Horizon Europe Info Days 2021 | Euratom | Session I - AI and developments in nuclear medicine

[Music] yes good afternoon everyone and welcome to uh the first session of this afternoon of the earth info days um which will be about artist intelligence and applications in nuclear medicine and we have uh two very interesting speakers um first of all we will hear professor habib zaider and then dr al makota so professor zaida is a chief physicist and head of the pet instrumentation and neuroimaging lab at geneva university hospital but he is also a faculty member at the medical school of geneva university he is also a professor of medical physics at the university of groningen in the netherlands in a junk professor of medical physics and modular imaging at the university of southern denmark but also a chunk professor of medical physics at beshte university and business professor at tehran university of medical sciences so he is very active on activities involving imaging solutions for cutting-edge interdisciplinary biomedical research clinical diagnosis and also of course lecturing undergraduate and postgraduate courses on medical physics and medical imaging and i would like to welcome professor but the floors is yours thank you thank you so much can i have the first flight please yeah just my name is habib zaidi not zayida and this is my first light thank you so much yeah so i've been given 15 minutes actually to summarize actually the innovations in the applications of artificial intelligence in the field of nuclear medicine and i think this is very limited time actually to cover all the developments in the field so just start by sitting on the scene next slide please and uh tell you that ai is is not actually so just fancy sometimes it's technical developments limited to academia but it's probably going to have a very huge impact actually in clinical practice and they suggested an initiative by colleagues from the european federation of the organizations of medical physics actually published a paper in physical medica highlighting actually the need to expand the medical physicist curriculum and medical physicist by training some nuclear medical physicists and we believe that er is going to have a big impact on our profession next slide please and you'll see that it's not only limited to europe but also um there are a lot of discussions actually taking place around the globe including actually in the united states usually the um the flow comes from the united states but with some exceptions in this case so we started the discussions in europe and there are also some ongoing developments in the united states and many other countries around the globe so ai is going to have an impact of your cm nuclear medicine and let's just give you let me just give you uh an overview of the basics of nickel medicine imaging with the next slide please so we start with the pets here so opacity is a hybrid imaging modality a component basically position emission tomography scanner with an x-ray series camera so what you see on the left side is uh um in the front of the scanner is basically an x-ray tube and the x-ray detector is on the other side and then you see that we basically start with the performing what we call planar scouts scan of the patients to get what you see on the right on the left side and the bottom and then we decide about the axial field of view that we are going to scan and then we start rotating the x-ray tube and the detectors around the patient to get the spinal ct scan and this power ct scan basically is going to be used actually for generation correction and also for anatomical mapping obviously the patient is already injected so we just need to push the patient to the other side if you look click on the movie on the right side we can see the patients actually been scanned on the pet's detector so the patient is radioactive because we already injected the patients with one of the molecular imaging probes that we use in uh position emission uh tomography and then we basically so um a fuse the pets and the x-ray city scan and then provide this to the md so that they can uh basically so uh reports on the status of the patients or plan the therapy or or reports also on the response to treatment and so on so next slide please so this is just one one on one side of the story and the second that the next side is combining patch with mri so there are a lot of developments in this area as well you see at least there are four scanning manufacturers they're coming with the instrumentation dedicated for combined spectrum or imaging so we're lucky to get the first one in europe they about 10 years ago so we saw the philips beta mri system in our facility you see so it's like a la pet city it's a sequential system where we start with scanning the patients on the mri and then we push the patients on the uh detectors to get the pet's image and then refuse those two images together and provide these to the mds so a couple of other manufacturers basically siemens and g so came up with the uh kind of solid state detectors which are not compatible so we can acquire the images sequentially instead of simultaneously sorry instead of sequentially and then more recently so another manufacturer so called the united imaging based in shanghai and china uh came up also with the recent system which is now commercially available since they got the c mark and also the fda approval so just to see some images on the next slide please you see some representative images of the brain so basically mri fuse it with the um within fdg pets image so fdg is the most widely used tracer in nuclear medicine so basically produce images corresponding to like glucose consumption in biological tissues and then in the bottom you see basically the myocardium uh with the gated image so showing basically the myocardium against an ftg path image fuse it on the mri scan just to show you what we can get in terms of images from those uh instruments the next slide so we start actually so talking about the applications of ei and equal medicine just going to give you a flavor so some works actually published by our group i'll just give you four examples so the first one is basically how to reduce the activity