications Mammography is an excellent examination Black race generally correlates with socioeconomically Use machine learning/ computers to mimic human Leveling the Playing Field for Leveling the Playing Field of Difference in Performance Accuracy 140 Clear global disparity in mortality from breast cancer Higher reduction for low Equal Hope (Previously Metropolitan Chicago Breast Cancer Task Force) is a High false negative and false positive rates Most limited resource is human resource Kristina Lång • Solveig Hofvind3A • Alejandro Rodriguez-Ruizs • Ingvar Andersson Kristina Lång • Solveig Hofvind3Aa • Alejandro Rodriguez-Ruiz s • Ingvar Andersson More accurate mammographic interpretation (improve mammographic interpretation Al can improve the accuracy of interpretation of Black women have a 36% higher age-adjusted Longer time from diagnosis to initiation of treatment for Black 1.0 Table 4: Mean Sensitivity and Specificity across Radiologists INTERVAL CANCERS YES!!! Al can improve performance of General Al can decrease the number of Interval Cancers by European screening programs are double read Didn't even begin to address the enormous ND Detection of Breast Cancer with Mammography: Artificial Artificial Intelligence JNCIJ (2019) djy222 without Al with Al +0.03 +0.02 False negative Dominique • Anne Marie • • Garth • study design 3D DBT could be underprivileged benign. Study design 3D DBT cognition Can increase performance of less experienced radiologist to not-for-profit organization originally set up in 2008 after the publication of P=O.002 0.4 Using Decision Support boosts Minimal signs • David 40% reduction in breast cancer mortality over the past 2 decades in SIEMENS . SIEMENS Effect of an Artificial Intelligence Support System AUC t 5.7% 0.01) Using the latest deep learning Al Improve efficiency: Variable With Al Support Difference (Percentage Points) Value Una MS,- PhD MD MSc PhD MD PhD MD Women in the US Womplen in the US Radiologists in the Midwest and West >Northeast Reduction in time Significant reduction in time for interpretation of both DM Transpara Scores — read sensitivity and specificity) in less time breast cancer death rate than White women mammography avoided7 genomic information in a woman's mammogram detecting cancers earlier Radiologists to Breast Imagers What are the needs to address True negative more experienced radiologist more experience, radiologist more experienced radiologists Africa: Mammography units are not the rate limiting disturbing research that showed that Chicago had large and growing racial the US Most mammograms in the US and globally are Transpara Score reduced access to quality health care. Better than general radiologist Can the "second reader" be Al instead of a 83 851 3 83 851 [84, 3 Global Disparities 83 3 Average without Al Disparity in Breast Cancer mortality examined from 1999- Disparity in Breast Cancer mortality examined from 1999- 1999- 013 013 Machine Learning (ML) statistical models trained available data Transpara is continuously evolving Transpara is continualsly evolving performance 240 DBT exams read twice techniques Sensitivity t 8.4% (lesion level: p < 128 Without Al To triage cases by complexity With Al Benignant calcifications 06 In a screening population, the group of 1.0 77 (108/140) (75.791 [77.811 2 77 (108/140) [75.791 2 77 (108/140) [75.791 (77.811 2 77 (108/140) (75.791 2 in Mammograp:: winh Mammography in disparities in breast cancer mortality .046 .06 0.6 MD Performance Measure and DBT Fewer ACR Centers of Excellence in underserved communities step normal exams with Increase performance of a breast imager Female radiologists > male sets Difference Significant impact on mortality reduction from mammography -Al scored one in three (143/42 ) 95% Interestingly, the same is true for Hispanics in the US, yet Hispanics 0.06 0.086 0.08 0.082 Preceding mammograms of 429 consecutive H. MD • MD • MD. MD H. MIX MD • MD • MD. MD H. MD • MD • MD. 1-0% 00/0 -10% interpreted by general radiologists Harness the enormous genomic information in Breast cancer relative risk of death is 71% higher Both FFDM and DBT Most limited resource is human resource 1.0 0.01)' so that women and healthcare 0.01) 01010' EstimatedS age-standardized mortality rates (World) in 2020. breast. all ages Estimated age-standardized mortality rates (World) in 2020. breast. ages Al system Al System Al can improve performance of less experienced Stratify screening interval based on risk? ? Impact on improved patient survival radiologist (95% co in 10 US cities with largest AA population (95% Cl) 00/0 Transpara Score 1-7 (70% of screening Note.