welcome everyone I am suin Jo uh the director of medical physics of the Department of radiation oncology uh at the University of labrasa Medical Center uh often common refer to as UMC it's my great pleasure to address such a distinguished Gathering of uh professionals medical physicist radiation therapists medical dmetrics and Physicians fellows residents and students and from around the world and physically on site or uh virtually uh via Zoom we are here to uh share knowledges of the latest development in the field and also learning from each other uh how we solve the real world uh challenges together we can improve radiation therapy for cancer patient today marks the second day of 2024 middle West Middle East Midwest ran colle Symposium this event require months of uh planning and hard work uh I suggest we give the planning committee a big round of applauses to thank them for their time and effort at UMC our medical physics division stands at the Forefront of innovation and Excellency we are fortunate to have uh a diverse group our medical physics residence program has recently doubled in size now training four residents two masters and two phds we are proud of our residence uh including one recently honored uh with the best uh disertation award from MD Anderson uh PhD program in physics who joined us recently our Medics physics division serves two caner centers uh across using variance uh eclipse and ARA combination uh for treatment planning uh recording and verification and Tu beims to deliver radiotherapy treatment the consistence in Ure that our patients receive the same high quality Care regardless of which Clinic they visit our main campus offers Advanced bra therapy services using uh un Central TPS system for planning and flexr after loaders for delivery this Technologies enable us to deliver uh HDR treatments uh for cases where external being radiotherapy may not be uh as effective in addition to standard standard electron and Photon uh radiotherapy our sources include an extensive array of special procedures modalities such as image guided radiotherapy service guided radiotherapy uh stereotatic body radiotherapy stot unique uh to their needs and ensuring the highest standard of care we are actively exploring proton therapy I know Dr probably mentioned it yesterday uh it's a significant advancement in radio radiation Colley delivery and Promises uh something uh radiotherapy treatment with reduced uh side effects we anticipated we will get a great news soon and we believe this new technology will enhance further enhance our ability to provide cutting Age Care to our patient the rule of medical physics at UMC are multifaced encompassing Clinical Services Research education and community service and Outreach we uh collaborate with a wonderful team a multi-discipline team that includes radiation oncologist medical dmes radiation therapists use patient treatment plans to make sure uh the right uh radiation do uh is delivered and we also participate in specialized procedures like protherapy uh stereotatic radios surgery and inop uh uh procedures and at UMC the planning these procedures are planed by exclusively by our medical physics ensuring uh precisions and Care in every aspect of the treatment research is another Cornerstone of our work at UMC where we uh collaborate with uh experts from other uh various disciplines uh our research supported by internal and external founding focuses on areas promising areas in medical physics H you know just give you a few example like the artificial intelligence and deep learning applications in uh radiation oncology uh medical image based uh clinical uh treatment uh patient outcome uh uh uh prediction and Flash radiotherapy this Studies have a real world uh impact and uh will improve patient uh outcomes down the road our fa posters and uh residents have made a significant contribution uh publishing extensively uh in peer reviewed journals and presenting their findings at National and international conferences the work Advanced Medical physics and has a broader impact beyond our institution next education is a critical component of our mission at UMC we Mentor medical physics and red oncology residents and teach uh medical physics to highlight Health uh radiotherapy student and also help training uh medical physics students uh from Katon University our residents achievements including uh National recognition and uh participation in prestigious conferences uh reflects the commitment to education uh three of our last five uh graduate uh residents from the medical physics residen program and went to uh join academic center and all five are thriving in their roles as medical physics so their success speaks volumes about the quality of our residence program and also the dedication of the faulty our medical physicists are also actively involved in service uh to the broad Community uh you know contribution and contributing to organizations such as APM Missouri River vany uh chapter and variance Flash Forward Consortium this activities are vital and uh for the continued growth and development of our field one highlight of our uh service work last Academic Year um was uh hosting uh our first Global Health uh rotation and a multi multi institutional simulated patient communication training for physicist these initiatives prepared our residents for the challenges they will face in their career despite challenges such as uh staff shortages our physics medical physics group has continued to grow and evolve we have welcomed four new uh tented medical physics and are excited about uh numerous clinical projects uh on the horizon and Dr inc's leadership and with the support from the University of labrasa medical center and uh labrasa medicine we are Optimist the field of medical physics is advancing rapidly and driving by breakthroughs in technology research and new clinical application the pace of change is exciting and challenging a requiring Contin commitment section six and 10 will explore how AI is Revolution ize treatment prision accessibility adaptive and personalized radiotherapy and P therapy next we will discuss the potential of flash RT to spare normal tissue in section uh session seven finally session eight will highlight innovation in medical physics education and and including Global uh collaborations and active patient uh effective