 Mae'r gwellaeth i adael'r llefydd yw gwellaethai Na gael'i amser gwellaeth elona. Elona dwyfodd â'r gwella yw hwn yn drefwyr eich gwyl yn ddefnyddio ffyrdd. so llefydd eich gwellaeth elona roeddeni gwella honno. Fe o'r gwellaeth yw amddangos dyn ni'n bryd, mae'n bunig bywyd i'r gwellaeth. Mae'n eu gwellaeth yw'r gwellaeth erionnol ychydig yn dda. Mae'n ffyrdd arfod yn dweud am yr oedol dynnu sydd wedi dirwais o'i gweld fel yr ydi'r cofnedd. Fa that was always there, but your confidence has really come along, and your personality comes through really well now which is great to see. Over the last three and a half years, Ellen has to have an uncanny ability now to predict the advice I'm going to give her. In fact, a lot of our meetings now go, she'll start telling me about something, I'll then interrupt her, go, ooh, ooh, what about this? And she'll skip to the next slide and she'll show it to me the result of what I was about to suggest to do. Ac rydw i'n ddweud yn gweithio. Rydw i'n ddim yn cael eu gwybod ar y ddweud i'n ddweud yma'r gweithio'n ddweud o'r ffwrdd. Rydw i'n ddweud, Eleanor, mae'r ffwrdd ddwy i'r ddweud yma'r ffwrdd mewn ysgolion rwyf yn ddweud. Rydw i'n ddweud, mae'n ddweud i'n ddweud i'n ddweud. Rydw i'n ddweud Eleanor yma, yna'r ddweud ei ddweud. Rydw i'n ddweud, Mae Rydw i'n ddweud yma, yna'r ddweud yma. Rydw i'n ddweud i mae'r hwn yn ddweud. Rydw i'n ddweud, mae'n ddweud ar draws a hwn yn ddweud. Rydw i'n ddweud, mae'n ddweud yma'. A fynd i'n ffeir i fod y Taxiwch yma yn ddweud lle'r ddweud eMA Gw iddod gwyddo'n ddweud i'n ddweud, felly mae'r ddweud ei ddweud, byddwch i'n lle'n ddweud, But it will be a really big deal for them. It's really important. Elinor's project is actually working on being able to treat lung cancer on the MLDAC. Lung cancer is one of the most deadly cancers still in the world today. It's not only important for me, it's important for Elector, just clinically as well. Before I hand over to just one more quick thing to say. Last summer we were at a conference in Montreal. Y Llector yma yw'r rhwng ymddiad cyflwyno ar Lleinarch yma yn Mwngtril. Byddwn ni'n ddylay i'r cyflwyno i'r fath o'r Llyfriddol yma. Mae'n cael eu gweithio ar yw'r gwrs. Cymru'n gweithio ar y gwaith, ddim yn ei gweithio ar y postwc, a wew'n gweithio ar y gweithio. Yr gweithio ar y gweithio ar y gweithio ar y gweithio. Rydyn ni'n cael ychydig ar y gweithio ar y gweithio ar y gweithio. a that for the first evening the other and i'n wasi i spent the last two days going through with ELECTR all the details of everything and they turned around and said yep, you're right the work that Eleanor's doing is a better approach that EleCTR were pursuing themselves so that was one of the proudest moments hearing them say that and finally just to say of the back of Eleanor's Ph.D EleCTR have recently agreed a new research agreement with UCL will by now funding more PhD's to come and it's no exaggeration to say that Alan PU is a massive part in securing that. So yes so without anymore, I'll hand over to Eleanor. Thanks Jamie, for your kind introduction so hi everyone. In my presentation I'll start with a little bit of a background on radiotherapy and ond gan y tarif y cynhwys cyfnod o gyffredigau Llan Cansir Yn Cymru. Mae'r Gwymor amdano, yma, yn edrych y Gwymor am y cyfnod ar y cyfnod Ammarlinac yn ei ddweud y gwerthu ar y Cymru, ac ein bod yn gallu'r cyffredigau cyfnod ar gyfer y cwyllgor a y ddweud y cyffredigau am ar Yn Cymru, ac rydyw'r cyffredigau cyfnod ar gyfer y cyfnod am y dyfodol. Rwy'n fyddwch gyffredigau'r cyffredigau, y cyffredigau'r cyffredigau er mwynhau o'r adnoddio cyllid yn unig am cael y cyfnoddau, ac yn ym 50% o'r adnoddau cyfnoddau yw'r adnoddau ymlaen o'r adnoddau. Fy gennym ni i gyd eich cyd-ddiweddau o'r adnoddau'r dda'u bod gyda'n gondol ymddiol. Efo'r adnoddau cerddoddau argynnu'r adnoddau o'r adnoddau cyfnoddau ar gyfnoddau ac gynnyddur yn cael y cyfnoddau ar gyfer ddoddau Plans in order to maximise the dose to the tumor and minimise the dose to the healthy tissues and in particular to the organ set risk. After a couple of weeks the patient start the treatment and the treatment is delivered over multiple days or weeks and the treatment is actually delivered over multiple sessions called fractions, Y rhaid o'r drwy flynyddiad yn ei ddweud yn ymweld i gael y planau oherwydd, a'r problemau yw'r anodololio cymaint, yr anodolio anodolio'u gyda'r rhaid o'r rhaid o'r rhaid o'r rhaid o'r rhaid o'r rhaid o'r rhaid o'r rhaid o'r rhaid o'r rhaid, er mwyn fydyn yn cael eu cyfnod. Mae'r gweld gan gymuned gennuriaeth yn ffocwst o'r gyda hwnnw gwirioneddau radd-radiol i gyda'r cyfrannu L-Lanc, Llancancillerau is the leading cause of death for cancer patients worldwide. It's characterized by a poor survival, equal to 5% at 10 years after diagnosis. So we really need to improve the way we treat these patients and we would like to do so by improving radiotherapy. One of the challenges when treating cancer patients with radiotherapy is given by respiratory motion. Yn dweud i ddechrau yn 10-15 dyn nhw, ac yn