false
Catalog
Practical and Advanced Spine Imaging (2023)
RC10519-2023
RC10519-2023
Back to course
[Please upgrade your browser to play this video content]
Video Transcription
We'll spend a little bit of time talking about artificial intelligence and spine imaging. We will focus on things that are real here and today, although at the end of the presentation, if I haven't gotten the hook by then, we'll talk a little bit about where we're going. So a little bit on what is AI, and then we'll talk about the impact in spine and a little bit of a preview of what we're going to cover over the next 20 minutes. My timer is set at 10, and it's not running. Can we change that to 20 and make it run? So AI, basically, is a branch of computer science that allows machines to do things that are normally associated with human intelligence, reasoning, learning, and self-improvement. AI is all around us, as you know. Virtually everything we do is controlled, in some extent, by machine learning and the like. As a matter of fact, this was last year's panel at the arts, and you can see she's not looking, but that's Dr. Gibbs over there in the red, and Carlos and the rest of the team are over there in the right. So this is the way things are moving these days. This is a really nice slide that you, Harvey, made that shows the use cases in radiology for artificial intelligence, and it can involve anything from the front office, things like utilization management, no-show management, things that we have going on in our institution, as well as affecting the way reporting is done and harvested. But we're going to concentrate our work around these last three here, from order-weighted acquisition through segmentation and a little bit of computer-assisted diagnosis. As you know, walking around the AI theater, you'll see there's an awful lot of companies that are out there working in this space. So we'll talk less about companies and more about some of the applications, and we'll start off with automated scan acquisition. Now automated scan acquisitions are really based on a form of machine learning, and they do things like auto-aligning to the angle of the spine, adjusting coverage so you don't have to do it manually, labeling the spine levels, as you see here, even doing curve planar reformatting scouts. And it has been shown, this is an example of one of those from the Siemens family called the dot engines. It's been shown that they were able to make less fuss with the exam, 74 fewer repositionings, 26% less user reaction. It's a more efficient process. In our institution, we had the same software, the same protocols, the same readers, the same schedulers, the same scanners, everything the same except for one piece that was different between the two institutions, and one had the automatic scan acquisition and one didn't. And you can see from the bar charts there with an end value of about 100 apiece that there was a significant difference, a couple of minutes difference, in terms of how long the exams lasted. So just a nice tool for accelerating spine acquisitions in MR, and pretty much ubiquitous. I think GE is really the only major company that doesn't have something in that space, and they're working very actively in that space to get a tool in there. And things like automatic fusion of the spine, I loved reading these when I was at Mount Sinai looking at cancer cases, so you don't have to keep looking back and forth, back and forth, back and forth to get the localization. I liked looking at these, and I would have a blown-up view on one side, the holistic view on the left, and it made it a lot faster. Now one area where we're seeing an impact that I think you can see every single day is an automatic scanning and reconstruction for CT of the spine. This is actually a movie from a GE system. I'm going to show you some of my personal work from Siemens systems, but you'll notice how the axial slice follows the angle of the spine, and even in the coronal plane it is angling to the disk space in a way that you'd never be able to do manually yourself, but it just does this automatically. You can actually see the way this plays out now in this particular demo. I've selected slices on the left to show you, but you can see how challenging this spine would be and how awkward it would be to try to read these slices, and it just comes out that way. Here you can actually see from the Siemens interface what goes on. If you look, you'll see a little mouse moving around, you know, no point to my hands, leave my wrists, and you see this happening on the screen. You see it sort of estimating the angle of the spine. It gives you the chance to verify, and then the images just come out. This is actually the direct feed off of the scanner, so you can see the kind of thing that goes on. Now, again, this has been out there for a while in the Siemens camp. It's called Fast Spine, and again, as I showed you, other companies like GE are working to create this, but you can see how nice it is to get those images in the appropriate plane. It just makes it easier to read. Now, we did a study when I was at Sinai, which we presented at the Eastern Neuroradiologist Society, I think also at the Spine Society, where we actually looked at the value of these adjusted angles by the machine. We gave it a fancy name, but it's basically what you just saw, and showed you the difference. We looked at the difference between the slices that we would have gotten orthogonally and the ones that we would get adjusted by the machine, the machine learning algorithm to, you know, parallel the spine. Here you can see from this particular slice the difference in the angle. In one case, we propagate the bone spur behind, looks like C5, into the spinal canal. The other one, you don't, and it turns out in this particular trial, there was a significant improvement in the qualitative ease of interpretation, accuracy of depiction of the canal dimensions, and greater effectiveness of communication by disc level in cervical and lumbar spine CT with these tools. So, a real available tool, machine learning based, that we can use. Now, there are also adjuncts to this in terms of bone reading. It will do things like label the spine for you, just as you'd expect, and if you end up reading any of these trauma surveys, or even oncology surveys, these views of the ribs, which are also part of the same type of an algorithm, are kind of useful. And, interestingly enough, you'll look at this particular study, you can see my mouse, can you see my mouse? You can see there's a met here, there's another met there, there's a few others, none of which were called on this particular exam, because it's hard to read ribs. But, when you see them like this, as a matter of fact, you know, you can think of these as ribs that you can spin them around as you eat them. See, you can spin them around, look at them in three dimensions, and actually look at them obliquely, and you can see the met inside of, looks like the left sixth rib. It really does facilitate review. So, automated spine, automatic bone, all things that are available out there to a certain extent based on this type of technology. All right, so let's talk about image reconstruction, which is a topic near and dear to my heart and the spine. Again, something that is either here, or is imminent, you know, imminently available, and we start off with compressed sensing. As you know, as opposed to collecting every line in K-space, which is fairly time consuming, we've been doing various forms of undersampling over the years, to the point now where we do this pseudo-randomized undersampling, extensive undersampling, and we count very, very few points. Okay, this is essentially an iterative reconstruction process, and when the data spars pretty much white and black, you can pretty much accelerate, you know, at least 100 percent, as you can see here, where our four and a half minute angios now take about two minutes. We use this across the board in our techniques to get about 30 percent faster speed than we normally would in day-to-day practice. This is from a Philips scanner, and Philips probably has the best, most extensive implementation of compressed sensing on routine applications. You can actually see what our protocols looked like before, and what our protocols looked like after compressed sensing. You can see we picked up about 30 percent across the board. You can see here our routine protocol with and without contrast, and our routine protocol with and without contrast after compressed sensing, and you can see this technique, iterative reconstruction technique, in a sense, gave us that 30 percent boost in scan time. You see the same thing in the lumbar spine. We're a five-sequence lumbar spine, which tends to be a long exam. It also tends to be the most common exam done on clinical MR scanners, so the time benefits are, you know, really scale quite a bit. You get down to about 11 minutes, but these tools, all of these tools, will give you the option to go faster, to do better quality, or even higher resolution, depending on how long you want to scan, in this case, running from less than five minutes for a lumbar spine all the way through about 11 minutes. But you can see the example. This would be our fast lumbar spine with compressed sensing. This is a sort of a standard lumbar spine with compressed sensing. If you look, you can see the matrices on the bottom all the way through what you would call a high-resolution lumbar spine with a scan time just a little bit longer. You know, a sacrum can be an exam that can be exasperatingly long, really taxing the patient's ability to tolerate, and you can see from this example, excuse me, we've gone from 20 minutes to about 10 minutes. When you start doing these plexus exams, you know, it can really get hairy, and it's nice to be able, again, use compressed sensing across the board to make these acquisitions reasonably, concise, and short, so the patients will hold still and they will be tolerated. It's very nice to make pretty images when you can get them, and when the patients can't hold still, this can be a significant retardant to just, you know, getting what you really need, which is the