patient uh communication strategies we are delighted that two of our former colleagues are among the distinguished speakers in today's presentation we are eagerly for a day of uh incitive press uh presentations and discussions next let's welcome uh our Kino speaker Dr Shaun Yang Shaun Yang PhD daab is a poor do professor in cancer research and serves as Vice chair uh for medical physics research in the department of rad oncology uh at uh em University University School of of medicine where he received he completed his post and medical physics residence training Dr Yang is a aent faculty member in the biomedical informatics department at emry University and the Wallace uh C Department of biomedical engineering at Emir University and Georgia Tech as a board certified uh medical physicist Dr Yang specialized in medical uh Imaging analysis uh image guided radiotherapy artificial intelligency and M medical image analysis he is the leader of the deep biomedical Imaging lab where he and his team uh focuses on uh improving the Precision of radiotherapy and precision of radiation therapy Dr Yang has published over 240 uh Journal appear reviewed Journal articles papers and uh received over uh 15,000 uh Google citations and has received numerous uh several uh ni1 grants as principal investigator and numerous scientific Awards including the John llin Young scientist award from The aapm Today Dr Yang will share his Insight on leveraging AI in Precision radiotherapy advancement in medical imaging analysis we I don't have any interesting so um you can see this figure actually they prly show the basic workflow for our C radi therapy so if you can see uh very carefully uh detailly a has be uh applied for each step of the cancer radio therapy a can use for deing support before treatment a can reduce radiation expor um and it has image quality can use for the tumor organ country and they can expertise the QA process and Generator plan and enhance the iMed gardeny and motion management finally we still can use a to Accurate predict response after treatment so specifically uh AI can be uh used to improve cancer radiotherapy by uh preventing more precise cancer detection diagnosis and more personalized and precision treatment strateg and less image radiation expor and image artifact better image quality and a more accurate image res and Target cations B and faster preing planning and curity process you has guiding motion management more conven frquently accurate patient followup by you know can provide better earlier and more accurate outcome prediction so if you take a look at this workflow so medical image plays subtension rules each uh phase of entire cancer radio therapy use AI technology Especial deep learning to advance the quantitative image in the cancer therapy could uh increase Precision of cancer radio therapy so in this presentation AB share several deing based image project which include image snation image deform image res uh medical IM synthesis as well as there are potential clinical application s so uh as you know so for our um cancer radiotherapy contrary Target and uh critical organ is required to quantify uh the dose receiv received by Target at normal tissues however manual control is time consuming and labor intensive so in the past maybe five 10 years deep learning has been rly used for the develop for the organ syation and many publication damage their are great performance even several commercial deep learning based software available for the clinic practice recently deeping has been applied for the tumor sub region or organ substructure segmentations so using deep learning method with limited annotation or without La annotation to explore the accurate tumor sub region or critical substr syation will be one of the very interesting uh res ination uh research Direction so uh in the fin F SL I will quickly share some uh medical mation which is different with the current commercial software so here is uh daily cbct country uh example as you know the daily cbct play important roles in providing latest 3D information of patient uh position at Anatomy during cancer radiotherapy accurate cbct snation is much needed for our dose modification or Anatomy change modification especially for CBC CED radiother therapy however uh the soft tissue contrast in the cbct is very poor you can see from this daily cbct image so this uh this patient original Mr or this synthetic Mr so our idea is we use cbct to generate sythentic Mr which like this image so which has better image uh soft tissue contrast which can help for the uh syation so you if you take a look at the profile from uh blader prosty and Rector especially take a look at the pr region so there is a low boundary information in the cbct image however we can see some kind of boundary information in either original Mr or is stima which this information actually is very important for more accurate processing patients so and rectors this is a second example for the cand substra syation so in current Clinic practice you already especially for soric radiotherapy the whole heart you be sigment to the whole heart has only especially for left at descending Lottery uh entery so we develop AI framework to perform uh uh up to 15 cardiac substructure syn from planning CS so the basic ideal we use multiscale feature to localize the V region of interesting for each organ there subst structor then perform sign accurate signation with the uh HRS so here you can see some uh 3D compar uh spare T planning if you take a look at this three plan comparison for the tradition plan this tradition plan this heart spell uh do planning that's uh substructure cardiac substructure SP planning we can see we can significant reduce dose for heart or use some critical substructures which could U we think can improve the cardiovascular toxicity so uh this tool actually also very useful for the Post treatment cardiovascular toxicity evaluation uh maybe uh also very useful for uh vric picardia rap so here's the last examples I want to show here is as you know more and more publication demonry is the advantage of Advan P image in the cancer radiotherapy but we Face AI framework to incorporate personology based information into P image to generate more accurate Works wise tumor prob probability map we think that probab map can help our uh do prostate treatment with do