CAD for Breast Cancer Screening

So far there is no defined role for CAD in mammography screening, at least not in the screening programs which also use a double reading. And of course, there's also no reimbursement, and this influences one another, of course, so there is no incentive for using it in screening programs. However, looking at the future with more 3D data with more images to be read. I think there will be an increasing need to use the cat cat systems. Also there is an increasing shortage of experienced radiologists and experienced screen readers and for that purpose it will be important to use or to test cat programs. We've had the chance of testing 2 very different programs or also in very different settings. The first one was just 2D. Cat system this is from screen point Transpara and we integrated that into our screening practice. Our screening practices. Mere Trudy screening is not allowed to do thermal synthesis for screening in the screening program and the screening program has mandatory double reading. An triple reading is requested whenever we have a new reader and when you can do it. Optional for supervising. Two readers, for example, if one reader is weaker or whatever, and the factors of course it would be nice when you do a triple reading, which is a lot of work to choose the cases with a higher probability of malignancy. So we found it quite interesting. This screening packs offers a rating and or score, and the score is given to a population in a way that score score one is the lowest probability of malignancy. Score 10 the highest. Probability of malignancy and to distribute among the population with every 10% of the population having score 1234 and so on. So what we did, we connected the screening packs to our risk system. The screening risk system we obtained the scores which were given to the cases, and we sent all the cases from our screening for about 10 months to this cat and the cat transmitted the numbers to the rest system. This was only made visible through the 3rd reader, so the first reader in the second reader were not influenced by this rating. They could also couldn't see the marks for the supervising reader an for the consensus reading. We could see the Marks and yeah, so we could integrate that into our screening and. This is how it looked like. This is just the list of patients we get during supervision. Here the patients are erased here an you have the score and then you can sort by score. So I started reading the pens and then the lines and afterwards I stopped because I didn't want to read to triple read 18,000 cases. That was the number of cases coming in within 10 months. About 18,000 cases were still have some exclusions the exclusions concern. Patients who are not yet completely documented or who didn't show up for the recommended examinations altogether. We ended up with 15,000 about 15,500 cases without a consensus, and we had 2551 cases with a consensus and they contained altogether 125 cancers 118 and one breast and seven in the contralateral breast. We gave a rating. We evaluated per patient, but the few patients with two cancers, one on one, one on the other side side, we gave two ratings. The results are presently only verified by the result of the consensus reading when they were normal cases or by the result of the imaging and histopathologic assessment. When reading the 18,000 cases, the cat for the triple reading would have assigned A9 or A10 as predictable. Yeah, in about 20% of the patients, exactly 20.4% of the patients. This positive reading was only due to the cat in 2901 cases, which is 16% among the nines and 10s. The others were the positives made positive by the readers, yeah? So if a triple reader you want to read these 20% of cases, yeah, then you can, you know change your or make a triple reading. Yeah I was not able to read all of these cases these 3000, but I read just consecutively 1600 cases and in these one 600,600 triple reading cases I got an additional 6 recall. So this is a very low number which you decide to change. They consisted of four benign readings which didn't change the specificity of our reading, so it was almost stayed the same, but we also found two more malignancies, so this is something which we detected even by only reading half of the nines and 10s with us increased our sensitivity from 92 to 93.6%. If we had read all the cases, it might have gone up to 95 or 6%. Which are the Kansas which we couldn't find at all? They were two cans is actually, which was only palpable. One cancer only detected sorry to only palpable cancers 3 Kansas, which were detected in the other side and the contralateral side during assessment. And two Kansas which