administrator to the patients or how to reduce the radiation dose of the patients using ai techniques so what we did in this work here is next slide so we basically acquired the data fdg brain fat data in list mode formats and then we decimated so the standard those projections to low-dose projections are basically taking only five percent of the acquired accounts for the standard of those and then we train the network using a hundred data sets hundred clinical studies to uh uh teach the system how to convert the low-dose scan to the full dose can obviously so the aim is to reduce the activity administration to the patients or to reduce also the radiation those delivered to the patients and we did two implementations basically so two models with the training performance in the image space what you see on the left side and the training performance in the projection or in the synogram space so basically what you see on the right side so without further details they're just to show you what's what we get so next slide please and you see basically this is the standard full dose of skin with injection to the patients with 200 megabacro of ftg and scanning the patients for 20 minutes and then you see basically with the five percent so the low dose is obviously very noisy not of standard clinical quality that you are used to and then in the next slide you see the outcome of the ai based solutions as i explain it so which was already trained to produce the high dose from the low dose scan so basically what we in in this case we're not using the full dose scan anymore so only the artificial intelligence network which is used to produce the images that you see here and obviously so we perform two implementations as i mentioned so previously and uh the conclusion is that the deep learning implementation in the projection space outperformed the uh implementation in the uh in the image space as you see also on the buys maps on the right side the next slide if you click just on the mouse so you see the bias maps here with the high bias in the uh predictions and image space compared to the predictions in the synogram space and then in the next slide so we perform at kind of qualitative assessment so we asked a couple of neutral medicine physicians having long experience in the field who basically ranked the images that we produced with the neural networks and you see obviously that the local scans go to very low scores 1.4 on average compared to the implementation projection space with an average score of 2.8 and the predictions actually by the ai business versions and the projection space was very close to the scores assigned it to the four duels for those images so 4.47 and on the next slide so what we see basically is the quantitative assessment so we um did the assessments on twenty percent actually required data sets and we uh split it actually we the brain of the patients uh through alignments to the hammersmith attacks which consists of 83 brain regions and then if you look to the blood healthman's plot so you see that the the the bias and the um and the noise also in the in the quantitative estimates is much much smaller uh for the ai predictions in the in the synonym space in the next slide so we basically so uh this is continuation of the previous work so uh we performed the assessments only for brain imaging and then we did the we extended actually the application to whole body in the partici practice because a lot of indications come from oncology and clinical oncology so we need to perform the whole body apache scans on the next slide so we see some of the administration for one patient so you see the cracity scan on the left side the full dose pet scans with the patients injected with around 250 a megabacquerel and acquired for three minutes per bed position and then you see basically the pre the low dose scans where we took only a fraction of the full dose cancer corresponding to 1 8 actually the uh of the acquisition time corresponding to the full dose scan and then you see on the right side so the two implementations so using two resist two neural networks to residual network which is commonly i mean using medical imaging applications and then a cycle again cycle again so uh is also wide used in multimodality imaging and is known to perform uh quite well in medical image analysis applications so we basically see that the cycle again so it does a good job as you can see here and then we performed also qualitative assessment so similar to the brain study in the next slide you see the scores assigned by the nuclear medicine physicians and we split it actually to the whole body uh pet scan to two regions so brain and the hand and the neck and then the um basically the torso on the abdomen and you see that the scores assigned by the read by nuclear medicine physicians um enabled us actually to conclude that the cycle again does a good job and then the predictions performed by the newest neural network so um approximates quite well before those scales in the next slide so you see basically the quality quantitative assessment so we we did not limit actually the assessment to the scoring assigned bionicle medicine physicians but we asked them also to depict malignant lesions so basically cancerous lesions on uh the four scans that we're producing and then we counted the number of lesions on the uh on the full dose can the low dose scan and then the predicted uh the predicted scans as well and then what's important here is that uh the residual network actually misses a significant number of lesions so actually eight plus fifteenths of 23 religions were misled by the residual network on the 80 percent 80 pensions actually included in the study cyclogan only missed four lesions which which is which is good actually so compared to the logo scan so we message a high number of lesions and then on the next slide so we see also the a a very good example here so we see basically two malignant lesions here indicated by the red arrows uh on the full dose scan but they are missed on the uh residual network on the full on the loadout scan but nicely depicted on the pet scan or predicted by the cycle again and similar observations also were made on the uh on the plots that you see here so