value of the exam out in a patient-focused, tolerable amount of scan time. All right, so let's talk about what we associated with iterative reconstruction. Anybody know iterative reconstruction from CT? Everybody's familiar with that, the stuff that's out there? This is the technique that allowed us to use really low doses. Doses so low that with our old techniques, if we reconstructed them, they would be too noisy, but with modern techniques, you could get that noise down and get your dose performance really low, right? This technique is actually available for MR right now. It's commercially available. This is a machine-learning-based iterative reconstruction tool. It does image enhancement, you know, by cross-validating across your sequences, using one to refine the next. Dr. Gibbs, our moderator, and Dr. Johnson, and me, actually, and a few other folks, did this particular trial where we did brain, lumbar spine, and cervical spine. I'll share some examples with you from the spine. Here you can see the example, the standard example at two and a half minutes, the faster example processed with MR iterative reconstruction in about 50 percent less time. The image quality looks pretty good here, but qualitatively, excuse me, qualitatively blind read, three being equivalent, you see the ratings came out about the same. So, no, the iterative reconstruction tool was able to create the image quality we expect, even though we accelerated the scan time. The cervical spine was, again, part of this. This particular example is a 32 percent reduction, but again, you can see with three being equivalent, we managed to hit equivalence in this circumstance, and for the interest's sake, the brain part of the study was also at equivalence. So, the study proved that you could use iterative reconstruction to accelerate images and maintain quality. All right. So, thus far you're saying, this is not a lot of AI, give me some stuff that I associate with AI. Well, let's talk about deep learning-based reconstruction. You may know from walking around the floors that there are several companies now marketing FDA-approved CT image reconstruction done with deep learning technology. You've got GE's True Fidelity, you have Canon's ACE, and the Korean booth is over there. I mean, there's a booth called Clarify from Korea. They also have a tool. And you say, well, why do we need more iterative reconstruction tools, and what do we get out of them? Well, let's talk about what these are for. So, let's look right here. This is filtered back projection, our standard image. It's a CTDI of 80, which is, you know, about what you'd do for a head CT in the old days. And you can see the standard deviation, which is the way you measure noise, is 15. If we go down to today's doses, our noise goes up to 32. So, with only about 10 percent of the dose, our noise goes to 32. If we use traditional iterative reconstruction at the same low dose, we can bring the noise down to 20. It's not the noise level we had before, but it's much lower, and it's generally tolerable. If you use deep learning reconstruction, however, this True Fidelity Recon, we can bring it down to the same noise level. Now, we're accustomed to the noise level performance, so, you know, that's okay. We're not shocked by any of this, but, you know, we know that we've always been a little unhappy with iterative reconstruction techniques. Here's the filter back. Here is the Acer iterative reconstruction, Acer V. We can use these tools to get an image quality that is more akin to what we associate with traditional images. Right? So, message one, deep learning reconstruction will give us CT images that look like the old images, as opposed to the new images. Now, this is, you can quantify this, by the way. You'll notice with FBP across different materials, the behavior is very consistent, whereas, with iterative reconstruction, whether it's Acer type or model-based, is across the board things change. So, these are some of the benefits of deep learning reconstruction. You can see filter back on the left, standard, deep learning on the right. The phantoms look almost identical. It maintains the traditional behavior across tissues. You'll have the traditional look, the traditional counts for the unit values, and we know that's not the case with iterative reconstruction with traditional techniques. More importantly, when you say, I don't like the texture, what you're really saying is the noise power spectrum has changed. Okay? And it turns out, if you look at this chart, you've got your standard iterative reconstruction as a significantly different noise profile than what you see here, which is the overlay of the deep learning and the traditional filter back projection. So, what this chart says is the noise power spectrum is the same for deep learning as it is for filter back projection, and it doesn't create the weirdness in the image. And what's the weirdness in the image? There's the weirdness in the image. That's Acer V on the left, and that's deep learning on the right. Notice there isn't any of that bizarre texture on the images, and that's the advance that deep learning iterative reconstruction brings. Your grandpa's CT scans at your nephew's doses. That's really what you get here. Now, well, don't I lose resolution? Well, no, I don't lose resolution. There you see different settings in the deep learning recon. You can see how well the phantom is rendered. And if you want to do this, you know, sort of a Flintstone fashion, let me give you an example of a filtering algorithm that does throw away resolution. So on the far left, you see the standard, you see a very smooth oriented denoising, and you see a sophisticated deep learning denoising. If you subtract this from this, you see there's a lot of structure loss. That's the structure you've thrown away. Notice when we subtract the deep learning from this, the image is empty. We haven't thrown any structure away. And this is a quantitative way to say we've preserved all structure and all resolution. Here you see a filtered back projection. Here you see the deep learning reconstruction. And here you see if you subtract the two of them, you don't have anything left, meaning you're not throwing any information away, a big advance over where we were before. Just for the sake of it, the spine section, here's a deep learning spine image. They look like what you'd expect images to look like if we even remember what they looked like before iterative reconstruction. You can see here the differences. This is the true fidelity in the lower half, the standard on the upper half, and they look like spine images. You don't have any of the weirdness, any of the odd texture that you associate with these techniques. Thus far, I've shown you GE techniques. Canon has a tool for iterative reconstruction for CT, which they do in image space, much like the deep learning you saw with MedicVision. And I mentioned also Claripi. I don't really know much about their software, except I believe this shows the same noise power spectrum. This shows the lack of alteration of the structure, the same type of thing I was just extolling. So, you know, deep learning reconstruction is a new version of CT reconstruction that doesn't give away the things we gave away before. Now, what about MR, deep learning reconstruction? And again, you're starting to see these things out there now. You can see it from GE as their prototype scanner. I believe Toshiba or Canon is already showing their ACE scanner, and SubtleMR is out there, you know, doing their work, and I'll share some of their work over the next few minutes. I would like to point out that this is a .9 millimeter 2D FastBand Echo scan, .9 millimeter two-dimensional FastBand Echo scan that has been restored to excellent quality with deep learning. Now, one of the nice things that deep learning tools can use, particularly these ones that are built on neural networks, is they can be trained on artifacts and can remove them. If we turn off the apodization filter on a spine, we get these truncation artifacts. So, you know, why not turn it back on? Well, if you turn it on, you lose some resolution. So if I turn it back off, the resolution comes back, but I get truncation. These tools can be trained on that truncation and remove it, preserving the resolution. Here you see that in another case. We've turned off the filter here. We've got truncation here, but we use deep learning and restore a nice image. Turned it off here, so we've got truncation. The deep learning has restored that. Focusing in a little closer, look at the shape and the size of the spinal cord. That truncation makes everything look old, makes it look low resolution, and it interferes with our ability to interpret. When you use this tool, you can get rid of that artifact. So it turns out we have a scientific poster on this particular tool at this particular meeting. I encourage you to go look at it. We looked at 28 patients. We compared what came off the machine to what the deep learning algorithm could do with the raw data off the machine. So this is case-based data. Here's some example of a stir. You know, it's hard to make a beautiful stir. Well, you cheat and use deep learning, you can do it. Here you can see your typical 3D gradient echo. Excuse me, this is a 2D gradient echo, and you can see how nicely the image quality is restored. Again, you see, and if you look at them, you get a sense that the spatial resolution is a little bit higher. If you really want to get a feel for the difference in these images, on the left is traditional, on the right is deep learning. Look at the noise in front of the spine, and look at how the noise has been effectively removed without smoothing other structures. It's a nice-looking image on the left, but look what happens when I get into noise. Notice how it really is not conveyed into the image, and the image is actually, you know, in our trial, you know, the perception of sharpness was higher on these tools. And if you look closely, this is a nice-looking spine exam. You wouldn't complain about that. But look at this annular fissure here, look at this high-intensity zone. How much more of it you appreciate on the