Hello hi good morning my name is Antonio community and I work here Microsoft Research in Cambridge and the dam. I’m going to be talking about today is about the analysis of medical images in particular is about how to apply modern machine learning.
Technique to the automatic analysis of medical scans medical scans are like you know this one they were showing on the screen this is a CT scan of. A patient city stands for computed tomography and you can you know even if you’re not medically trained. You might not be you can see structures like the spine you know.
This is a view of the same patient you know horizontal view and this is a frontal view and you can see the pelvic region. And the autumn now what we’re trying to do here is you’re trying to. Come up with tools that can help doctors you know diagnose the problem you know with the patient and also treat you know the problem more efficiently in particularly what we’re trying.
To do you know we’re trying to make you know the analysis of medical images quantitative. So that you can really assess how big say a tumor is. You know how effective a certain drug is so we have two demos want to do with the analysis of your CT scans and the other one to do with the. Analysis of magnetic resonance scans of a patient’s brain so in the first demo we have already. Loaded in my patient image again you can see the different you know horizontal views here and the different side views of the patient here what we do is first. Of all we want to automatically identify you know the spine so the spine you know it’s important to.
Identify not just you know to be able to analyze whether there are broken vertebrae but also because the spine is a natural coordinate system for the patient so it is a very helpful thing. To do for any kind of doctor the you know wants to look at these sort of images so you have seen that you know I have clear. From the run button and something has already happened very very quickly but it is not you know quite as you know refined as it could be so then I launched a second. Round of refinement and you can see the position of the spine and the individual. Vertebrae you know been refined you’re on the go so now it’s done after a few seconds you already have a very good idea.
Of where the spine is and not only that you have the exact name for each of the vertebrae so you might think that this is something that any. Trained radiologist will be able to do very very quickly it turns out that it is a little bit you know time consuming. For them to to be able to do these sort of things and yet they have to change between different views and sometimes they have to resort to actual counting.
You know from you know the head towards the bottom to figure. Out which vertebra they’re looking at so we can also compare and really. Quantitatively measure how well we do with respect to some ground truth labeling and so now you can see in red they automatically detected vertebrae and in yellow that the. Ground truth position you can see that there is very good you know alignment between them what’s important is that you know in this in this demo we can detect the spine.
In which vertebrae were looking at no matter what type of CT scan you know we are looking at no matter what the resolution is or no matter what the scanning machine that we actually. Used and more importantly no matter how much we can see of the human body in this case we can only see the. Abdominal part and the thoracic region is missing we do not need to see the whole body we can have a very you know tight cropped view and the system still.
Works very effectively so this is quite important the journal ization probability. Of this is quite important so this is the first demo the second demo is to do with something a little bit different so we. Have a different type of image we have brain scans and they these scans are acquired with a different type of you know scanner it’s called magnetic. Resonance so here is an example we have four views horizontal views through.
A patient’s brain and once again any of you can you know easily figure out that there is something wrong with this brain. In particularly in this region where we see a lot of you know white and textured material where that is a brain tumor it’s a. Very nasty aggressive type of brain tumors called glioblastoma and anything that we can do with you’re measuring the size of the glioblastoma and its progression will be extremely.
Extremely welcome in a medical world because honestly this is a. Very bad in fact incurable disease so what we want to do.
Is we want to you know develop an automatic analysis tool which can tell us not only that this patient has got glioblastoma and how big. It is but also we want to figure out what are the different component regions in the tumor this is color coded here in this view where you know blue means healthy. Tissue and then green means necrotic hole which is the already damaged in fact dead region of the brain red is the actively proliferating region where there is a lot of you.
Know oxygen you know blood being drawn into the area because you know the the. Tumor is feeding itself and growing very rapidly and yellow is what’s called a edema is the inflammation where it might be possible they you. Know you find you know tumor cells in there as well but in general is. Just you know healthy inflamed region but it’s something to watch you know very carefully and so what do we have here we have a another application and just gonna you.
Know select a patient drag the images there the images get loaded and here you have you know the tumor so what you see here. These black regions this is all healthy they’re called ventricles all this region here is all healthy but this. Region here where you see this black and you see in the other views as well. The the lateral and the frontal view that’s the tumor in particular the black region is the necrosis where the.
Brain cells are they died and around you know you see an enhanced rim a white rim and that’s the proliferating edge so once again what I do is I just you know run. My segmentation tool and all I did is you know click on rank segmentation and what happens is. The in front of you this segmentation you know is you know color-coded it’s computed color-coded and refined as we speak so this works for. You know any patient you know no matter what the size of the brain is no matter what the size of the tumor is all the location of the.
Tumor the other thing is that these tumors have got very different appearances and so you really you know the only way you can tackle this problem is by using machine learning techniques which. They learn from labeled data and so that’s precisely what we’ve done for both projects we have a lot of you know images of patients which have been labeled by our. Medical expert collaborators and then we have devised some you know algorithms to you know exploit that and to. Learn models that can be applied to new previously unseen patients so we really hope that this technology can be adopted you. Know soon in hospitals we already have a number of trials going on and your help you know the you know improve the quality of.
Life of patients who suffer from very nasty diseases thank you very much you.