compared to the brain study next slide please all right so now we'll say we have a couple of other examples uh for the um basically to produce quantitative images that result that you use for either chinese diagnosis for therapy planning or for monetary response to treatment so we need to correct the best images for a physical degrading factors that we call photon attenuation and there are different ways of doing this and the special and pestamorized system so the mri signal does is not proportional to the um linear attenuation coefficients and then it's very difficult to get the attenuation maps i mean patient specific attenuation maps so what we did in this study here on best mri system next slides is that we um next slide please so we use it actually so um a neural network called a dl ad ss here which consists basically of two cores and it was traded by pairs of uh mri and the x-ray series scans as you can see here so the two cores of the first core basically is called is called syngan and contains 22 layers whereas the second core that you see in the bottom here is responsible for the segmentation it's called segment and it consists of 16 layers so in in total so our deep learning based uh network involves over 54 million trainable parameters and the idea basically is to be able to generate synthetic ct scares from the mri scans of the patients and on the next slide you see next slide please you see basically one example so this is the mri of the patients acquired on the patamurai system this is a patient-specific x-ray city of the same patients uh acquired on a city scan and this is the deep learning solution and this is obviously so the crude approximations were performed by the smri system supplied by the scanner manufacturer and on the right side if you click on the mouse please you see the vise maps which are very high in in in this case and you see that the deep learning or the ai basic solution does a good job next slide please and then they so obviously so what the ai visa solutions it works quite well in 99 percent of the cases but you have some up files where the deep learning network fails so it doesn't work and we don't know why because of blocks black box nature of air solutions so in this case you see that we have a high overestimation of the soft tissue linear alternating calculations and obviously so this impacts the past images generated and this has to be taken into account for the clinical implementation of those networks next slide please so let me just uh yeah please uh last yeah go ahead you please just extension of this next slide please so this is a robotics application so we use also deep learning network that was trained in three different ways so slice wise and then we use it a number of slides at 32 slices with a 3d implementation and then a patch basic implementation each patch consisting of 64 by 64 by 64 voxels and then you see on the next slide so basically the outcome next slide please so what we see here is the x-ray city image of the patient that was not used basically for the training of the network this is the gold standard what we call the city-based attenuation correction we'll use the reference city to correct the pet images for this phenomenon called attenuation this is the incorrect image that was used to train the network so what was basically provided to the network is this the incorrect images and this corrected one for a number of patients and then we predicted those two three images that you see here for a number of patients just for the sake of time you see the bias maps here and obviously the implementation in 2d did a good job so next slide please you see uh this is a good example here uh basically showing the potential of here raises solutions in the sense that if you want to use the ct scan which is not well matched to the best image so we see some distortions here some artifacts here related to these respiratory motion and the deep learning network is capable of compensation for this uh respiratory motion as you can see in these images here in the in the bottom next slide please so um uh the last example i have i think i am out of my time so very quickly so this is the dos calculation example next slide where a lot of work has been the performance on this area so the most widely used technique here is called the milk formalism or we basically calculate the average those of the different organs next slide please and then a number of recent developments so use the voxel-wise implementation to calculate the voxel-wise radiation those distribution and the next slide please you see here so basically the principle so we get the best image and then we convolve with this with what we call those deposition kernels to get the voxel distribution of the irradiation those actually derived through the different organs so this is the conventional way of doing this and in the next slide so you see the ai based solution that we developed so we choose basically a patient-specific attenuation map that is used actually to calculate those depositions so voxel-wise and then we train a neural network to be able to predict the radiation those delivered to the patients and then the predictions match it quite well so the gold standard which is basically montreal calculations which are widely used in the field so with the next slide i think i came to yeah let's get to the conclusions please i think i i'm out of my time so let's get to the next slide so just as conclusion i think ai is going to have a huge impact on this on the field it's not going to be some fancy developments using academia and we project that within the next five years so the number of clinical implementations will explode and this will be one of the actually uh uh uh clinically relevant applications in the field probably within less than a decade or so so thank you so much and then the last night so we just have the my team and uh who did the developments that i basically presented in this talk and then we have some funding from the european commission and then from the sly swiss national foundation thank you so much and happy to answer any questions um thank you very very much for the society it was very interesting um i would like to open the floor for questions