higher-resolution deep learning reconstruction than you do on the really high-quality image in this particular case. Taking a zoomed-in look, you can see more of it. You know, in general, low noise, higher resolution is a good thing, and in our particular trial, you can see the ratings. We add statistical significance for all six or seven of these criteria, six of these criteria. Just to show you that Canon has a tool like this, too, it is a lot more like the Medic Vision than it is like the GE, I understand, but you can see what you can do. You can go in and make an image that is impossibly high-res, and actually use these tools to denoise effectively, still maintaining structure. Okay, I did mention Subtle Medical, because Subtle Medical has an image-based-based deep learning tool based on neural networks, and we did a trial with them at our institution. Dr. Gibbs and Dr. Basher in the audience participated in this one, and we simply said, let's take our scan time down by two-thirds and see how well deep learning can restore the image quality. Okay, you see the standard of care on the left, the accelerated scan on the right, and the deep learning process scan, and this is an image space in the middle, you know, how well it's restored. Here's another one. Again, standard of care, fast in the middle, deep learning, fast, and in this particular trial we managed to get the image quality numbers that we expected to see. We ended up with these scores, and I'll show you again here, just another example. Here you can actually see the noise in the spinal column that you don't see here, but in this particular case, in this trial, we showed that the images were equivalent, even though we accelerated it by two-thirds, so only a third of the scan time, and as I mentioned, these guys have this. I didn't mean to show this twice. I think I moved it. Didn't show it. All right, a couple minutes left. As you know, there are AI-based tools that are being used in urgent finding triage. A.I. Doc has a cervical spine tool that will pick up these subtle fractures, often before they're even read, which might alert the team at this patient or this patient. See the arrows here? Or this patient who is walking around the department getting the rest of their work up maybe should be in a collar before definitive management comes up. You know, these tools will actually do basically one trick. You know, in this particular case, the trick is identifying cervical spine anomalies or fractures and let the user know that there's something to pay attention to on the horizon. Now, in terms of going to the next level, I'm going to put up, you know, bring up a couple papers here that I think are really nicely attuned to what's next in the spine. Both images are there. We are seeing that things like localization and labeling of spine structures out there. We're seeing things like segmentation, both disc and end plate and vertebral bodies. I'm sure many of you have seen this paper on quantitative analysis of neuroforamina, things like segmentation of foramina, segmentation of brain structures. It's really low-hanging fruit for a machine learning algorithm. It does a very nice job of those tools. And when you think about what AI is going to do for us, do we really want to spend our time measuring and segmenting? Or if you start to get these measurements that are more standardized, would that help us with common data elements and structured reporting and data mining? All of those things. You know, how do you manage a scoliosis? Where are the vectors? Can you say enough words to inform everyone appropriately on a scoliosis case? It's nice to have these AI tools that can help. And there are tools looking at things like the generative changes, Furman grading, you know, getting more quantitative with your spine assessment, again, as a prep before you get to the exam. We've also seen, you've given you some hints of other areas where people are working on CAD for spine imaging, helping with fractures, looking at lytic and blasting mets, differentiating tuberculous from pyogenic spondylitis, and looking at plaque and multiple sclerosis. There have been studies looking at biomechanics. This pretty diagram here is actually using AI to highlight this altered mechanical stress in the pedicle, as well as looking at things like outcome prediction, even clinical decision support, in this case, meaning utilization management and content-based image retrieval. So I hope that was a worthwhile run through some of the applications, mostly the ones we have today, as opposed to the ones we are dreaming about tomorrow. But bear in mind that before you know it, it will be tomorrow. And the next 10 years, I doubt we'll ever look at any of these exams without having some computerized intervention to get us on our way and focus our talents where human talents are best devoted. So I'd like to thank our moderator, the rest of the panel, and all of you for listening. Thank you for your attention. We're going to be talking about advanced imaging of peripheral nerves. And we're going to focus on the lumbar sacral plexus, as well as the peripheral nerves of the lower extremity. EMG and nerve conduction studies lack the anatomic detail for precise localization. They are difficult to perform in infants and young children. And furthermore, there might be a delay of several weeks after nerve injury for the test to be positive. And that's where MR neurography has a very useful problem-solving role. MR neurography is really tissue-selective imaging that is directed at identifying and evaluating the characteristics of nerve morphology. The visualization of the vesicular architecture of the nerves is made possible by exploiting the differences in the water content and the connective tissue structure of the fascicles and the perineurium compared with the surrounding epineurium. And what can MRN be useful for clinically? Well, it can help to localize nerve injury, confirm peripheral nerve entrapment, help in detecting and characterizing underlying neoplasm, and is particularly useful when you have indeterminate results of EMG or nerve conduction studies or lumbar spine MRIs. So at UCSF, we have mostly GE scanners, and this is our protocol for the lumbar sacral plexus. And we rely on these IDEAL sequences. IDEAL is really a useful imaging technique that provides uniform fat suppression while optimizing the signal-to-noise ratio, and it can be used with a variety of pulse sequences. And this particular sequence nicely discriminates between normal and pathologic nerves from adjacent vascular structures. We don't always give gadolinium to patients. That will be administered in patients with suspected neoplasm or infection or radiation versus malignancy question. DTI is done on every single patient that gets an imaging of the lumbar sacral plexus and peripheral nerves. We prefer imaging at 3-Tesla because of the greater and higher signal-to-noise ratio, and we'll prefer to use 3-Tesla MRI scanners even in patients that have hardware, just because of the advantages that 3T offers. Now, one thing I'd like to mention is that when you're designing a protocol for a particular patient, remember that you need to tailor the imaging protocol. You need to be, and then when interpreting the study, you need to be familiar with the anatomy. That's a given. And you must know the results of the EMG and nerve conduction studies. So here is a picture depicting the morphology of the peripheral nerve, and we see the different connective tissue layers of the spinal nerve, the outer epineuron, the central perineuron, and the inner endoneuron. And it's thought that normal endoneural fluid, when protected by an intact perineural blood-nerve barrier, is responsible for the normal signal intensity of the nerves on T2-weighted neurography. And it's thought that when there's an injury to the nerve, that increases the signal of the endoneuron and the periphysicular connective tissue, and that increased signal is what we're seeing on our T2-weighted images. When interpreting these studies, I have a checklist that I think is useful to keep in mind. You, of course, want to look at the nerve of interest, but you also want to look at the surrounding fat. You want to look at the muscles that are innervated by the nerve in question, and you want to look for anatomic variance. So this is my interpretation checklist. So I'm looking at the size, the signal intensity, the fascicular pattern, the perineural fat, the course and continuity of the nerves. In terms of size, the peripheral nerves should be similar to the adjacent artery. And on T1, they are relatively iso-intense. They might be slightly increased in signal on T2, our ideal sequences. And the fat planes around the nerves should be preserved, like in this case of a normal sciatic nerve. And there should be no focal deviations of the nerve. When you see a focal deviation of the nerve, that's abnormal, and that can be from scarring or other pathology. And the nerve should be continuous. And normally, there should be no enhancement of the nerve except for the dorsal root ganglion. And here's just a coronal T1 image showing that nice, normal sciatic nerves with the fascicles identified and the perineal fat preserved. Here's a coronal ideal sequence showing the nice, normal signal intensity of that sciatic nerve. One word of caution here is that remember that normal peripheral nerves may have variable T2 hyperintensity, which may occasionally lead to a false positive diagnosis of neuropathy if you're not careful. Do not, if you see isolated T2 hyperintensity of the nerves, be careful about calling them pathology because we can see focal isolated T2 hyperintensity of the nerves, particularly at the level of the sciatic notch, sometimes attributable to magic angle artifact, or sometimes just due to varying degrees of endoneural fluid. Therefore, mild T2 hyperintensity in isolation should not be used to make a neurography diagnosis of neuropathy. Here is a coronal T2 ideal sequence showing an abnormal neurogram in this patient with Bechette's disease, a systemic disease with diffuse hyperintensity of the sciatic nerves, and that's what we see with systemic diseases. We