were just detected at preoperative MRI. So you don't expect that a care finds a cancer. Which you can see on the mammogram, and not on the term. Oh yeah, so anyway. So that's just that's why we end up few percent lower than 100%. Now we looked at the single reading performance. You know, we just thought could can be used as a single reader. For example, in a double reading setting. Yeah, and here we found that the CAD achieved the sensitivity of almost 80%. Same is true for reader one. Same is true for reader to, so the sensitivity is quite compareable. The fact was the false positive rate was definitely higher for the cat than for the readers. However, because you had to read the 20%, but after reading them and after consensus it appeared quite compareable. So in order to see whether the care provides different information than reader one or reader two, we just try to combination and with the combination reader one and two we have this combination because whenever reader one or ever reader 2 red, we had a consensus and the consensus yielded a sensitivity of 92%. Now we played a little bit and we said that if the cat finds something or reader one, this is considered. Positive and or decision. So this is the maximum sensitivity we could achieve with cared and reader one and it's 87.2% slightly lower. Not statistically significant with Reader 2 is exactly the same sensitivity, so we're approaching state where a cat could be considered as a double reader, yeah? And it could be considered as a double reader for reading the 20% cases, not the 80% cases, yeah? Where are remaining problems? I mean cared as you can see, has an 80% sensitivity, so it has its good side and it has its drawbacks. These are two cases with Microcalcifications which were missed, but to readers just catch constantly. Does the same whenever it detects it is not diverted by anything else, was just care to detect it. Here are two groups of Microcalcifications. This is actually rare. The cat was very sensitive for the calcifications but they were not correctly recognized. Their wardite. ACTED but the score was too low. It was just the score of seven. Generally, can systems obviously do have more problems with masses and they have more problems with just a symmetries and they have more problems in very dense breast tissue or sometimes with architectural distortions in very dense breast tissue. Here you see this is just bump here. It was also just called by one reader. The cat did not recognize it, instead it called here the powder on the skin. Not very helpful, but we have this with readers as well. And here this is an asymmetry. That's the only case where I say it's not really understandable that it didn't detect it. Obviously it considered that an asymmetry, but yeah, this is something no cat will probably deliver presently, because this is just recognition by comparison. The readers had the 2D in they could compare before and after, so the priors and you can see there is a faint abnormality which appeared here. It's not characteristic, so how should the cat detected? We just detected? Because it's new and here you see, we also have a dynamic evaluation in mammography screening. It's not just unique contrast agent for everything. Yeah, it's just whenever something is new, so that would be an interesting field for any cat system to learn. Of course, you think can we do further steps? Could it get better with 3D? And actually we have the chance of evaluating a 3D CAD? Of course the setting was different. We were able to give 386 cases to the cat system. They were consecutive cases from screening assessment and all the cases also obtained term a synthesis during screening assessment and now we compared the sensitivity and specificity of the reading before these patients were assessed. Reader one and reader two. They of course had. 2D mammograms, two planes and the priors. And we compared it to the term or reading of the cat only. I mean it's just playing games, but you get a feeling of what you would expect in this patient selection. And in fact it is really amazing. The sensitivity is really compareable, you have reader one and reader 2 and you have cat which has two different thresholds. The difference is not much, it is just one more. Cancer detected with the more sensitive threshold and the sensitivity is exactly between both. It's very good with 85% and the false positive rate. In this setting. I mean they were all considered positive by consensus. Was really amazingly good because reader one had a false positive rate of 41%. Reader 2 of 67 and the cat between 26 and 43 depending. On the false positive rate, so it was lower. Yeah, then you would expect it from the