y pethau, ychydig yn iawn ar gyfer y gallwn y byd. Mae'r ffyrdd ymgyrch yn ymgyrch yn cael ei ddweud, ac mae'n amlwg iawn yn gwneud ychydig yn ymgyrch. Yn ymgyrch ar gyfer y byd, mae'n gweithio'n cyfrifio adeg ychydig a pob gwrdd ymgyrch ar gyfer y cerddau'r cyfrifio. Mae'r cyfrifio ar gyfer y peth yn ymgyrch o'r cyfrifio ar gyfer y peth sy'n gweithio'r cyfrifio. We've got also inter-cycle variations, so we can have a deeper breath and a shallow breath, and this would be the path of the tumour. But we've also got day-to-day variations, that means that the tumour and the other structures may move in a different way from one day to the other. And all these variations can cause, in the context of radiotherapy, some problems. So the first two type of variations are called the inter-fraction variations, and the last one falls under the inter-fraction variations. And these variations, as I said, can cause errors when planting and delivering radiotherapy. And one of the ways that the patients are treated nowadays is using, for example, gated treatment. In this case, the treatment is delivered at breath hold when, for example, in one specific phase of the breathing cycles, in particular the end-exhale. This means that the beam will be switched on only when the tumour is at the end-exhale. And this is problematic if we think about the breath-to-breath and day-to-day variations because the motion of the tumour during planting may differ from the motion of the tumour during delivery. So the end-exhale position may be different, especially if we think that the position in the respiratory cycle is actually estimated using external devices. So we don't have information about internal anatomies itself. Another way, actually, if we knew the position of the tumour during treatment, we may implement tract treatments in free breathing, and this means that we would shape the beam according to the tumour motion during delivery. And this is actually the idea behind the new technology, which is called the MRLINAC. So this has been implemented in some clinics now all over the world, and these systems basically combine anemar scanner with a linear accelerator. And for the first time, it provides us real-time images during treatment so we can image the patient's internal anatomy. And this is an example of some images that we can obtain from the MRLINAC, and actually this is myself and my internal anatomy during free breathing. So anyway, coming back to the MRLINAC, this is actually a really valuable tool and system because it may allow us to adapt the treatment to the intran and interfractional anatomical changes that I've mentioned before. However, not all the problems are solved yet because of the limitation of the MR imaging. So MR imaging cannot acquire data fast enough to provide 3D high-resolution images in real-time to monitor the tumour during treatment. So during treatment, just to the cinema images are available, and this means that we can follow the implant motion of the tumour, as shown by these movies here, but the motion of the rest of the anatomy remains unknown. And actually the knowledge of the 3D motion of tumour and organzaet risk now is still missing. However, it may be useful for accurate gated or tracked treatments and also is essential for accurate estimation of the delivered dose, including also the motion information. So to fill this gap, so this missing information of the 3D information, we propose to use surrogate-driven motion models. So these models take surrogate signals as input, and as output they can estimate the 3D motion of the internal anatomy. So in the case of an MR-Linac, actually these surrogate signals can be extracted using the internal anatomy itself from the 2D cinema images which would be acquired during treatment for tumour monitoring. And in this case we could have MR-derived signals input them to the model and get 3D motion estimates of the tumour and organzaet risk during treatment. So to build these models, we propose to use a motion model framework which was developed by Jamie, my supervisor. This framework combines image registration and model fitting into a single optimisation, and this allows us to fit the model directly to the answer or separate to the images which cover the whole 3D area of interest. This framework doesn't assume regular reading, so we can model breath-to-breath variations, and the model fitting can be iterated with motion compensated image reconstruction. So now we can have a look at how we build these models in more detail. So during model generations our framework takes as input, as I said, the answer to the MR images which cover the whole 3D anatomy and they can be acquired from multi-slice acquisition, and also the other input are seroid signals that in our case are MR-derived. As output during model generation this framework gives us a respiratory motion model where the motion is described as a function of the signals and we also obtain a motion compensated super-resolution 3D image reconstruction. So now let's have a look a bit more about the input and