see often bilateral and diffuse findings. In abnormal settings, we may see focal or diffuse increase in size of the nerves. We may see increased signal intensity of the nerves, and we might see that the perineural fat is effaced. And we might also see enhancement, like in this case, of an autoimmune sciatic mononeuritis, where we see this diffuse enhancement of the sciatic nerve. And the enhancement can be due to breakdown of that blood nerve barrier from inflammation, infection, or tumor. Now, denervation changes in muscles can be seen quite soon and acutely, within one to two days after nerve injury. In the acute setting, we might see edema that is reversible. In the subacute setting, there might be edema and fatty infiltration, and this finding may also be reversible. In the chronic setting of denervation change in the muscle, fatty infiltration and atrophy might be present, and these changes are usually irreversible. In general, the edema pattern in the muscles is distal to the insult, and there is no fascial edema. So this differentiates edema from other causes. So before we look at some clinical examples, I'd like to briefly talk about DWI and DTI of the peripheral nerves. So diffusion tensor imaging of the peripheral nerves exploits inherent anisotropic diffusion of water along the peripheral nerves to generate signal. And DTI really provides a functional imaging of peripheral nerves, and tractography can also be very helpful, particularly to facilitate pre-surgical planning, as well as for post-operative follow-up. Remember that the fractional anisotropy is really an estimate of nerve fiber tract integrity, and it can be decreased in nerve injury, particularly from axonal degeneration or demyelination. And ADC gives us an estimate of immune diffusivity, and it can be increased with injury to the nerves as well, as in states of edema or inflammation. So this is the DTI protocol at UCSF. We use a single-shot EPI, typically 12 to 28 directions. And the B value, this is a typo here, our B value for the lumbosacral plexus is 600, for lower extremity peripheral nerves is a B value of about 1,000. And it's about a three to five minute imaging sequence. So let's now talk about some of the practical uses of MR neurography and DTI for lumbosacral plexopathy and lower extremity peripheral neuropathy, particularly for assessment of peripheral nerve injury and monitoring treatment response, for preoperative planning, and for characterizing peripheral nerve pathology. So let's talk about nerve injury. So there are three major types of nerve injury. One could have a neuropraxic injury, which is essentially the mildest form of nerve injury, which is a myelin sheath injury. And here on MR neurography, all we might see is mildly increased signal and size of the nerve without any muscle denervation. A more severe injury is an axonotematic injury, where there's an injury to the axon, but the surrounding connective tissue structure is preserved. And the most severe injury is a neurotematic injury, where there is both an injury to the axon and the surrounding connective tissue. It is critically important clinically to determine the type of nerve injury, particularly because neurotematic injuries do not spontaneously recover. Furthermore, EMG may not differentiate between severe axonotematic injury and neurotematic injuries, but MR neurography with DTI can. And as I mentioned, with a neuropraxic injury, all we might see is mildly increased signal of the nerve. With an axonotematic injury, we might see a greater increase in size. The fascicles might be effaced. The perineal fat might be effaced. And we might also see muscle denervation. And with the most severe type of injury, a neurotematic injury, we would see nerve transection or a neuroma incontinuity. A neuroma is essentially scar tissue that intrudes itself at the site of nerve trauma, and it prevents regenerating axons from advancing into the distal nerve. So it's very important clinically and radiographically to be able to identify the presence of a neuroma, as in this case. This is a 32-year-old unrestrained driver that was unfortunately injured in a motor vehicle collision. And post-operatively, the patient had low back and left leg pain. And we see on the coronal T2 ideal sequence, we see increased signal and enlargement of the L4 and L5 nerves. And we also see a mass-like soft tissue enhancement consistent with post-traumatic neuromas. So this patient underwent a surgical reconstruction. Here's an example of a traction injury in a post-operative patient. This patient was a young patient, again, who had a total disc arthroplasty at L5-S1. Immediately post-operatively woke up with a right L5 and sciatic neuropathy. And on the MR neurography that was done on 3T, despite the artificial disc, notice we can still see the extraforaminal L5 nerve really well, and it's enlarged and edematous. The FA values on that side were reduced, and the tractography shows the nerve to be thickened. So this was presumed to be a traction injury, and this injury resolved on its own in a few months. Here is another patient with a sciatic, and specifically a common perineal neuropathy after a total hip arthroplasty. And the images that were obtained a couple of months after the total hip arthroplasty show that there is this very abnormal appearance of the sciatic nerve, which has low signal intensity on the T1, which indicates fibrosis, and this is high in signal on the T2, and it's affecting mostly the lateral fibers of the sciatic nerve, which contribute to the common perineal nerve, which explains why this patient presented with a common perineal neuropathy. And remember that total hip arthroplasty is the most common surgery that results in a sciatic neuropathy, and it may either be due to an intraoperative direct injury, a traction injury, or an injury from the hardware, or a postoperative complication such as compression from a postoperative hematoma or fluid collection. So this was very important for us to show the referring clinician, and this patient underwent a surgical neurolysis of the sciatic nerve. Here is another example of a traumatic injury to the nerve, in this case, a perineal nerve transection that was confirmed with DTI. We notice on the axial MRI images that it's very difficult to identify the nerve and to see if it's intact because of the edema and hematoma that was surrounding the nerve, but the tractography shows that there was actually a transection of the common perineal nerve with about a three millimeter separation between the two ends of the nerve, and we see that the nerve is focally thickened at either end, suggestive of neuroma formation, and so this patient also underwent a surgical reconstruction. This is an example of a neuromedic injury. So DTI can be very helpful for detecting early nerve regeneration. This is a very nice study in which rat sciatic nerves were subjected to contusive injury, and in this study, the fiber tracts returned to the pre-injury level by three weeks. On the left panel here are fractional anisotropy values, and here are the neural tracts by tractography, and in this study, the FA values at the lesion epicenter strongly correlated with the parameters of motor and sensory functional recovery, and the DTI results were consistent with histological and clinical findings, and we're seeing this on a practical basis as well, and here is an example. This was from a study published from our institution a few years ago. This was a young patient that suffered a laceration injury from a saw in the popliteal fossa, and this patient lacerated the deep component of the common perineal nerve, and we notice on the preoperative MR neurography images with tractography that there is a stump where we would normally see fibers of the deep perineal nerve, and one month after a sural nerve graft was done, the tractography can actually show these fibers regenerating across those sural grafts, and 13 months later, this nerve is completely regenerated, so we use DTI clinically in this postoperative setting, and in this case, axonal regeneration occurred within one month of nerve repair, consistent with the study, with that rat sciatic nerve study I showed you. Serial DTI studies are useful in following the course of axonal regeneration, and remember, the EMG and clinical exam findings are limited in this setting, because until the axons reach their target muscle, EMG and clinical exam will show no signs of functional recovery, but DTI tractography will, and regeneration is anti-grade at the rate of one to five millimeters per day. Now, in the few minutes I have left, we'll talk about peripheral nerve neoplasms, so remember, the tumors of the peripheral nerve can be characterized as intrinsic or extrinsic, depending on whether they are of neurogenic or non-neurogenic origin. Intrinsic tumors, which are the most common, can be benign, for example, neurofibromas and schwannomas or perineuromas, or maybe malignant, such as a malignant peripheral nerve sheet tumor. Extrinsic tumors are primarily malignant, and examples include metastatic disease, of which the breast is most common, but there might be other benign lesions, too, such as lipomas and lymphangiomas and fibromatosis. In this setting, DWI and DTI, particularly tractography, plays a very useful role in preoperative planning, in helping us tell the surgeon whether the fibers are displaced or infiltrated or disrupted, and ADC can also be a very useful biomarker in helping to diagnose benign versus malignant tumors and in selecting patients for biopsy and also for monitoring response to therapy. So here are some examples. This is an example of a left popliteal mass, which was actually a schwannoma involving the distal sciatic nerve, and we see that the mass is closely related to the proximal tibial nerve, and the tractography shows that the nerve fibers are intact, but are displaced peripherally, eccentrically along the mass, medially and posteriorly, and this is characteristic of a schwannoma. In a neurofibroma, the nerves would be infiltrated, the nerve fibers would be infiltrated through the mass. So we could confidently say this was a schwannoma. We also measured ADC values, and in