readers, so this gives a good indication. And obviously we are getting really additional information from this 3D data set which we don't have from the 2D data set, even with comparison with priors. Again, I provide a combination, you know, and the combination always shows whether the information. Matches yeah and here again we can see the information that is matched in mammography screening is reader one and reader 2 and together we gotta sensitivity of 97.7% in these patients. And if you combine the CAD with reader one or with reader two, you also got 97.7%. So this is really good basis for counter checking how this cat for example, will do in asymptomatic patients. Or on a screenings setting, of course we didn't have this possibility. Here again, you can look at the cases. This was a case priors and present films the readers did not notice that here very faint architectural distortion is starting to come up here. Very faint. And the term mode detects this architectural distortion in the cat correctly also detects the architectural distortion. So this is a different information which really helps. Similar setting, difficult case only one reader called the case because of microcalcification. The CAD called because of microcalcification and this architectural distortion, which in retrospect can be seen on the mammogram but wasn't seen prospectively by the two readers. So this is also quite good. The cat also detects false positives, as you can see here. This is of course the cancer here is selected another area dens little module. It could be something. Why didn't the reader call it? Because it existed for years was a small nodule. Yeah, there are difficult areas even for the 3D CAD and this concerns architectural distortions. Here you see an architectural distortion up here, which was seen by the readers. Nothing was seen on the right breast. Other cat correctly detects this even though it is really within an ACR for area and however, when you did the term oh and went through the term or you could see a very small architectural distortion here which was not detected by the cat. So there are limitations to everything and this was only seen during assessment because we didn't have the term. Oh, there are two other fields where even the use reddiquette has problems. These are well circumscribed around lesions because they look more benign. Yeah, and here the readers have the advantage of making a comparison of priors and present films. Here you can see it is visible on the cat, but the cat under estimated that here you see again this problem multiple multiple nodules. This was the only new module which was then correctly recognized by the readers, but not by the cat. But this is something. Where you really have extremely high demands to a cat system. Here again, architectural distortion when you are really honest, the architectural distortion, or actually the asymmetry is less obvious on the term role. I mean, you could even misjudge that on the term low as compared to the mammographic normals. And this was also not detected by the cat. Overall, I think the results are good, cared is becoming compara bulto the readers the need for cat support is increasing. I think we don't have to discuss that. Cat capabilities seem to increase from 2D to 3D quite significantly when you work with a cat, you have to try and understand it. I mean, the reader also doesn't is not always in the best forms. It's the same with cat. Depending on the case it gets. Yeah, overall it has high reproducibility, high reliability. It is not distracted by anything else like we just could be. However, it also has limitations like some readers. With very faint architectural distortions in very dense breast, it has problems with a symmetries that something you can easily check for rounded masses. Be aware to counter check with prior films and of course presently it still doesn't have a comparison with priors if you have the comparison with priors, you really have that piece of information that's lacking because then you see something gross you don't need contrast agent then which also has its limitations, yeah. OK, there still exist possibilities for further improvement and. Presently, I think it is evolving to become able to do double reading or double reading in a certain number of cases. I mean whether you want to do a triple reading. This is always an individual decision. How many cases do you want to read in order to have which yield? So thank you for your attention. I think future stays interesting.