escalation to the dominal leion so here is some two example so you can see this are two case one from the patar is PMC P another one is pamar image from facbc twers so uh you can see this Fusion image between M and the P if we only use e Mr or eer pet to generate tumor map there has to here uh those result we combine Peta with person data use a model to generate some uh tumor probability map the result actually very good we show our physician uh currently this ouro study again so we just try to include more patient and more data to test this concept consuming and label intensity especially for Med physis so which prevent real time chat planning and do symmetry during the alasa or Mr guard needle placement procedures so of course AI method can help so we can see this manual uh country from our media phasis those are the result from several deep learning based method either you can see the uh uh the green circle triangle show the uh either wrong detection either more detection or less detections but uh we can see actually overall the Deep learning method especially the proper meod get a very decent result so this two figure show either 2D Arrow distribution uh and also 3D uh needle deration you can see the the red part is the tips because in breaket therapy tips location is very important overall if you take a look at the the table we can get a 7 millimeter um tips arrow and also over have we can have like over 95% accuracies so the this multi needle detection method holds um significant clinical value in stream HDR treatment workflow with improve treatment plan quality and it sets a stage of real time adaptive HDR Pro C break therapy in the operating room so last year uh around in the AIL The Meta a release a uh segmentation model they call sigment elic s which is a deplo model uh is where the TR use Bing samples so uh several group include multiple group include our group perform some evalation for this a tools to dis imitation so here is the example is for we test use our clinical data you can see actually Sam perform uh can detect you know n part and reap isus very well um this 2D segmentation actually very fast but you have to do some like a PR process for your specific segmentations so uh then uh just like early this month the the the group also release another same two uh CHS these two can quickly uh real time segment even 2D videos so here's another example uh we develop for the 3D segmentation so in this uh application we develop one for all signation use self super self self-supervised large scale Network which one four Bing parameters so once we TR this model this model can perform snation for abdominal pelvic even chest CT and also this model can perform snation for the M even for the P image so that's uh use uh if we can treate a UniFi model this model can perform signation in different site of for different modality that be something will be very exciting for our Clinic future application in current AR in one model never be much convenient uh deformation has be uh used wonderly used in the each phase of the cancer radiotherapy so in next few slides I like quickly show some how a can help for the deform image positions so this figure show the the the Deep learning based uh uh imtion uh articles by the year we can see the publication number significant increase and the Deep learning has been successful used to enhance the performance of iMed res so this figure actually shows the over view of seven uh category uh in deep learning uh Bas method in uh iMed res I'm not going to uh go to the detail but I would like quickly share uh some uh funding based on our previous research so firstly deep learning based onestep transformation prediction technology could accelerate the traditional immedate recision make many of previous deform recision can be done in real time so this which may stimulate deform restriction to apply for mainly real time image guarded intervation uh procedure our physician request 4D C image to monitor to investigate the patient breast informations so after uh physician uh go through the moving pick up which pH we will use for treatment uh uh a 4D image stion could offer a quantitative motion analysis for precision radiotherapy here is example we use you can see those Fusion result between two phas before res those all the uh Fusion uh result after res with different loss function if you take a look at the final uh result actually the small Vel match very well which demonst except this uh good performance for AI based deform Institution significant much faster or faster than the conventional method which can facilitate a comprehensive motion management for the patient even on the CT table immediate after for this gu um such speed and accuracy can greatly assess our radiation colist in make a precise objective decision regarding treatment modality uh treatment phase selection or even this can also support for the accurate fast restriction between daily cbct or weekly uh between daily cbct and between qct and planning city is much needed to quantify the interraction anatomy change like tumor Shing or maybe or variation and also perform Contour propagation and this kind of retion also can support support for the aaty change prediction some kind of studies so we same concept we apply deep learning for this kind of restriction you can see this one example for pacreatic cancer radiotherapy this you can see the cbct image for each days and we use our Clinic uh retion tool and air based method so here's some result for the uh Fusion you know after res the fusion image after resion this another example for the head protal therapy we try to redress the qct with planning c those fion result we can see actually after re we can get a decent result so this fast and accurate deep learning based Administration actually is very useful for the for performing longital quantita analysis during the cancer radiotherapy especially I'm sure a lot of group try to work out the patient Al Anatomy change prediction study for those kind of study you have to enoll a bigger number of Pati but you have to find some very fast accurate image restation tools to support this kind of project uh except uh you know uh previous I mention we use cetm to help for the image snation so here I still want to show some example we use sity image to add our Multi modity Image rests so uh the