general with the peripheral nerve lesions or tumors, if ADC values are 1.3 times 10 to the negative three or higher, they're more suggestive of benign entities, and less than one or less than 0.9 times 10 to the negative three, we see with higher grade or cellular tumors and are more likely to be malignant. So this is another study from our institution where we showed that DTI was very accurate in demonstrating the relationship of nerve fibers relative to the tumor, and in our study, it was about with a 96% sensitivity compared with intraoperative electrophysiological simulation, and this is very helpful for the surgeons when they're planning their surgeries so that they can safely excise the tumor. Here's another example of the useful role of ADC. So this is a patient with anal cancer who has bilateral femoral neuropathy, and the question presented to us was, is this perineural spread of tumor? So when we look at the MR neurography images, we see diffusely hyper-intense femoral nerves, which are also enlarged in size, not only in the pelvis, but also at the level of the groin, and then when we actually measured ADC values, the ADC values were elevated. They were above two times 10 to the negative three, and this allowed us to be more confident that this was radiation neuropathy and not tumor infiltration, and that proved to be the case on follow-up. Here is another example where the question presented to us was, is this, in this patient with metastatic breast cancer who had a left S1 radicular pain syndrome, whether this is tumor infiltrating the nerve, or is this post-radiation change? And when we look at the images here on the left panel, we see that there is mass-like enhancement and very low ADC values along the left S1 nerve root, consistent with tumor along the nerve. After this patient was radiated, we noticed that the ADC values are elevated, and this is post-treatment change. Here is an example of a tumor mimic. Remember that not everything that's mass-like along the nerve is tumor. This was a young female, 34-year-old, with a history of gluteal pain radiating to the right dorsal thigh and the lateral lower leg, and we see this mass-like soft tissue along the right sciatic nerve that enhances, and the interesting thing in the history was that this sciatic pain typically started a few days before menstruation, intensified progressively, and showed relief a week after menstruation was over. So this is actually catamenial sciatica. This is extrapelvic endometriosis, which is often associated with considerable diagnostic delay and morbidity. So remember, not all masses along the nerve are tumor. Keep this entity in mind. Remember, it's important to know the history. And in the last couple of examples I'll show you of tumor mimics, here is another one. This is a 61-year-old male with a left common perineal neuropathy, and we notice that there's this cystic mass with peripheral enhancement along the course of the left common perineal nerve. When we actually follow the MR neurography actual images down to the superior tip-fib joint, we notice that the cystic lesion extends into the joint, and this is not a nerve C tumor. This is a very characteristic example of an intraneural ganglion cyst. These cysts most commonly arise from the superior aspect of the tip-fib joint, and they can grow along an articular branch either into the common perineal or tibial nerve. And here is that, interoperatively, that markedly enlarged common perineal nerve in which that intraneural cyst had grown, and here is the cyst after resection. In my last example here, I want to show you another example of a large cyst within the tibial nerve. Notice, again, as we go down, we can follow this little cyst into the superior aspect of the tip-fib joint, and when we look at the coronal images, we can actually see this little articular branch through which the nerve is growing and then connecting along with the tibial nerve. And, oops, let's go back. Okay, and so this is just a depiction showing that these articular branches are present in the tip-fib joint, and these cysts can grow along these articular branches into the tibial or common perineal nerves. So, in summary, I just want to talk very briefly about some DTI challenges and troubleshooting. Remember that the peripheral nerves are small in size and complex, and the DTI images may be affected by motion, susceptibility, ghosting artifacts, and low SNR, and, therefore, some solutions that I've offered are scanning at 3T, using an auto-shimming, tighter echo-spacing, parallel imaging, higher bandwidth, and at least six directions. We use 12 to 28 directions for our peripheral nerve imaging. So my takeaway messages are that MR neurography and DTI is complementary to traditional clinical and electrophysiologic evaluations of patients with lower extremity peripheral neuropathies. DTI is ready for clinical use for assessing the severity of nerve injury, for monitoring early nerve regeneration, for preoperative planning, and differentiating benign from malignant entities. And the MR neurography and DTI findings may substantially influence patient management. Thank you very much for your attention. So let's start with the anatomy. So the brachial plexus is formed by the ventral rami of the nerves from C5 to T1. It is responsible for the motor and cutaneous innervation of the upper extremity, except for the motor innervation of the trapezius and levator scapular muscles, and the cutaneous innervation of the axilla, suprascapular, and scapular regions. It is important to remember and recognize the segments of the brachial plexus, which from medial to lateral are going to be the roots, the trunks, the divisions, the cords, and the branches. We teach our residents and our fellows back home, the pneumonic radiologists and technologists like to drink cold beer or cold beverages. So again, from medial to lateral, you're gonna have the roots, the trunks, the divisions, the cords, and the branches. So the roots of the brachial plexus extend from the neural foramina all the way to the edge of the scalene triangle. If you remember, the scalene triangle is formed by these big muscle, the anterior scalene muscle and the middle scalene muscle right there. So these would be the location of the roots of the brachial plexus in the axial plane. In the coronal plane, you can see the roots medial to the anterior scalene muscle extending from the neural foramina all the way to the edge of the scalene triangle. And we see them here surrounded by fat very nicely within that scalene triangle. Then at the edge of the scalene triangle, the roots are gonna merge to form the trunks. So we have C5 and C6 that merge to form the upper trunk. C7 continues alone as the middle trunk. This would be C7 and the middle trunk. And C8 and T1 are gonna merge to form the lower trunks. The trunks run for a very, very short segment just at the edge of the scalene triangle. So this would be the location in the coronal plane and this would be the location in the axial plane. Then each trunk is going to split into two to give an anterior and a posterior division for a total of six divisions. Three anterior, three posterior. They're gonna be located from the edge of the scalene triangle to the mid-clavicular level. So you always see this image right here. This is the mid-clavicular segment. So the divisions are gonna be located from the edge of the scalene triangle to that mid-clavicular level in the coronal plane and this is gonna be the appearance in the axial plane. On the sagittal images, you can see all these little dots represent the divisions of the brachial plexus just above the subclavian artery. So if we review a little bit of anatomy, this would be the first rib. This is the mid-segment of the clavicle. This is the subclavian vein. Again, subclavian artery and these little dots represent the divisions. When we're scrolling on PAX, as soon as these dots pass in between the clavicle and the first rib in the sagittal plane, we know that we have entered the cords. So the divisions become the cords of the brachial plexus. So the cords are going to extend from the mid-clavicular segment right here to the level of the anterior coracoid process. Those are two very important anatomical landmarks. We see here the cords in the coronal plane and this is gonna be the appearance and location of the cords in the axial plane. Then at the level of the anterior coracoid process, those cords turn into the proximal branches of the brachial plexus, i.e. the musculocutaneous, axillary, median, radial, and ulnar nerves. Again, the anterior coracoid process is your key anatomical landmark. And here we have this sagittal image at the level of the anterior coracoid process. And we can see the proximal branches of the brachial plexus surrounding the axillary artery. Very good. So let's move on to imaging. So MRI is definitely the method of choice to image the plexus due to its multiplier capabilities and exquisite soft tissue contrast. There are multiple ways of imaging the brachial plexus that will depend on the protocols that you have in your institution and the magnet that you're using, 3T versus 1.5T. I'm gonna show you the two protocols that we use in my institution. The first one is the protocol that we use in our 3T magnets and in our newer 1.5T magnets that have the appropriate software. I call it a short and sweet protocol. It only takes 20 minutes only for sequences, as you can see here, three coronal, one sagittal. My two goal sequence is the first one, a coronal 3DT to space. Why? Because it's isotropic, so it allows me to reformat in any plane. And it also allows me to compare with the contralateral side. Here's an example of how we acquire those images. And this is an example of the type of images that we get. And you're gonna see how here I'm starting to see all the segments of the brachial plexus on the right. One more click and I see all the segments of the right brachial plexus all the way from the neural foramina to the axillary region. I'm starting to see the contralateral one. One more click and here you go. All the segments of the left brachial plexus from the neural foramina all the way to the axillary region. So