uh through slido so please send in your questions or queries or any anything that would you would like to mention in the meantime i would like to inform everyone that uh in the horizon europe program we have uh just launched the uh a topic uh on uh that is very related to what you have just been uh talking about it's about artificial intelligence solutions for treatment and care and this topic is is about clinical validation of of those applications so uh for everyone who is interested in in this please go to the thunder and tenders portal of horizon europe so have we already received a question for professor zaidi so then i would uh propose that uh we proceed with the presentation of dr cotter and then uh come back with uh with questions so people have time to uh to think of questions uh following this very dense presentation of professor zaidi so dr elmar carter the second speaker of this session is an md graduated medicine in montpellier france but also interestingly uh in computer science in paris trained then in radiology in freiburg in germany he is a chair at the esr's e health and informatics subcommittee but also president of the ub society of imaging informatics and a consultant at the department of radiology in freiburg and an associate professor of radio also at the freiburg university and i would like to invite dr el makoto to to start his presentation thank you thank you very much for the introduction uh may have the first slide please my my talk will be less technical i'm going to talk about ai and developments in radiology i'm a radiologist and not a in nuclear medicine next slide please so um as it has already been said in the by peter drewell there are many many papers he said like i think one paper every 30 seconds uh this is uh statistics from from pubmed just showing the increase on papers on ai in both the medical and the imaging domain and as you can see this is an exponential increase next slide please we have a very large number of ai companies on the market this slide i think it's like two years old there's really a huge number whoever went to the rsna the radiological society of north america exhibition last year there was one hall dedicated to ai in radiology with about 200 companies only in ai for radiology next slide please um and if you have a look at the website of of my friend abraham van gineken and this group ai for radiology it's called you can check what kind of applications or products for ai in radiology are on the market and sought them by applications so he lists right now a total of 144 products and 63 companies on his websites but which is quite an interesting number next slide please um we have to be aware that [Music] when we are talking about a ai in radiology or nuclear medicine most of us think of the diagnostic process this is what what is on the lower part of this slide so it's really about detecting or classifying or making the diagnosis using images but we have to be aware that there are many many applications within a within radiology many applications of ai outside the purely diagnostic process i will come back to this later next slide please what are the challenges we face today in radiology as most of you probably know there is a true gap between the number of radiologists and the data we are producing the data volume we are producing in radiology is increasing in an exponential manner while the number of radiologists stays approximately the same or even in some countries is lowering so we really have a gap there as a second point is that the data we are producing are becoming more and more complex today we use many multi-dimensional data like well pet city or pet mr is already one example but when you think of prostate imaging or when you think of a radiomic features which humans are simply not able to perceive so this is an additional challenge the radiologist not being able to process all the information that we are able to put to produce today with our machines then we have many repetitive tasks in radiology when you think of detecting nodules in the lung searching for metastasis and lungs this is something which is very repetitive which is boring for the radiologist and where radiologists even well trained are not very good and we know for many years that machines that computer-aided detection can do better than radiologists especially when combined with a radiologist um then also radiology is becoming increasingly increasingly quantitative it's not anymore only qualitative like saying well we see something you describe it we also have to measure everything and ai algorithms can help us in measuring in a more efficient way and also in radiology we have an increasing number of guidelines we have to respect in our daily lines that they they daily work and this also is something where we really need guidance which could also be something where ai might help us in the future next five days um we have to know that ai is not a new phenomenon it all started around the 50s where the concept of ai came up saying the computer doing having behavior which is indistinguishable from human behavior behavior and at that time specialized expert systems were produced with hand extracted and coded knowledge then starting in the 80s appeared what we call machine learning with some kind of semi-automatically learning of the ai systems and since the 2010s approximately we have deep learning systems that are really working with those deep neural networks which are able to learn nearly in a fully automated way next slide please we distinguish between you cannot really probably read it here but on the left side this is the very specialized ai what we would call narrow ai and on the right side you have the more general ai and the curves you see on the in blue here are showing the a human ai was a human intelligence sorry there is a i and at one point probably this will cross and ai will be superior at least for certain aspects uh to human intelligence when general ai becomes superior to our intelligence that's what we call would call the singularity when ai is superior to human intelligence next type please so we have some challenges when using ai this is in a very simplified way how deep learning works you have an input on