how we tailor them for the MR-linac. So the serogate signals, as I said, are generated from the 2D cinema images, and in this case these are basically from a fixed-slice location and they would be the images acquired during treatment by an MR-linac, and we call them serogate images. To generate the signals what we did was to apply the formable image registration to the serogate images to obtain the formation fields, and then we applied principal component analysis on the deformation field and we used the first and second principal component on the deformation field as serogate signals. So now that we know how we get the serogate signals we can have a look at the acquisition pattern that we proposed to acquire all the images that we need to build the models and which is tailored for the MR-linac. So here you can see the slice position of the images over time and these are the serogate images from a fixed-slice location over time which we used to generate the serogate signals. Now interleaved in time with the serogate images we've got the answer to the MR images which covered the whole anatomy and I hope that this animation helps. So the red frame indicates the serogate slice location which is fixed and interleaved in time with the serogate slice we've got motion slices in a sagittal orientation and now we've got interleaved with the serogate slices motion slice in axial orientation. In this way we covered the whole anatomy once and after doing that we applied a 2mm offset indicated by these arrows so 2mm offset to the motion slices and then we repeat the whole acquisition pattern again and we do this five times in order to facilitate super-resolution reconstruction. So all the images acquired with this pattern is what I would refer to as one repetition of data and it takes around three to four minutes to acquire those and I'm mentioning this just because in the next slides you will hear about two, three repetitions so you know what I mean by that. So now after all this background I can show you my current work actually I'm mentioning a bit about my previous work so in our previous work we've built the 3D models using the framework that I mentioned to you with the acquisition pattern and we built the models from MR data for voluntary data sets. We started with using long acquisition time so 10 repetitions up to 30 minutes and long processing time to build the models up to six hours and these times are not really suitable for the patients. However we assess the motion compensated image quality and the fitting error also for fewer repetitions and we saw that actually we could get good results with fewer repetitions down to five, three, one. So this allows us to move on to the patient data sets. So my current work aims at identifying the minimum amount of training data that we need in order to build motion models which give us accurate model estimates and this will be done for lung cancer patients. So basically our collaborators at the Institute of Cancer Research and the Royal Marsden are currently recruiting the patients for our study and so far we've got data from one patient, lung cancer patients, Canon and Marlinac. We got six repetitions of data so around 24 minutes and the specialist resolution of the acquired images was equal to 2x2x10mm cube. So in order to determine how much we can decrease the computational time and the acquisition time but still get good motion estimates, sorry. What I did was to use the first one, two, three repetitions as training set to build our models. So I ended up with three different models and three different motion compensated super resolution images with isotropic voxels of 2x2x2mm cube and then to assess the models, I've used the last three repetitions as a test set. So we can have a look at some of the results here. These are qualitative results. This is just a sagittal slice from the 3D motion compensated super resolution image that we can get from our framework and these are the three different cases with one, two or three training repetitions and hopefully you can see that the diafram and the vessels are quite sharp which means that our models were able to recover the motion quite well. And the image quality is pretty similar. Probably in this case we've got a bit more noisy image but still there is in the blurriness that we would have, hopefully you can see it, without motion compensation and here we've got the same diafram here and then the vessels with much more blurriness. So I just want to mention to you that the acquisition time for the three different cases was between four to 12 minutes and the model building time was go from 24 minutes to around 80 minutes and this was without any optimised code. So after that I can show you how I evaluated the models on the test set. So basically for each model we had from the building phase a respiratory motion model and a motion compensated 3D image and I basically used the surrogate signals from the test set and I input those to our models. And as output we obtain the 3D estimated motion in the forms of displacement vector