it's a very robust, short and sweet protocol. The other protocol that we use is a protocol that I developed while I was doing my fellowship at McGill University some years ago. It came out of frustration, this protocol. Why? Because we were not doing a very good job. Images were taking too long. We were scanning the patients for an hour. There was a lot of motion. No one wanted to report these images. So we tried to tweak the parameters and we didn't succeed in decreasing the scanning time. So we tried several things until we came up with a very simple solution which was to increase the number of slices that we use for the planning of the first image that we get which is the coronal localizer. So allow me to explain that. When the patient goes into the magnet, this is what the technologist is gonna do. They're gonna use these very thick slices in order to get the coronal localizer and the rest of the exam is gonna plan of this coronal localizer. But you can see here that the brachial plexus is not clearly seen. So the tech, what they're gonna do in order to make sure that the brachial plexus is included is they're gonna scan the patient from C4 to T4 including all these tissue and including also the lungs and the heart which will also increase the amount of artifacts, respiratory and cardiac artifacts. So what we did is we increased the number of slices in order to get that coronal localizer from seven to nine to 35 and by doing that, this is the localizer that we obtained. Now, you can see the brachial plexus, I can see the brachial plexus, the technologists can see the brachial plexus and if all of us can see the brachial plexus, then that means that we can plan the study according or along the course of the brachial plexus. So that's what we did and to our surprise, we obtained this beautiful image. This was a bonus. We were not expecting to find this. We can see this axial oblique T1 weighted sequence, all the segments of the brachial plexus all the way from the neural foramina to the axillary region. So of this axial oblique data set, we plan the coronal images and look at this, a single image and we can see all the segments of the brachial plexus, again, all the way from the neural foramina to the axillary region. In the same fashion, we plan the sagittal images and this is just to show you the differences that we use between the conventional technique and the modified technique in one and each one of the planes. At the end of the day, we were able to decrease the scanning time by about 14 minutes. But in addition to that, given the images that we obtained, we were able also to remove three of these sequences. So this is a protocol that nowadays we're able to do also in less than 20 minutes. So this is the way we obtained the images. And as you can see, these are not true axial images, but axial oblique images. And you can see how here I'm starting to see all the segments of the brachial plexus, one more, and I see all the segments of the brachial plexus all the way from the neural foramina to the axillary region. And then in the same fashion, we planned the coronal oblique sequences. And similarly, you can see how I start to see all the segments of the brachial plexus from the neural foramina. And then in the next image, all the segments all the way from the neural foramina to the axillary region in one to three images alone. So we published these back in 2013. If any of you is interested in getting this article, I'll be happy to send it to you. You can talk to me at the end of the session. So now allow me to illustrate the benefits of this modified technique with a nice example. This is a young patient who presented with left ulnar neuropathy. You can see the significant atrophy of the intrinsic muscles of the hand on the left side, in particular between the first and the second fingers. So he was seen by neurology. They diagnosed an ulnar neuropathy and neurologists thought that this was likely related to a mass in the proximal ulnar nerve. And he even said, I think this is going to be a schwannoma. So he requested an MRI of the brachial plexus. This is the MRI back in 2005, using the conventional technique. It was reported as normal. And I agree, we don't see any particular abnormality. Then year after year, this patient had a follow-up MRI of the left brachial plexus. All of them reported as normal. And then in 2011, the patient who happens to be the husband of one of our MR techs found out about our modified technique. So he said, can we try it? So we put him in the magnet and these are the images using the modified technique. And sure enough, we identified this fusiform enhancing mass lesion in the proximal branches of the left brachial plexus, suggestive of a schwannoma as the neurologist had suspected from the beginning. So what happened? I think that with the conventional technique, the lesion was not included in the field of view. That's why it was not detected or was at the edge. And there was probably volume averaging with the adjacent vessels. So moving on to some pathology, brachial plexopathies represent always a clinical challenge because the symptoms are often vague and non-specific. Trauma is by far the most common cause of brachial plexopathy, followed by neoplasm involvement. In terms of trauma, this is a slide that you've seen already. Vinil showed it in his lecture. These are the three main patterns of nerve injury. Neuropraxia, which we call also as stretch injury, where the axon is intact and you're gonna have injury of the surrounding endoneurium and perineurium. Then you have axon admesis, which is basically like a partial tear where there is discontinuity of the axon and of the surrounding endoneurium. And then the worst one, which is neuroadmesis, where there is complete nerve transection. So here you have a couple of examples of post-ganglionic injury. The patient on the right was involved in a motor vehicle accident, and you can see that there is increased signal intensity and swelling of the upper trunk and the proximal divisions of that brachial plexus in keeping with stretch injury of neuropraxia. The patient on the left was involved in a ski accident, and you can see that there is significant swelling and increased signal intensity of the C8 nerve root within the scalene triangle. And the lower trunk is also involved. And you see that there is an area that we don't see very well, and this is secondary to a partial tear or severe axon admesis. Examples of post-ganglionic injury. But we also have to assess the possibility of pre-ganglionic injury. And the best sequence for this is three-dimensional high-resolution T2-weighted sequences as we see in this example, because they allow us to see very nicely the ventral and the dorsal rootlets like in this particular normal individual. Different from this patient where a pre-ganglionic injury was suspected, and you can see the rootlets on the left, but you do not see them on the right side, secondary to avulsion injury. It is very common that these patients also present with tears in the dura, so we're going to find as well post-traumatic pseudomeningoceles, large ones like in this particular patient, and we can also see muscle edema, which could be seen as fast as within 24 to 48 hours after the injury. You have an example of a patient who sustained a motor vehicle accident. He was riding a motorcycle. These patients tend to fall with hyperextension of the arm, and you can see that there is significant thickening and increased signal intensity of all the segments of the brachial plexus on the right in keeping with stretch injury. This patient also has a couple of post-traumatic pseudomeningoceles. We do not see the nerves within those pseudomeningoceles in any plane, so there's also nerve root avulsion. So this is an example of pre- and post-ganglionic injury. So as you can see from these examples, imaging studies play a key role in making the difference between pre-ganglionic and post-ganglionic lesions a difference that is crucial for determining the management of these patients. As patients with pre-ganglionic injury will benefit from nerve transfers, while patients with post-ganglionic injuries will benefit from conservative treatment or from nerve grafts. Moving on to neoplastic processes, the most common neoplastic entities that are going to affect the brachial plexus are the nerve sheath tumors. You have a patient with a mass lesion in the divisions of the left brachial plexus. Look at the pattern of enhancement. It's like a target sign. This has been described in the context of neurofibroma, and actually that was the diagnosis in this patient, neurofibroma, but the truth is we're not very good with conventional imaging to tell apart neurofibromas from schwannomas. Here's a more dramatic case of a patient with NF1. You can see multiple lesions along the right brachial plexus, but also in the subcutaneous tissues of the neck and on both sides of the spine in a patient with NF1. Now, as Vinil showed in his previous lecture, there is an important role for tractography. This is from an article published in Neuroideology by the Geneva Group where they imaged several patients with benign and malignant lesions of the brachial plexus, and you can see that there is a mass here in the course of the left brachial plexus, and you can see how, on tractography, the fibers are going around the lesion. So with this, the authors of the paper suggested to the neurosurgeon this must be a schwannoma. The surgeon went in and was able to enucleate the lesion, and this was indeed a schwannoma. Different from this other patient with a mass lesion in the roots of the right brachial plexus where you can see that all the fibers are embedded within the lesion, and this ended up being a benign neurogenic tumor that, of course, was not resected. Here's another case of a patient who presented with left brachial plexopathy. There is a large mass lesion in the left upper log extending through the chest wall into the axillary region. You can see the central necrosis. There's also thickening and enhancement of the rest of that brachial plexus in a patient with a Pankow's tumor. And again, going back to that article, another patient