the left side of this image which is a the image of a dog and on the right side you have the output which is which is dog and um i'm sorry um and uh in between i'm so sorry excuse me um and in between you have the neural net which you want to train so what you do is by presenting many of those inputs with the appropriate output you train the net to build associations between the input and the desired output this works but next slide please we do not know what is happening in this deep network one is once it is trained so you will be able to train the network you will get the right associations once it is trained but you don't know what is happening inside this black box next slide please um and this is something we call the transparency challenge we don't know what's happening and how can human keep the oversight on what is going on with artificial intelligence and this also has been addressed by the proposal for an ai regulation by the european commission which states that users should be able to interpret the system's output and use it appropriately and also that a high risk ai systems which medical ai systems are should be designed and developed in such a way that natural persons can oversee their functioning next verb please another challenge is it has already been said the validation of ai which is something really crucial to the development and the introduction of ai into our daily routine there are several papers out now saying that ai on the market even when ce market is not very well validated this paper here saying in a conclusion that of 100 ce market ai products that that have been tested only 36 out of those 100 products have had peer-reviewed evidence and that so they are not very well validated next slide please and this is another paper coming from the acr the american college of radiology stating that despite all predictions there is still very inconsistent performance of ai algorithms on the market so both papers and there are many more over the last month say that really validation is a challenge and validation of algorithms today is not very good next time please and this again is a point which is addressed also by the european commission proposal for an ai regulation saying that high data quality is essential for the performance of many ai systems and the training validation and testing data sets should be sufficiently relevant representative and free of errors and complete in view of the intended purpose of the system next slide please so it is very clear the validation of ai algorithms is needed for the development and the validation of ai algorithms in medicine it should be accessible to scientific and commercial developers for ai algorithms the data sets should be available to both scientific and commercial developers we should be able to provide curated and annotated data sets accessible to developers and we need to use broad data sets which are representative of the populations and the scenarios where the ai algorithm is to be used in and of course all this must be gdpr compliant next slide please there is another challenge which is liability who is going to be responsible for the use of ai algorithms the society of automotive engineers has defined that's what you see in the upper right left corner of the site at different automation automation levels for self-driving cars and there this has been transposed to radiology defining several five different levels of autonomy for ai algorithms in radiology going from level one and two which are only assistance to the radiologist and where of course the full responsibility is in the hands of the clinician going to level three where the responsibility is mixed between radiologists and ai and then level four which would be highly automated and level five with full automation where of course the liability would be uh would be on the side of the ai developer next slide please as i said we are concentrating today on the diagnostic parts of a radiology when we talk about ai that's what you see on the middle of this chain processing perception and reasoning but there are many many other steps like patient recording like a scheduling of examinations like acquisition that's what we heard in the talk of professor zaidi also the reporting helping in producing radiology reports and in the communication or the information of patients and i want to give you a few examples next slide please this is a rather old example from my department when where we have been thinking about how we could increase the performance of our patient glucose logistics in the hospital using rfid tracking and automotive autonomous agents today we would use ai systems to optimize patient logistics so this would be one example of application next type please another example would be to better use the slots that you have available at your machines for examinations and this could be helped by ai systems able to predict the no shows in radiology so this has been published by uh by mr harvey in the journal of the american college of radiology next slide please and also by the group of oleg kian york from harvard predicting waiting times for patient would also be something very interesting and could also help us to better use our modalities in the departments next slide please and of course we could use ai systems as a protocoling assistance meaning to produce optimized protocols for each patient for each examination for instance for an mri examination this is not a very simple task and it would be a great thing to have ai algorithms assisting the radiographers in setting up those protocols also ai algorithms could help radiologists with hanging protocols when we get in examinations typically we have a large number of images which we need to arrange on our screens in a certain way and this is something where ai also could help us which would not be a diagnostic process but still helping us in simplifying the way we work next slide please so here are my take-home messages we have many many applications for ai in radiology of course we have many applications for diagnostic purposes in radiology but you also have many applications outside the diagnostic process the introduction of diagnostic ai why we have a real hype today around a.i is