—-Breast Imaging Reporting and Data System category 3 or higher was used the recall threshold. Note.—Breast Imaging Reporting and Data System category 3 or higher was used the recall threshold. Reporting and Data System category or Was Reporting and Data System category or Was used ?better than breast imager by 18 MQSA qualified To serve as a 2nd reader in Europe 0.895 0.86 06 0.6 0.4 02 0.0 Higher stage (a—SS) with breast cancer have a prognosis that is better than non-Hispanic High-risk calcifications Malignant calcifications interval cancer with risk score O From 2005 to 2007, Black women in Chicago died of breast cancer at a rate Healthineers that are more often predominately Black communities Numbers in are data. numbers in brackets are confidence Numbers in parentheses are data. in brackets are confidence Numbers in are raw data. in brackets are confidence in parentheses are data. in brackets are confidence Detection and characterization of breast lesions with Imperfect examination Availability of Radiologists to interpret mammograms is 2D and 3D confidence 0.87 +0.03 0.6 volume) has a negative predictive value of providers can have peace of Independent of your level of 21 /24 improved sensitivity Academic Radiologists > Non-academic screen detected cancers analyzed with Transpara Decrease time for interpretation with Al greatest on least 96 Whites for non-Hispanic Blacks and 14% higher in mammography for accurate individualized risk assessment AUC standalone Al = 0.89 Can get machines- what about radiologists/ increased interpretation radiologists similar to more experienced Accuracy improves using 0.4 0.2 240 DM exams (100 cancers, 40 FPS, 100 normal) Decrease interpretation time 0.0 85.9% 62% higher than White women even though they were diagnosed less often. At 85.9% 812.7% 84.0% radiologists; once with and radiologists, once with and determination of likelihood of malignancy OK OR Rachel F. Brem, MD FACR FSBI Even experienced breast imagers have Stratify screening as a limited resource based on Addresses limited resource: Radiologists to Breast Subspecialists All increased except Chicago 99.97%. These can be confidently read (-1.0%, 9.4%) YES!!! ! Overall sensitivity: 85% 0.0 1-0 v-, 0.4 0.8 0.9 0.47 Decreases interpretation time (74.1%.89.3%) ( no difference of average of readers and Al) Fewer Black Women in clinical trials Poorer Prognosis (76.0%, 92.0%) (79.8%, 92.1%) Performance of Radiologist experience, radiologist experienced radiologists Specificity t 6.9% (p <0.01) -67% (96/143) were either classi ed mind that they are always Could we set up a spoke and wheel system where experience, radiologists' more experienced radiologist the same time, there were only very low breast cancer mortality disparities in Leveling the Playing Field for -15 Transpara as a second pair of Standalone Al General Radiologists Sensitivity Hispanics How do we leverage technology to address this critical need 140 Deep Learning: Uses enormous computational power with Radiologists radiologist radiologists faster because it is very likely normal. Breast imagers > General Radiologists improvement with General Radiologists once without use of 0-2 04 Radiologists 2 experienced radiologists classified the cases (in 1.0 improvement in sensitivity but with greater Increase the availability of high quality mammography/ 1.0 risk? Interpreted by 14 radiologists with and without Al interpret mammograms -10 as