fields at each time point. So then we use this displacement vector field to warp so to deform the motion compensated 3D image and obtain an estimated 3D anatomy at each time point. So the problem now is that evaluation is quite challenging because we don't have the ground truth 3D images and we don't have the ground truth 3D motion to compare our model estimates against. However, from the test set we've got the unsorted motion images which covered the whole anatomy that we didn't use so far. So basically what we did was for each time point we simulated from the estimated 3D anatomy the 2D image which corresponds to the acquired ones and then we did this comparison. And before going to this comparison I just want to show you a qualitative result in terms of the motion estimated by the models. This is the case of three training repetitions and here you can see a coronal and a sagittal slice of the 3D image which was animated by the estimated motion. And over there you can see the two signals that we used to drive the models. And from the first surrogate signals you can see that we've got breath to breath variations that our model was able to recover quite well. And so you can see that there is like a plausible respiratory motion that was modeled by our models. And here at the top now you can see the comparison just qualitatively between the acquired motion images and the simulated motion images at the right. In the middle indeed we have the comparison in the form of color overlay where basically the colored pixels indicate intensity differences between acquired and simulated images. And overall you can see that there is a good agreement except probably for very deep breaths. But obviously I want to say to you that many pixels that are colored here are just due to the intensity differences but they are not due to this alignment of the structures between the acquired and simulated images. So what we did was to try and quantify this alignment. So we performed to the imaginary station between the acquired and the simulated motion images and basically what the imaginary station gives you is a displacement vector field which tells you how much you need to move one image with respect to the other in order for the structures to be aligned. And this displacement vector field quantifies for us the residual error between the acquired and simulated motion images. So here I'm showing you the results in terms of mean and 95th percentile of the residual error in millimeter for along the different axes so superior and inferior, anterior, posterior and left to right for the different models built with one, two or three repetitions. And I just want to remind you that these are values computed over all the test sets which included the last three repetitions. So here you can see that we've got pretty good results with mean values around one millimeter so below the voxel size and values which are actually comparable between the different models. And the same applies to the 95th percentile with values just above three millimeter. So in summary, I'll describe to you our unified motion modeling framework that we used to build the 3D motion models from unsorted to the images and we applied this to a lung cancer patient from Anne-Marlinac using a tailored acquisition pattern. I've shown to you that we can get accurate reconstruction of the anatomy and its motion since we've got mean residual error around one millimeter over 12 minutes period which is actually comparable with the treatment time. And ongoing and future work will further assess and quantify the image quality of the motion compensated reconstruction and the model accuracy. And obviously we will use much more patient data sets from the Anne-Marlinac and the aim, as I said to you is to determine the minimum amount of training data that we need in order to build models which gives us good motion estimates. And finally, we will continue to work with our clinical and industrial collaborators. So with the clinicians, the focus will be to try and fit our motion models into the clinical workflow with our industrial collaborators. As Jamie mentioned before, we are trying to see if and how we can incorporate our motion modeling framework in something that is potentially suitable for the Anne-Marlinac. And if I have just 30 seconds, I want to just say that I'm really glad to be part that we're involved in this project because this is really a multidisciplinary project with clinicians, and my thesis engineers, computer scientists, but also with the presence of the company. And I think that this is really valuable and essential. And we are working towards trying to translate our academic research into something which can make a difference for the patient so I'm really glad. I know that there is still much work to do, but I do think that we are going in the right direction. So with this, I would like to thank all the people involved in this study, and also the founders, obviously. And I would like to thank you for your attention. I'm leaving you with my images. So, d'enternal anatomy.