basic oh is a s Mr to Mr res so here is the example again this this patient original Mr so this stima we generate use CT we can see actually they have very good soft tissue contrast which can could help for accurate mrct resion so this some result you can you can take a look at this Fusion result we directly use mrct resion those Fusion R we use syth sythetic Mr added resion if you take a look at Hana region so you can see actually the stic image added res could overcome this mismatch in some regions and again this another example for the head leg case for protal therapy so uh if you take a look at H region so the direct we we can see a large mismatch for the directly mrct recision however if we use sens CT to add mrct recision you know you can see actually bone and soft tissue match very well um break therapy um AR breaking therapy could can be significant benefit if we can incorporate multi parametric Define dominant leion into real time alasa to guide the needle placement so however for this procedure we have to uh develop a real time Mr alasa rests so you you can see here this PR Mr image T2 image this Pro ARA image there toal different gr level intensities and also the uh image feel of ultraa is very small if you take a look at Fusion image this much smaller compar Cloud correspondence now we incorporate the proy material property with this surface Port surface part correspondence uh into F element model then we can generate volumetric Port Cloud U uh correspondence Lo Port Cloud volumetric Port cloudon could use for our letter worker training so after training our letter worker is able to predict the port Cloud deformation between M and arasa with build in bi mechanical constraint so in summary is this method try to use a deep learning model to learn the fine element simulated War metric Point Cloud correspondence to expertise deformable Mr tion with biomechanical constraints so here's some result you can see so you can see those AR Imager are the different direction those M imer are the different direction if you take a look at the system so we compare several method with our method so actually this match very well in this uh propos a based method so this table show the overall the Restriction Target arrow is around 1.5 mm which demonstrate the air based meth could uh accurate redress the Mr to alasa to support uh IM guarded prostate interation this not only for the breaket therapy but also can use for the prostate biopsis so uh in the previous several slides I I show how we can use air based deformation to support image guarded radiotherapy here I just want quickly show another example how can we use deep learning to predict HDR needle positions so because as I mentioned several time the learning AI based uh uh deformation uh is much faster which could enable or rampid multi analist restriction make this study clinically practicable so this figure show uh the perform generator new need new needle position for the new patient of the new patient through the at restriction deep learning add an this restrictions so here is some result you can see uh here is the 2D 3D comparison as DH comparison so this actually is a good case you can see the DH match very well there Aver case you know the the the pro cover is okay but uh you know for the predictor uh location that the case will receive a little bit high dose to Rector and blad but still Clinic acceptable but this worse case you can see so the deploying method means some kind of uh uh me Miss uh they miss some needle in the left of the prostate so after you normalize uh the dose for the prostate we can see the rect blood receive significant high dose so maybe we need a better uh needle regulation algorithm to uh improve the final predictions so this two actually has a greater Clinic potential since it can provide a cancer location estimation before um procedures so which could reduce dependence on the physician experience in CET placement and enhance quality of the prostate HDR treatment plannings so lastly I want share some uh example for the medical maor cences so with the development of the AI technology medical IM synthesis has been wonderly used in Radiology at radiation IES so we uh classify the medical major syntheses in two category uh first one is into modality synthesis uh which include data argumentation and include from low to high quality uh quality image synthesis like a delos study like a metal artifact correction or scan Corrections uh another one is low to high resolution Improvement in the image domain and some group still use to low to high frequency improve in the frequency domain this especially for the key space for Mr imag and also we can perform cross modality transfer like use T gener T2 and also we can use low field Mr to generate High field M image uh of course include image reconstruction uh for the in modality since is obviously we can currently most of people know we can use Mr CT we can use cbct GM which we already talk about help our segmentation or image resion uh some group still use M pet use PET CT and also another important part for our radio therapy we can directly use either Mr and CT to generate do plan so here's a first example for the data argumentations so if you work on deep learning model training uh data commentation is required the traditional augmentation include TR translation flipping rotation and scanning deformation this Mees cot increase the diversity of the data that's the issues so uh deep learning uh especially current def model uh C regenerator use different kind of uh deploying method so that's uh result use our propos method use diffusion models uh this another example for the Mr we treat our test our model is a several larger database after we generate those uh image we give our faces our phys the you col separate which one is real which one is sythentic so that means the Genera image actually looks very realistic so we can use those C imager to treat our model especially current the larger Foundation model is very popular but the problem we need a huge database to do majority time most of time we clarify the very bigger database for this kind of training but we think uh this deformation model based like like data uh synthesis could help this uh augmentations and also you can see this figure show the recently because Mr gued the redo therapy Mr based sensity is