minimal signs or false negativ versus Radiologist + Al New York and San Francisco Up to 50% of breast cancers can be "missed" in women with dense > 0.6 20.6 1000) Highly effective as a second reader rbrem@mfa.gwu.edu getting the best second pair of 64 0.4 (n Global Disparities accuracy improves using most mammograms would never be interpreted by Substantially decreased Metropolitan Chicago Breast Cancer Task Force (2008) Metropolitan Chicago Breast Cancer Task Force (2008)ished Metropolitan Chicago Breast Cancer Task Metropolitan Chicago Breast Cancer Task Force established hundreds of layers of parameters All subspecialist improved specificity Genomic medicine far-reaching consequences General Radiologists Average radiologists experienced radiologists AUC (0.787-0.841) +12.5% 712.5% 812.7% ASR (world) per 100 000 Complex in the US Sensitivity improvement was statistically significant (4.7%, 20.4%) 120 20 0.0 breast tissue Average of radiologists experienced radiologists How do we optimize radiologist interpretation (59.7%, 83.2%) T transpara@ Transpara@ transpara• are Transpara — Transpara reading DBT 11-25 years in practice > 10 or less years consensus) as negative, minimal signs of cancer or ASR (world) per ASR (World) per improvement with General Radiologists mammographic interpretation in the US and globally to Decreased time for interpretation greatest with lower Radiologist (n — 95) ff. per This suggested to the founders of Equal Hope that instead of biological 0.896 0.89 eyes available. Decreases Interval Cancers Survival and Cost implications of Transpara reading DBT — Transpara (original publication) Transpara. Wit Transpara Specificity improvement trended positively a radiologist, and have a centralized regional Transpara (Screenpoint) Al provides lesion, case-level scores (1-10) Radiologists 20217 -19% of the interval cancers visible Radiologists reading DBT 2257.1 Spoke and wheel system in emerging countries — AUC (0.820-0.860) unaided (0.866) Initiated due to breast cancer health disparaties Recall 7.2% 0.01) @ Siemens Healthcare GmbH, 2021 Ongoing learning Enormous opportunity in emerging countries Rachel F. Brem, MD FACR FSBI Mortality reduction unequal among differing populations Recent study demonstrating significantly higher recurrence rate with only O Radiologists reading DBT unaided With 0.01)' 0.8 0.89 Metropolitan Chicago Breast Cancer Task Force established Metropolitan Chicago Breast Cancer Task t transpara• Increased reading time with Al support Transpara (today) Transpara Score General Radiologists 100% 0.2 0.0 0.6 06 Average radiologists Average radiologist General Radiologists HI 188.8-257.1 differences, this health inequity was driven by structural racism. In other -15 Breast (158) detecting cancers 3 years earlier 80 1.0 100% 60 AUC Rad unaided = 0.87 -10% -20% 32 0.2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 20 For CDR2 percentage of radiologists with acceptable performance was 77.0% 1.0 false negative —e— Breast specialist (individual) Differential in world-wide training/specialization Stand-alone Al system (0.887) complexity cases with 2D as false negative or with minimal signs improve outcomes from breast cancer 3. Differences in sensitivity the (Al) 3. Differences in sensitivity the intelligence (Al) center for those with suspicious mammograms 14 individual radiologists 140.4-188.8 Oncotype DX in Black Women 16.6-19.4 Cervix uteri (23) Mortimer K. Conant E et al https://doi.org/10.1016/j.jacr.2020.12.033 Quality of care especially mammography varies a lot in Chicago Director, Breast Imaging and Intervention Risk assessment- "Precision Screening' Reader Number 1.0 1.0 1.0 The computer continually teaches itself via continuous feedback 0.46 0.8 0.8 0.69 1.0 Reading time 1.0 Mortality reduction in US Black Women 40% less than Caucasian Women Mortality reduction in UJS Black Women 40% less than Caucasian Women 0.4 — Breast specialist (average) 52.7% (p s 0.01) 52.7% (p < 0.01) 52.7% (p 0.01) Outcomes: AUC, Specificity, Sensitivity, reading