still hindered by the insufficient validation of ai algorithms the black box problem and the problem of how to help humans to have still the oversight and also liability issues i think that there are many non-diagnostic applications and also some self-explaining diagnostic applications without these issues and those might be the low hanging fruits for a of ai for the times to come thanks for your attention yes thank you very much uh dr carter also for uh for your presentation um i would uh like to uh to to open the floor for uh questions maybe uh if i could start with uh with one so we of course in the work that we do we are closely following uh particular applications when it comes to radiation therapy so the traditional photon therapy but also proton therapy which is uh more and more uh used for certain cancer indications and i was wondering when it comes to to to protein therapy for example photon therapy what are the the current limitations with ai i mean uh professor society you already i think you showed already some some potential issues but for example is there an issue when you uh have to work with images where there's a lot of air in the in the organ for example in the lung or is that which is typically an issue in protein therapy or is that another issue so my question would be what what are sort of the limitations at the moment in the clinic when it comes to let's say cancer in your experience maybe first professor society and then dr cotter i'm sorry i i didn't get your question i mean the question is link it to frozen therapy right for example because there we clearly see some uh some issues that that field is a bit lacking behind in terms of clinical evidence um and so we what we see in europe is a considerable uh increase in in uh the construction of such sites but uh we don't see that this is going hand in hand with strong clinical evidence for prototherapy for a number of especially common cancer indications so ai talking about ai which definitely is of course going to be a game changer in a lot of aspects when it comes to treating patients either for example particular issues when it comes to organs that are have for example a lot of uh air in them or any technical issues uh in in what i a i can do uh yeah yeah yeah there are some technical issues actually for instance i mean proton therapy is slightly off my field but i mean given the experience by colleagues i mean some of the issues that they face either on mr linux or also on proton therapy devices is how to check and verify actually that the radiation dose calculated by the treatment spinal system is is accurate enough actually so that we can uh do some uh predictions link it to the patient's outcome and so on so this error is basically in the uh radiation those calculations when it comes to treatment planning so obviously come from some of the approximations which are basically ignorant of the algorithms that are used in the clinic and uh i know that we have a proton therapeutic facility here in switzerland so it's based in in in the german speaking part so it's in vitigen in other cultural institutes and some of the issues they have is to be able actually to uh to verify that the radiation dose delivered to the patients correspond to the calculated one and that they use actually best imaging for this purpose and i i think that ai is going to play a big role in in the image reconstruction aspects because there are some limitations linked to the geometry of the pets of the pet's device that company can be implemented on such systems so it's to limit an angular tomography problem technically speaking and the reconstruction of the images is is very challenging and i think aion is probably going to play a big role at least in this area of proton therapy thank you and and and for example in do you see particular limitation issues when you use ai for example in in patients with comorbid disease let's say for example if obesity is it does obesity pose a technical issue when you uh when you you try to use ai in in nuclear medicine applications yeah well actually it causes problems in case you're training data deviates considerably from the actual patients you are actually using in your in your in your data so basically if you train at your network only with slim patients so less than 80 kilograms and then you happen to have in your in your in in your pool of patients or patients weighing 150 kilograms so the the strong deviation between the actual patient size and the cohorts of patients using in the training of the a business solution is going to create problems and probably you're you're going to um yeah to uh to get them out flyer i mean like the example but i mentioned it in in in one of my slides so this is the only problem i can see i mean otherwise if you're uh if your cohort that is using the training of the network is is uh there is a large variability in the data says so i don't just switch any problems i mean handling or categories of patients okay um maybe dr carter would you like to complement on this on this question yes i might add to this that really validation is really the big issue today um because you can have um like we have we have one algorithm in our department and trained to detect the cell radius fractures we have developed it ourselves or we have trained it in our department and it works very well in our setting but this doesn't mean that you can take this algorithm and transfer it to another hospital and it will perform in the same manner so this is really a big challenge and my my personal opinion is that that we need really we need repositories with data with annotated data allowing the manufacturers of algorithms to test their algorithms versus data coming from different places not only from one two or three places because i think yeah that i think is a very important point that you raise now um so um i would like to uh go back to a question on on on slido so um the first question that came in was that for present time where the uh participant says that i may have missed this point but is there any link without those symmetry based quant quantities considered in these ai calculations of doses is that something that you could answer for first society yeah sure i mean