more and more popular in our field however how we how we can use Mr to generate sensity is lot new this usmr to sensity was developed in the uh first petar scanner in the 2010 that time we use segmentation based or either traditional machine learning based method to generate sity actually we call sudo City that time to provide the tissue aention information for pet altion Corrections same concept so if we can use Mr as only image modality for T planning which we can El limited synthetic elimi system itic Mr C cotion error can reduce medical cost because we don't need a CT can spare CT from CT spare patient from ctx3 exporo and also if we only use M we can simplify the clinic workflow but the problem is as everybody know Mr cannot provide electron desity electron density information which is very important for our accurate do calculations so use Clinic routinely Clinic use use Mr to generate sensity actually is uh is very designable to uh for the Mr guarded radiotherapies so our group actually have work on this topic for many years we published tons of Pap about how can we use Mr G reasonable s c uh so here I should two example you can see this for pelvic side those all the uh planning CT image Ser as reference those synthetic uh uh CT we generate from Mr different Mr uh those are different image between uh CT and S uh if you take a look at the zoom in image or E the uh difference and profile actually the generated sity is very close to the planning City okay another application for head region you can see this original TM those planning City those S City reg generate those are show the difference and even from image we can see they are very close so uh that's a quick example we already use our mbased sity for am prot Center in 2018 so this first case we we D this patient Mr only have Mr SC for the patient this a sity reg generated uh and also our do to make a proton and a photon plant is those comparison so this kind of two mrb soy particular valuable for Pediatric patient because it help reduce their exper to additional radiations so as I mentioned before so uh Mr basy can was early developed in the PMR when PMR was developed and here is example we show how we can use uh uh sity for the pet attention correction um as you know for PET CT scan us already CT S as uh CT can be used to generate attention map for pet attention correction but however in current combine petm ma m can D cannot directly provide this information that's uh we can use Mr business in the different uh uh volume of interesting in the brain region we can see actually the percent difference less 5% which demonstr the S City we generated is good enough for the pet attenuation Corrections so that another example I want to show is uh for the cbct based s c which we also work for many years e GRS and those sity we we gener to use a model if you take a look at you can see the uh the sensity with very good soft tissue contrast and reduce a lot of scanter especially they can actually perform very well they can significantly reduce metal artifact which actually is uh is a problem for our protal therapies this figure just shows the his gr compar we can see actually the S City the hisr of s City and planning City are very close so that's again another example for the abdominal regions so this planning City with very good stive contrast here is our daily cbct for pacreatic treatment we can see that's a severe artifact because air po Pock pocket so there a sensity we generated from the daily cbct we can see actually we significant improve the IM quity at least our physician is confident to use this sensity for their country and also for our correction we also correction the H numbers that means our sensity can directly use for the dose calculation once you have control uh we continue work on this project because we want to improve uh the sensity generation for different s here is a very interesting example I want to show here this is for long spt treatment the the circle region is tumor this different five days F fractions cbct uh this is the again deform plan City we seriz GRS uh those actually sensity we use previous General model if you take look at some region especially Hana region you you can see the Deep learning model get rid of the tumor because the model think this just artifact that's ridiculous we cannot use this kind image for our uh treatment so Bas our uh multif fraction treatment cenal so we propose if we can use previous cbct to fune our general model make this model have some patient specific information they may improve performance so we that's the some result we use our idea you can see so here is some like uh all the sensity with uh different uh F because we have multiple fraction we use uh for this result for this result actually we use this result we use first uh first fraction cbct okay for this re we use all the previous cbct for model Trea you can see if after we fune this model we call Patient specific model they always can keep the tumor and also significant improve the image quality that something will be very useful for Sam renal uh this here is another example so for current uh cbct based uh sens most time we have because most time we use a supervised deep learning model for supervis deep learning model you have to build a match CBC C and planning City use this database to Tri model but in some cenal like abdominal it's very hard to perform a good restriction between daily cbct and planning CT you can see because air poket issue the artify issue so you can see those all the daily uh cbct those the deformed the planning city after deform res this still you know the the if you take look at the high Anatomy the air paret canot we cannot perform perfect deformation so that's one of the motivation for us can we develop a unsupervised method for this kind of application so we are work on so here is some result paper under review you can see this those are the sensity reg generat use unup method if you take a look at this s they significant reduce artifact scatter but they keep same Anatomy as cbct which was you know image which could help acquire Real Time 3D volumetric image for the tumor tracking even currently we use MRX but you you may know that they only capture few slides real time they can all to capture real time volumetric imager for the motion tracking so of course deep learning