time 0.8 0.0 100% 0.0 10 1.0% -10 words, Chicago's inequitable healthcare system was to blame Rad with Al = 0.89 3D The statements by Siemens Healthineers' customers described herein are based on results that were 8A. RodriRuez-Ruiz. E. Krupinski J. Mordang, K. Schillinw S. Hevwang- 8A. Rodriguez-Ruiz, E. Krupinski J. Mordang, K. Schillinw S. Hevwang- A. Rodriguez-Ruiz, E. Krupinski, J. Mordanq, K. Schilling, S. Hevwang- A. Rodriguez-Ruiz, E. Krupinski, J. Mordang, K. Schilling, S. Hevwang- tem and each at the of each considering tem and each at the Of each considering on imaging were classified a high 111.9-140.4 111.9-140.4 Not applicable Not applicable 1.0 0.8 1.0 0.2 0.9 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Non-melanoma skin cancer (2) 1 - Specificity Highly effective for stand alone interpretation of most mammograms radiologist (individual) 1 - Specificity Radioloøst 10 years ex*rierxe Radioloøst 10 years ex*nerxe Radiologist 2 10 years experknce Increase not just amount but QUALITY of care to AA women in Chic, Increase not just amount but QUALITY of care to AA women in Chic Increase not just amount but QUALITY of care to AA women in Chic. -20 to improve Kobrunner, l. Sechopoulos, R. Mann. Detection of breast cancer I - Specificity Kobrunner. l. Sechopoulos R. Mann. Detection of breast cancer determined by Al evaluation of mammograms? 20.8 1.0 20.0 MDS three and Over recall. — Breast Imaging and MDS three and Over — Breast Imaging and AUC Rad unaided How do we answer the critical radiologist/ breast Professor of Radiology 0.4 0.8 0.0 €111.9 Thyroid (1) Conant et al. doi: 10.1148/ryai.2019180096 Variation and fragmentation of care in particular on Chicago's south No data Not = 0.87 -400/0 Not applicable — General (average) I - Specificity 1 - Specificity Not every woman has access to mammography achieved in the customer's unique setting. Because there is no "typical" hospital or laboratory and many P = .002 Harness enormous genomic/proteomic data available in mammograms for Reading time reduced by 52.7% (p < 0.01 )With DBT using mammographv: Impact of an Artificial Intelligence support C. Balta. N. Janssen. A. Rodriguez-Ruiz. C. Mieskes. N. Karssemeiier. S. H. Hevwang- risk with Transpara score 1 using mammographv: Impactof an Artificial Intelligence support More availability of radiologists to interpret mammograms 10 1.0 -10 a. e,. -10 1-0 -19% fewer missed cancers 4. No data Data System. unaided (0.866) -19% fewer missed cancers on average Liver (l) 1 - Specificity side, is likely to affect stage of diagnosis and survival system. Radiology 2019;290 305-314. https:,'/doi.org/10.1148/ Köbrunner. Can Al help to increase the PPV of screen-recalled biopsies on In a screening population has a NPV of 99.7% In a screening population, the group of I -Specificity I-Specificity 201+290 305-314. variables exist (e.g., hospital size, samples mix, case mix, level of IT and/or automation adoption) there 6. AUC Rad with Al = 8. 4. individualized risk assessment Kim E, Moy L, Gao Y, Hartwell CA, Babb JS, Heller SL. City Patterns of Screening Mammography 3700010B 3700010B 37000108 37000108 37000108 0.8 0-8 0.2 0.8 02 0.2 1.0 1-0 0-2 Transpara Score transparæ transparæ@ 04 The George Washington University 160 20 0.0 1-0 0.4 t. Receiver characteristic the Receiver characteristic the •Internal imager shortage Transpara Score calcifications? Presented RSNA2020 Screen-detected cancers ASR Normal cases A' The and the Of the do not the of any radiol.2018181371 A' The and the Of the do not the o f any The and the Of the do not the any A' The and the the do not the o f any The and the the do not the any h" The and the the do not the any Can Al elevate the quality of mammographic interpretation with Al support (0.886) source: FDA approval number Kl 93229 FDA approval number Kl 932291 0.4 20.4 0.8 0.2 20190430 20190430 20190430 Mortimer K, Conant E et al https://doi.org/10.1016/j.jacr.2020.12.033 Mortimer K, Conant E et al https://doi.org/10.1016/j.iacr.2020.12.033 https://pubs.rsna.org/doi/pdf/10.1148/radiol.2018181371 https:/[pubs.rsna.org/doi/pdf/10.1148/radiol.2018181371 can be no guarantee that other customers will achieve the same results. ifici I-S ifici on the put o f t he Wo H on the O f or on the put o f t wo on the o f or on the O f t he W O r" on the Of or on the put o f t he wo H on the Of or on the o f t W on the Of or on the o f t We wo H on the Of or on the o f t he wo H on the Of or on the o f t he H on the Of or Uptake and Disparity across the United States. Radiology. 