with a Pankow's tumor, and you can see how the tumor went up through the apical fat pad into the scalene triangle, and you can see the destruction and disorganization of all the fibers in that particular location. So tractography is a very important tool as well. Another common reason to image the plexus, and this we get at least two per week in our institution, is to make the difference between radiation fibrosis and tumor recurrence in patients with history of cancer, history of radiation, who months or years later present with brachial plexopathy. So here I have an example of a patient with a history of breast CA and radiation, and you can see that there's significant smooth thickening of all the visualized segments of the right brachial plexus in a patient who proved to have radiation plexopathy. We followed this patient for many years and these never changed. So diffuse thickening different from focal speculated masses is more in keeping with radiation plexopathy. And this is a companion case, another patient with breast CA and radiation. You can see that there's significant thickening, clumping of all the visualized segments of the left brachial plexus. But the other interesting finding, this is a coronal stir, this is a coronal T2, note the signal of all the segments of the brachial plexus is very low. This is something that you would see in radiation plexopathy. If you think about it in tumor, what you're gonna see is increased signal intensity both in the stir and the T2 weighted sequences. And then what else about, what else can help us to make that difference? Diffusion. This is actually a very nice paper that comes from the group at UCSF published in AJNR in 2015. And they used diffusion imaging to try to tease apart the benign from the malignant lesions. And here we have two examples from that paper. This patient had a mass lesion here along the T1 nerve. You see that the nerve is thickened, increased signal intensity, there's significant restricted diffusion. And this is another patient on your right with a history of sarcoma that was resected, received radiation and this is the coronal ADC map. So they measured the ADC in these two patients. The patient on your left had a very low ADC of 0.95. They biopsied the lesion and this ended up being metastatic disease from breast cancer. While the patient on the right had a very high ADC of 2.59. And they just followed up with imaging this patient and these remain stable for many years. So this was considered to represent radiation fibrosis. So in their article, they say that an ADC equal or greater than 1.3 favors radiation fibrosis. And when the ADC is lower than 1.1, that would favor neoplastic involvement. And last but not least, don't forget to look at the spine because often the problem is there, such as in this patient with right C7 radiculopathy. You see that there is a disc extrusion extending into that foramen and compressing the exiting C7 nerve root. So take home messages. Remember the anatomy. Remember the mnemonic. Radiologists and technologists like to drink cold beer. Roots, trunks, divisions, cords, and branches. MRI is definitely the imaging method of choice to image the plexus, but it needs to be interpreted in the context of the clinical history, the exam, and the neuroconductive studies. I'm going to be talking about the osseous spine. And in particular, I'm going to be talking about osseous spinal metastatic disease. I think everybody here probably reads some spine, maybe, or you wouldn't be here. And everybody who reads spine probably sees some osseous spinal metastatic disease, the most common spinal tumor. Now you all know what this looks like, so I'm not going to tell you how to diagnose it. What I want to do instead is try to convince you to assess and report it in a certain way. And I think you're going to like it because that way of reporting will add value, it'll highlight our value, we already add a lot of value, and it's going to optimize patient care. And this applies not just to spine, but it's kind of a broader overview of a way of reporting that we can use in all of radiology. All right, so this was a case that came into our ER about six months ago, L.A. County ER, if I can say that. The fellow said, you know, the only history we got on this case was back pain. Should I call down? Should I find out anything else? I said, no. You know, there's only one thing this can be, we don't really need to know. This has to be osseous metastatic disease. Sagittal T1, you can see T1 infiltration of the L1 vertebral body, pathologic fracture, retropulsion in the canal. We probably have a few more lesions down here in three and four. Our sagittal fat-saturated T2, or STR, we can see the edema in this body, see this fracture a little bit better, see our other two lesions. Fat-saturated post-contrast image, surprisingly not enhancing as much as I would have thought, but you can see the epidural disease. And here on our axial post-contrast also, you can see the epidural disease and the paraspinal extension of the tumor. So I tried to convince my fellows to really be succinct in their reports and get the information that the surgeon needs to know, that the treating doctor needs to know. And my fellow did a good job. So he said suspected metastasis with pathologic fracture L1, retropulsion, moderate compression in the conus, 75% at loss, additional lesions L3 and L4. And that's good. I mean, that's what we saw there, right? But I threw it away. I put in my own. Tell me what you think about this one. My first point, unstable spine, SIN score 13. Those of you who don't know what the SIN score is, I'm going to explain that in this talk. Recommend prompt surgical consultation. Now it doesn't matter if you're the ER doc, if you're the intern on your first day, or if you're in the custodial service, you know reading that point number one, the importance of this case, the severity of this patient's condition and exactly what you need to do about it. Number two, lytic metastasis L1, pathologic fracture with retropulsion greater than 50% at height loss, involvement of the posterior elements. Each phrase in number two is part of this SIN score. And whether it's a neurosurgeon or an orthopedic surgeon, when they get this patient, when they get this consult very soon, they're going to know what each of those phrases is saying because this comes from their literature. And these are the things that they are going to assess on this patient. They don't even need to see the images. They know exactly what to do. Same thing with number three, high-grade conus compression, epidural spinal cord compression grade two. I'm going to be talking about that scale as well. Again, you know, our surgeons, they like to look at their own images, but this, they know immediately, looking at this impression, what's going on with this patient. So oncologic instability, very common for people with osseous metastatic disease to have catastrophic failure very quickly. When you have catastrophic failure, pain, paralysis, death, you could argue that in these patients with cancer, this unstable spine is actually more important than their cancer at this time. Pain, paralysis, death, they need to get that taken care of first, and then they can get their cancer treated or whatever's going to happen there. Differs from instability and trauma, the failure patterns are different. These patients have poor bone quality that impacts surgical considerations. They need longer, stronger constructs. Their healing will be delayed. Lots of things to think about. And we can predict impending failure based on the pattern of involvement with this SIN score. So like I said, pain, paralysis, death, cord compression is what matters here. This can happen two ways, pathologic fracture with retropulsion, for which we have the SIN score that I'm going to talk about or epidural extension of disease, for which we have the epidural spinal cord compression scale. So my goal in these patients facilitate rapid triage of this patient with an unstable spine to surgical consultation. The spinal instability neoplastic score comes from the orthopedic surgery literature. It has five radiologic components and one clinical component. So think about that for a second. Five parts of this scale, five parts in determining whether this person is stable or unstable, comes from our images and our reports. So like I said at the beginning, I'm making the assessment of stability. Think about that. That's not the surgeon. I'm making that assessment and the recommendation. This describes both current and potential instability, like I said, from the pattern of involvement. The higher the score, the more unstable the spine and the more urgent your surgical consultation needs to be. Now, the significance of the SIN score, this is relatively recent, published in 2010, but already this was immediately adopted by so many surgeons and oncologists, everybody in the spine oncology treatment team. For radiologists, it was a little bit slower. When I started talking about this maybe three or four years ago, pretty much nobody or very few people had heard of it. Now, I think it's more widely known when I went to the Spine Society meeting last February, there were at least five talks that addressed this. So it's becoming more common to see this and to use it in the radiologic community. It's well-validated, excellent inter-reader reliability across multiple specialties, radiation oncology, surgery, radiology. There are a number of papers out there. Incorporated into multiple treatment guidelines as surgery becomes more data-driven and algorithmic. This is the kind of information that goes into these algorithms. And it contributes to a standardized approach to reporting, consultation, and treatment. Now that, I can give you the references for those people taking pictures too. This is my version of that chart that you just saw on the last slide. I've just made it a little bit more simple. A lot of these things you're already talking about in your reports, I'm just organizing them a little bit differently. And also another difference. You're answering six questions right here. And your answers are not just whatever you want