AMS can use for this kind of application so here's the example we propos uh you know specifically we use under sample data from caspace domain to reconstruction low quality Mr then we use deploying based model to enhance the quality of on the sample Mr make it comparable to four fully sample Mr so here are some result you can see this ground TOS we use full fully sampled Mr those different result with different method and also with different accelerator uh rate if you take a look at here when the accelerator uh rate is eight uh that's our result you can see actually this result still looks uh good is comparable with ground to image so that means at least those this kind of a me can accelerate uh Mr position time up to eight times so that's previous I talk uh I discussed about U acceler Mr image in key space but we still can uh generate high resolution and acceler Mr uh Mr or CT uh image uh acquisition time is the immedate way so uh as you know in current our image protocol we usually capture high resolution image like 1 mm within slice however with syn SL cist either because radiation dose because the patient motion issues so if we use if we have a 3D war with 3 to six mm SL cist that image if we use that image for the contary for Downstream analysis they may affect our cont accuracy and do calculation so to deal with this issue we propose a self super self-supervised model to uh learn the high resolution information with the slice and then perform uh improve the course slice resolution you can you can you can think that's like a interpolation in the the directions so this model actually they don't need any gr chose high resolution image for the model trainings so uh here's some examp example for the CT you can see this CT ma a capture use our current clinical protocol with 3 mm slist to test our method we done sample the sist to six here is the high resolution imager we use B cubic uh interpration which is one use for our clinical practice here is a high resolution image use a method uh if you take a look at the the zoom in image you can you you can find a stair sever stair artifact in low resolution image uh in the high resolution image with inter interpretation the image become very BL but if you take look at the high resolution image in the you know use am method they keep a lot of details actually is is is very good those IM will be very good for at least for contrary those calculation so here is another example we apply same concept for Mr image so you can see that's the whole body Mr uh with current clinical protocol they have 0.9 0.9 and 6.54 millimeter uh uh resolution so we can see s St artifact in the 3D imia so again we we generate high resolution use very blur so however we can see a lot of detail in the high resolution imager we we generated use air method this high resolution uh detail could be used for Al romics or maybe contrary U process uh another quick example is you know study have show multi contrast mrr is very useful for tumor defination defination and the response evaluation in radiotherapy however acquire multi contrast mrr use multi sequence is time consuming and also very passive so even sometime we have to use contast agent this make the clinical flow more complicated and also increase medical cost so of course we can use AI TR to synes is multi contrast Mr this may be a convenient and efficient way for the fast multi parametric image especially for Mr gued adaptive radio therapy so here is some example you can see we try to use the a model uh to synthesis the multi parametric Mr met from the cross modality synthesis here is example the input is T2 and you can see this t t image the synthesis T image the synthetic t with contrast image and the flare image if you take a look at the tumor region actually they match very well that's another example I think yesterday some speaker talk about use structure image gener function image here is concept we try to use T1 and T2 flare to generate the ADC map so you can see this ADC map we generate those original ADC map actually they are very close so another example here is um for enhance low field Mr which is will be very useful for some Mr Linux machine so as you know the high magnetic field M could offer better image contrast at a special resolution but the problem with high field M currently suffer from limited Clinic availability at higher cost uh and also they have some technical issue especially for Mr Linux so again we try to use air model to S is high quality High field Mr from low field Mr so here is some example you can see so uh you can see this image is from the 64 M Tesla image those 3 as a reference we cannot see you know good soft TI contrast and a very ly blur but those are the synthetic image we can see so if you take a look at synthetic image that's a great soft contast and Sh very well so another example we try to use uh a model to help syntheses uh T seven seven for the 3D DC map you can see that's some structure is very BL because uh uh the special resolution issues so uh last two topic I quick want to share is how can we use air to reconstruction uh or real time monatic image uh for for iMed therapy as you show you see from this kind of moving so the taret motion during the radio therapy uh radiation beam delivery could smell the dose distribution and reduce those conformities however that's give us CH motivation is Real Time volumetric image is much needed if we want to J to real time we have to use one projections so this figure you can see the uh shows the comparison about the different reconstruction method we can see uh for the traditional method we require uh many number of the projections a deep learning model can significantly reduce the number of the projection as need um especially for the uh currently patient specific models so the patient specific deplo model can even reduce the projection up down to one projection but however we can see this still need a a high level of the priority so if we can use the one projection to reconize 3D image which make our 3D real time dose tracking feasible during the dose delivery so here here's example we use a patient specific model to deal with these issues the basic idea because for each L spt patient we have 4D scan we use 4D scan and there projection to tr hour model at the same time we use micol simulation to add some scatter