2019 Oct;293(I):151-157 SIEMENS and the artificial intelligence (Al) system in SIEMENS SIEMENS 9 FDA approval number K193229 Data held on file FDA approval number p = .002 A" rights The designations employed and the presentation ofthe this do not imp* the expression of any opinion whatsoever A" rights reserved. The designations employed and the presentation ofthe this publication do not imp* the expression of any opinion whatsoever A" rights The designations employed and the presentation ofthe in this publication do not imp*y the expression of any opinion whatsoever A" rights The designations employed and the presentation ofthe this publication do not imp* the expression of any opinion whatsoever FDA approval number K193229 Data GLOBOCAN 2020 Data source: GLOBOCAN 2020 Data 2020 GLOBOCAN 2020 Never have human eyes interpret 70%, 80%, maybe 90% of mammograms globally False Positive Fraction Lee CS, Moy L, Hughes D, Golden D, Bharganvan-Chatfield M, Hemingway J, Geras A, Duszak R, Rosenkrantz AB. Radiologist Characteristics Associated Conant et al. doi: 10.1148/ryai.2019180096 Al risk score Move forward from "population information" for risk assessment 1.0 (http://qco today) today) or of or the Of on or of or the Of or on or of the of on or of or the of on or of the on or of the of or on on the part of the Health Organization / International Agency for Research on the legal status of any city or area Of Of Co-Founder and Medical Director, The Brem Foundation Co-Founder and Medical Director, The Brem Foundatio on the part of the World Health Organization International Agency for Research on the legal status of any territory. city or area on the part of the World Health Organization international Agency for Research on Cancer concerning the legal status of any country. territory. city or area on the part of the World Health Organization International Agency for Research on Cancer the legal status of any country. territory. city or area on the part of the World Health Organization / International Agency for Research on concerning the legal status of any country. territory. city or area Graph production: (Outcomes by Race in Breast Cancer Screening With Digital Breast Tomosynthesis Versus Digital Ma O, Which may not yet be Which may yet be with Interpretive Performance of Screening Mammography: A National Mammography Database. Radiology. 2021;300:518-528 consistently both in the US and globally 0.89 0.49 the on md on (hitp:/ today) (http://gro.iarc.fr/today) (http:/'geo.iarc.fr/today) today) or of its authorities. or concerning the delimitation of its frontiers or boundaries. Dotted and dashed lines on maps represent approximate borderlines for or of its authorities, or concerning the delimitation of its frontiers or boundaries. Dotted and dashed lines on maps represent approximate borderlines for on the AUC the SC Conant et al. doi: Conant et al. doi: 10.1148/ryai.2019180096 transparæ transparæ• transparæw Impact on regions of the world with markedly limited human resources SCREENPOINT transpara• Transpara. transparæ. Transpara@ sc sq t transpara• Transpara With Transpara not b. b. Which there may be agreement. Which there not yet be full agreement. Which there may not yet be full agreement. World Health Organization World H Organization 4. I -Specificity I - Specificity https://doi.org/10.1038/s41586-019-1799-6 Sieniek, M. , Godbole, V. et al. McKinney, S.M., McKinney, S.M., 0 Medical Medical C' Medial 0 Medical 02021 ScrænPoint 02021 ScreenPoint 02021 ScrenPoint 02021 Screenpoint