no no the answer is no i mean what i mentioned is basically so it's a micro dose image application so it's definitely not none of those imagery i mean the scale of the uh organs that we use basically for uh organ based dose imagery is very large and the even the voxels for voxel-wise implementations for 3d dose distribution calculations are pretty large so it's basically micro dosimetry and not of nano symmetry so for sure yes and then uh the the next question that of course is is with related to what both of you i think have have addressed so the question from the participant is that how do we prevent artificial intelligence making mistakes and i i i think maybe dr carter you could maybe expand on your previous input yes that's well one one point has already been addressed which is the validation of algorithms um his second point is um i think um we should not consider a.i alone or radiologists alone in the end probably we will have a ai augmented radiologist who will perform better than the radiologist alone and this combination will be more efficient than that what we try to do today and the third point is and that's something ai in radiology is missing today um we are concentrating only on the images and all ai algorithms are only working on the image domain but when you look at how radiologists work like 30 to 50 percent of the input we need is clinical input to make diagnostic decisions so this is something which is today still missing in ai for radiology to use all of the clinical input to be able to make a diagnosis in the end that's i think that's that's clear so there uh is another question in in slido that is about the uh yes participants stating that in your presentations you focus on diagnostic applications of artificial intelligence but is there something that you could say in addition about the use of ai when it comes to treating cancer patients for example uh you know radiotherapy so cancer treating cancer patients um we know that uh you know treating certain cancer indications would require multiple visits to the hospital and every time the patient will have to be put under the scanner at least in that's in the the most optimal scenarios that i have seen and so what could you say about ai when it comes to treatment and and follow-up of these patients uh maybe professor said yeah for sure yeah i mean there are plenty of literature actually on the topic and uh so number of papers were published actually on the radiation those calculations in external beam therapy in brachytherapy and brachytherapy actually the concept i mean in terms of physics is very close to nuclear medicine so we also have one of my phd students who implemented in a business solution for brachytherapy those calculations and in addition to this i feel that some of the open questions on for instance mr linux so there are at least a couple of companies i mean producing this type of equipment for external beam radiation therapy this is alexa and varian uh no i'm sorry not valiant there's another i think german company for the name and one of the common problems on those devices is that the radiation dose is calculation based on the mr images and then the physicists do not have access to x-ray series images and there is no direct link actually as i mentioned it between the proton densities and the charm relaxation times that we have in the mr images and the electron density which is needed for the radiation dose calculation and there are different ways of converting the mri images as i mentioned if one in one of my slides to convert the mri to x-ray ct images and this is just another application that's actually the i think the major parts of the literature basically focuses on algorithms for grades and those calculations for external beam therapy and also for bracket therapy thank you very much dr carter would you like is there something you would like to complement uh on this question uh use of a cancer treatment as as compared to diagnostic applications is that something that you think would be you could say in this to the audience yes i think main applications would be on on one side the uh helping in uh contouring and optimizing the contouring of the structures to be treated and on the other hand also monitoring the response and using ai to monitor the response over the time then um there was also uh there is also a question on uh from a participant that who is working on ai and uh would like to participate to uh projects under the uh the new framework program called horizon europe that started uh in january this year and the calls for horizon europe that uh they are they were slightly delayed because of uh you know strategic discussions in in a small part of horizon europe program but how can you participate well uh there is now a call that we published uh just a few weeks ago is out on the thunders and tenders portal uh this call is uh is uh let me not state the the number or or but the title of the call is uh clinical validation of artificial intelligence solutions for treatment and care when you type in funders and tender portal horizon europe and you type in just artificial intelligence in the in the keywords the search function i'm sure you're going to find it the topic identity is a title that says horizon health 2021 disease zero four slash zero four and this topic is is is very much uh on uh addressing what has been mentioned by both speakers so if you are interested please apply uh this this call is open for uh so so-called cooperative research so cons a consortium that works together minimum three countries from the three different member states uh in in europe uh the deadline for this call is 21st of september now i think uh with this um i'm looking at the organizers i think we we have addressed all the questions and we have come to the end of this session if i'm not mistaken so i would like to thank everyone thank the speakers from society dr cotter and i would like to pass the floor again to the organizers thank you thank you thank you so much you

Channel: EU Science & Innovation

Views: 298

Rating: 5 out of 5

Keywords: Innovation, European Union, European Commission, Horizon Europe

Id: vfG-PZUAUSk

Length: 58min 38sec (3518 seconds)

Published: Fri Jul 16 2021