artifact to adapt our model to match the cbct uh uh imager in the LX here are some result you can see this planning City so ground tools we generate the 3D monometric imager use web projection from this projection so you can see actually with a different the final result if you take a look at the tumor region and the contrast actually is comparable to the planning City or region reference or you may argue said oh there still quality need to improve but think about this image we use over 360 projection but this IM this image we only use one projections so we continue work uh uh improve this method recently we introduced diffusion model to this project we can see the current 3D image we generate use one projection actually is much close to the planning City the difference very small so uh another example a quick water shell is just for because the previous study we have to use 2D X3 is a lot of radiation maybe uh for some patient with canuse but we still want to investigate if a can help us generate real time radiation free volumetric image that's one ex example here we just use surface image to g generate volumetric imit um we use a same concept like a patient specific model from use 4D CT scan with the surface as you know three armor models so here are some simulation result you can see this real CT image those prediction 3D CT image those similarity map and a difference and you can actually B on the histogram they are very close it's good enough for the uh motion tracking use this synthetic 3D image so lastly I have to talk about you know for the planning because planning is a very important part for our radio therapy so definitely a can help uh delate uh Tre planning so at the beginning uh many group have already use a deep learning based model to uh generate predict DH or either those distribution help you know those kind of PL cannot directly deliver by our machine but at least they can provide some knowledge to help our do symmetry generator quickly generate good plan then several group investigate how can we use aell to deliverable radiotherapy PL uh that's also published by several group this generation uh we have to uh use human perform plan evaluation at the Symmetry adjustment the optimization parameters but in this study they try to use a foundation model this model can automatically perform uh plan evaluation and automatic perform uh optimization parameter adjustment so this can because the plan use eded priority in radiation color learn from do symmetries and our physician to function as a expert planner guided uh finally guided uh treat planning process this method actually for the with Foundation model put a significant make our online adaptive feasible in future so in summary uh AA could improve L identify and address B to ensure the failes in a particularly in house scale is also very important not such I will quickly show very interes study so this paper published by Google deite it just get accepted by uh very good cence so you can see this table show how much parameter in this people they used with different experiment you can see uh this are different um experiment but somewh of the experiment these people use almost like 26 billion perit in their model so that's the reason after the people publish on I some group investigate how much cost comp computation cost uh used in these papers so you can see that's the follow 10.9 million just for one Pap for different kind of experiment this can make us reink about the balance between larg model even lar model very popular between larg model and computation cost finally I would like to thank all my collaborate my funding support uh and also my lb members without their uh hard work uh a lot of project canot going very well thank you so much for your attentions wonderful I know there are some questions from online audience and also we started with this okay the first one uh regarding pro project where the SR is generated from CT can you please comment on the accuracy of atmr do you uh for is a technology of being explained into a generator sity image for more M modality Beyond regular T1 T2 image I think that's good question you know so the S image you know I'm sure a lot of research concern about sity image so that's a problem but you know is it depend uh what kind of task if you just use S image for like uh diagnosis high risk task I think you should be very very careful but sometime for our radio therapy for our department we just use Mr or CT either for dose verification either for cont you know because all the cont will be reveal and approv by our physicians even I'm sure s image May introduce some error but this should be okay like if we just use for Country some kind of low risk uh application I don't see ex EXA issue and also you know they can some kind of see uh can significant improve like efficiency especially you know like I show the example the cbct directly if we want to perform direct CBC most time you can canot get a very decent result it's terrible but if you use CBC Bas santr to end the add that even we know since Mr may have some error but most time they get a very decent result that's in future CL application we need a separate which task another question about project where used to Fusion Mr pet for break of the PTV these are specific want to localize exactly location mag of uh I think I think I talk about Mr Ultra not Mr pet recision part or I little confused for the fusion Mr pet or maybe that's question about the uh tumor probility math we try to Fusion we try to incorporate uh bsy person data information with PMR to dat more accurate tumor prity map uh you know in current prostate treatment we we treat a whole PR that's true but majority time we still give additional do to dominal leion so if we can uh generate a more accurate dominant Le um I know that's a lot of publication question about even multi parametric CER they could not g to very accurate dominant or P still have some issues that's reason one of the application we try to combine you know P me m even with personology data can to generate a more accurate Works wise Dominion that could maybe benefit our Pro break therapy what's m CH accurate Rob yeah that's true I mean um um the the a lot of time you know stion is the reason you know we we even in adary clinic we use uh airas software sometime we know they cannot get very good