CAD for Breast cancer screening: first experiences Result of single readers and CAD Transpara 1: results of selected triple reading Transpara (Screenpoint integration) First experiences with 3D-CAD 2D-CAD combinations 3D CAD CAD for with 2D-CAD integration for screen-reading of 2D (86 primary cancers and random104 benign cases) 2D-CAD: 3D CAD 2D-CAD: Case 3: Case 7: 3D CAD Summary 75 Present state masses Score 9+10: first experiences with a CAD mammograms in 3678 benign cases/18029 cases (20,4%) in 3678 benign cases/18029 cases The need for CAD support increases strongly with the workload of images CAD read2 obtained using 3D volume imaging 3D CAD by iCAD based on deep learning (Al) case CAD-tomo 27 positive reading onlv due to CAD (RI and R2 were normal/benign): EUROPEAN SOCIETY OF RADIOLOGY CAD capabilities (lower fps-rate) have improved noticeably from 2D to 3D! read I-DM read 1-DM read I-2 read 2 - DM read I-DM SENSITIVITY 123%-thresh read 1-DM read 1+2 read I-DM Read 1 Read 2 read 2-DM read 1 read I-2 CAD+read1 CAD+read2 2901 cases with a benign result by RI/R2 2901 cases with a benign result by RI/R2 =16.1% EUROPEAN SOCIETY OF RADIOLOGY Try to understand your CAD: SIEMENS .. SIEMENS screening-assessment German screening program: 98 25 95 tpos 75 71 74 15 84 75 74 75 12 71 74 79 true pos Geb.Oat no defined role in scr-programs double reading - high reproducibility, no distraction (for achievable level) Mandatory double reading (86 primary cancers, 300 benign cases and 36 add. mal foci -identifiable) 3D CAD f Analysis of 1600 cases (RI or R2 or CAD pos): double vs triple reading Possible triple reading by supervision (new readers or random check) fneg 11 15 71 15 12 - limitations: (faint architectural distortion in very dense tissue) 04.08.1964 84 19.02.20 9 19.02.z0 9 84 74 75 74 84 true pos re-embursement Supervision (third reading) mandatory for „new readers" (3000 studies) asymmetry! 29.12-1959 Healthineers sensitivity 67,3% 86,0% 87,2% 86,9% 97,7% 97,3% 91,9% 26,9% 97,2% 41,3% 97,7% 86,9% 82,6% 87,2% 87,2%6 6 additional recalls '6 additional recalls and optional for random checks rounded or faint masses Heywang•K. 4 benign recalls 6 additional recalls 2 analyses: fneg lacking comparison with priors! CAD-integration for „random" triple reading: 2 malignancies CAD 12% CAD 26% CAD 26% Mammamanager (RIS-system) CAD 12% CAD+RI Shortage of experienced screen readers in Europe Results (Jan 20-Nov 15, 2018) Results (Jan 20-Nov 15t 2018) - Comparison of sensitivity/specificity of 2D-screening reading (2D and priors) FPOS-RATE threshold Read 1 Read 2 read 1-DM read I-DM Read 1 Transpara (Screenpoint) assigns a probablility-score of malignancy (1-10) fneg 10 10 16 16 Possibilities of further improvement o. overall sensitivity increases slightly (from 92.0% to 93.6%) 18029 screening studies were screen read 2x and obtained a Transpara-score Increasing load of images to be read (tomosynthesis) sensitivity 91,9% 41,3% 87,2% 87,0% 97,7% 86,0% 97,7% (2D-CAD based on Al —deep learning) versus 3D- CAD-reading (based on I-view tomo) 73,196 73,1% 58, 7% 32,7% % true neg 115 115 (I cancer only detected by supervision, not by CAD (I cancer only detected by Fupervision, not by CAD Use for double reading is evolving (48 patients noshow or refusal and 15 yet incompletely documentation 63 exclusions) R 15137 Obvious missed by 2 readers 10-point scale: low supicion and high suspicion variables exist (e.g., hospital size, samples mix, case mix, level of IT and/or automation adoption) there can 87,2% K15019 fine missed by 2 readers fine missed by 2 readers for primary cancers / 105 consecutive benign cases 15478 cases without consensus (first and second reading normal/benign). 2 only palpable cancers, 3 detected at assessment, 2 at preop MRI) 2 only palpable cancers, 3 detested at assessment, 2 at preop MRI) triple reading possible (advantage vs added workload) % false pos 43,3% 41,3% 26,9% This score is transferred to our RIS-case-list of „studies to be supervised" SIEMENS .. 2551 cases with consensus (118+7 contralateral malignancies) 2551 cases with consensås (118+7 contralateral malignancies) specificity double vs triple reading unchanged (first and scond readers are blinded) SIEMENS Per patient 1 rating (2 ratings in 7 cases of bi-lateral malignancy) Per patient 1 råting (2 ratings in 7 cases of bi-lateral malignancy) Per patient 1 Åating (2 ratings in 7 cases of bi-lateral malignancy) Supervision is performed on one computer, to which a pad is connected that displays Verifications by consensus (neg) , by imaging or histo-p assessment) the marks collaborations. E pleomorhic mc detected but understimated (7) SIEMENS .. Case: 1003 microcals detected by one reader; CAD detects RI 5021: POS for benign Of It POS for benign Of breast; It fr,eg Healthineers • Healthineers mass and mc behind inverted nipple: nicht erkannt detected