segmentation accuracy but you know well see I want to Han said you know a Bas method definitely the current state they can significant improve the efficiency for some kind of like signation work but the the major challenge for the signation accuracy because it's depend what kind data you use you know we we use physician control but you know we did some uh study before the inter variation between physician also were huge but process sometimes 30% we we did some study before so that's reason if you use the image with Contour with different physician to treat the model actually you know you give some a lot of information uh I mean I know in some institution they have some specific position created by Department this kind of position can support how can you translate the lab developable uh in house to for Clinic that's you know that's a lot of gap for me uh you know we are running lab I I really don't have enough budget time to let all these tools works very well for Clinic a one of the good news we just get why H uh early clinical trial funding let support I work with well our physician for the cbct gued spt treatment adaptive treatment so that's good because we get enough money some money to support some kind of poster uh students to make this way really useful Clinic I think for for me um two solution one if Department can create some kind of position can help us definitely we are very happy on other part if we can I try to work very close with our relation colies if we can get some support clinical trial funding from H that be also another good way for us to you know test what's really outcome I know that's very important a good questions uh what CH existing ACC from Mr particularly in term bone and self tissue interface um since City should all use under what conditions are definitely better than normal uh similar as first you know Society all always there has their limitation like I mentioned before especially if we want to use the supervise based method you have to build match you know e mrct but you know most time our deform resion cannot perform very well so that's a issue uh even sometime we have a database to tr our model but that database still include some uh hair we know that's not perfect but you have to use because you need a lot of patient data to train I think again my uh suggestion is it depend what kind of your application if you just use sity for those application I don't think that's issue you know a little bit uh you know variation will a lot to affect our those uh distributions but if you just use that kind of image for followup for response evaluation you should be very careful we don't know maybe it's better for for you combine you know even you have a sanity image but you still need some original image combine with this s image maybe can provide a more cisive evalation have you try generate quantity2 Mr from CT such T1 T2 IDC map every comments a little bit on possibility to generate accurate quantify uh quity Mr Data from City no we don't do it um the only application for us we use CT generator Mr either for segmentation either for restriction because if you your city try to generate quantity TMR they need a you know especially if you want to use this quantity TMR for some highrisk task we don't know that's good I think is a good point but we we we need a test use more details yeah oh second one oh what Q process less rity verify performance deep learning model in clinic well that's a bigger question for field you know I think at least as I know in wpm we have several task group they are work on either for signation evaluation either for retion evaluation I'm sure you know they will have some guid to help our field that's it okay well well that was an awesome and exciting Talk and of course I'm thinking like okay how how how do I get enough budget to hire three or four more physicists to do all this um you know a a practical question uh with with h uh with HDR Breaky therapy um you know so I saw the structures the the the rectum the prostate the bladder the the other the other one that needs to be put in that data set is the pubic Arch because um you know if I'm not paying attention um you know uh we get a great plant that looks great but oh the only problem is there's three needles going through the pubic bone and that's that's not going to happen um per AUM we we do that but um you know that that would be something that would be really helpful I think in terms of of a tool and bottom line is there's really no ideal HDR prostate break therapy planning system right now anybody who wants to go over low customer volume environment and really enhance something that there's room to do that um you know I I looked at the probably hours of sleep you would gain me when you were showing you know the auto contouring for all the cardiac structures and I mean I think that is really worth pursuing because I think we could really improve a lot of our plans effen if we took the time to actually do that yeah uh it's very time consuming um it would be interesting to see how that system performs in yeah either when it's not ideal pathology or U Anatomy yeah because of either they've had prior uh coronary procedures done or other but that that was really exciting so I applaud what you're doing and hope you continue this yeah yeah thank you than you thank you great to come now you know just listening to your thought it's eye open it's wonderful thank you you know tradition when we done sample the images we don't have much concern when you AB uh you use lot from other place yeah and in the future you want to use this in the clinic who would be the best person to QA those say I I was think thetic phys yeah I think so yeah not just for religion anology but for probably there are many other departments that do not have PhD likey departments and then I was thinking of your customer base and this is radiation oncology and this is very exciting but I think diagnostic Radiology for improving the accuracy of of their reports would be potentially profound I mean your customer base there is what probably 15 fold or more yeah just like we do annual QA acceptance commissioning for treatment delivery I think probably in the future we should do something similar to I think yeah I think they are early stage we should do some comprehensive evaluation especially uh CL trial your wonderful