to say. It's a multiple choice test here. You have to choose one of these words. And depending on what you choose, it's assigned a point value. So location, is it in the junctional spine, cranial, cervical, cervical, thoracic? This makes a difference. I'm actually going to show you examples of all these in a minute. But if it's in the junctional spine, it gets three points. Mobile spine two, semi-rigid spine one, or rigid spine zero. What is the quality of the metastasis? Is it lytic, mixed, or blastic? Depending on what it is, you'll get a different number of points. Lyme, is there already Frank's subluxation, deformity, or is it preserved? Whatever your answer is, gets a different number of points and on down the list. Like I said, I'm going to show you examples of all these. Last one is pain. That's the clinical component. That's mechanical back pain. You may know if you talked to the ordering doctor, you looked in the EMR, or if you're in a tumor board. If not, you just leave that score out and add up the rest. 13 to 18 points. Unstable spine. Urgent surgical consultation is needed. Again, this is from our initial read of the MRI or the CT. Seven to 12 indeterminate, but still probably a good idea to get that surgical consultation as soon as you can. Zero to six is stable. So that means they don't have to worry right now about seeing the surgeon. They can go to their oncologist and get worked up for their cancer. All right, so let me give you examples of each of those. Location. Some areas of the spine are more vulnerable than others, and the consequences of failure are more devastating. So from the occiput to C2, cranio-cervical, cervical-thoracic, thoracolumbar, lumbosacral. The worst consequences if those fail. Here's an example. Renal cell carcinoma, axial CTA, sagittal CTA, completely destroying the posterior elements here, C2. And you can see why this is a problem. Completely destroying that soft tissue mass. What's holding the head on the spine? Not much. So when this fails, that's going to be bad for this patient. Now contrast that with this large melanoma metastasis within the sacral ala. Now this is invading the S1 foramen. It's probably causing some pain, but if this fails, it's not going to cause cord compression. So that one in the rigid spine gets zero. Mobile spine, the cervical spine, lumbar spine gets two, and semi-rigid thoracic spine, that gets only one because you have extra stability from the ribs. Alignment. If you already have sprains, subluxation, or translation, you're already unstable. You get the highest number of points in the whole thing, and that's four. A new kyphosis or scoliosis. And this can be secondary to a number of pathologic fractures, like this woman who had breast cancer. This had been happening for a while. She's got some compression of her thoracic cord. Or an acute event, one pathologic fracture, as in this case with breast cancer. And you can see this large soft tissue component coming out, compressing the thoracic cord. Got her in right away for surgery. Got corpectomy, cage placement, and regained neurologic function. Lesion and bone quality. So lytic mets have a higher number of points because they're more prone to pathologic fracture. Lack of mineralization, you can't bear the load as well. And they are predisposed to burst fracture. And it can happen quickly, as in this case, which was about a month. Here we have a coronal. We have these punched out lesions. Renal cell carcinoma, large soft tissue mass. I like this path picture, so I put that in as well. Here, blastic mets, typical. Prostate cancer, coronal, CT, multiple hyperdense lesions throughout the spine, the pelvis. Can be a little bit harder on MR. Sagittal T1, sagittal T2, hypo-intense. Sometimes they don't enhance. Even harder when it's a case that's been treated and now they have rising PSA again, so you don't know which one is pathologic. Here's an axial. These are not as likely to have the fracture causing the cord compression as epidural disease. And here's a nice path picture. Now, I wanted to put this one in. This is kind of a new slide. I think I got it last spring. So this was from an abdomen pelvis CT. And here's a sagittal. Here's a little hyperdense focus here. Here in L4, there's another one. L, or T12, there's another tiny one here. These were described in the report, or called bone islands, which maybe you could say that. And that's, I guess, what your first guess would be. Although I don't think this one looks quite right. I don't know what it is, but something doesn't look right there to me. What they didn't put together though was this person also had bladder cancer. So never call bone islands when somebody has bladder cancer. I would say that person needed to be evaluated further, but unfortunately they didn't come in for three more months. And so indeed, these were metastases. So here's the same focus. And then pretty much showing you here the imaging characteristics of a blastic met again. T1, or yes, T1 hypo-intense, T2 hypo-intense, STIR hypo-intense, non-enhancing. There's some enhancement around it, and STIR around it, but the lesion itself has no signal. And here we have it on the axial images as well. Now, it's funny, because when I've given a talk about this before, people have said, I don't know if it's lytic or blastic. And at first I thought, that's a silly question. Of course you're gonna know. But then I thought, yeah, that actually is a good question. So anybody in the audience tell me, is this lesion lytic or blastic? I defy anyone here to find a lesion. I would not have, I still don't see one. I don't know where it is. But they did have one. So on MR, T1 hypo-intense, STIR hyper-intense, enhancing. Obviously this is not lytic or blastic, but we remember maybe from medical school that nothing is really all blastic or all lytic. It's always a spectrum. This is probably somewhere in the middle. So I would have given this one mixed one. So for those people who say lytic or, what if you can't tell? I would say give them a one. All right, so posterior element involvement. We all know that the pedicles are very, very important. We always talk about that. In the thoracic spine, it's the costovertebral joints that give that stability. So bilateral, so important, that gets three. Unilateral, one. And if they're not involved, you get zero. The amount of vertebral body involvement and collapse. If you already have greater than 50% collapse, you're in trouble, that's three. Say this one, no collapse, but greater than 50% involved. You know that one's going soon, so that one gets one. And you can have more collapse in the thoracic spine than the lumbar spine, again, because you have the stabilization of the ribs kind of holding things up even when that vertebral body is failing. And pain, like I said, you may or may not know this, and this is not cancer pain. It's not the periosteal stretch that's worse at night and better during the day. This is mechanical pain, like degenerative pain. It hurts when you're walking around, feels better when you lay down. All right, so here's an example, putting this all together. Real cell carcinoma, sagittal T2, lesion in T9. So we go through with our chart here, and I'm gonna show you, soon you don't really have to have this chart to do this. Semi-rigid thoracic spine, so that only gets one point. It's a lytic lesion, so that gets two. Alignment is still okay despite that fracture, so that's zero. Collapse greater than 50%, so that gets three. Posterior elements were both involved, so that gets three. And we knew from talking to ordering doctor that the pain was severe, and it was mechanical, so that was three. So total score of 12, top of the indeterminate range. Still a good idea, obviously, to see a surgeon. You would probably do that anyway in this case. Now contrast that with this prostate cancer case we already saw. We'll, again, which one is active, we are not sure. Probably get a bone scan, but we'll talk about T9. Again, semi-rigid thoracic spine, you get one. It's blastic, so you get zero. Alignment is still okay, so zero. No collapse, but more than 50% involvement gets one. The posterior elements were not involved, so that got zero. And pain, we didn't know. So we added up what we did know, and that was two. Even with the maximum pain, it would have only been five. That person has a stable spine, so they can get worked up by their oncologist and not worry about the surgeon right now. All right, so I said you don't need to worry about using that chart. You don't need to print it out. You don't need to tape it to your wall. You don't need to Google it each time. This is a macro I created in PowerScribe. You can probably use whatever reporting system you have. This was easy. I just made this list. There's a pick list with every element there. So I would click on location. I choose from my choices here. These are my only choices, so you really barely even need to think. I knew it was, what, T9. So I just click on this. It fills it all in, the description with the points. Same with lesion quality. Give me three choices there. Alignment, collapse, all those. I added my score. Click on which one it is. Unstable, indeterminate, whichever fills it in. Now, even easier, these are all, this macro, in addition to a number of other ones, are available on the ASNR, American Society of Neuroradiology website. We have these common data element macros which have this kind of reporting where you, you know, it doesn't matter if you're in New York, you're in Los Angeles, you're in London, you're Hong Kong. If you use these, everybody is reporting these the exact same way. That patient can go all over the world, get treated. The doctors are gonna know because the surgeon and the oncologist are using this system, what needs to be done for this patient. You kind of see the utility of that. People don't like structured reporting. So what I like about this is you can make a normal prose report, dictate whatever you want. When you talk about that particular aspect of the patient's study, you put in this macro and you can fill it in and then you can finish your report, say whatever you want. At my old institution, they used to ask me to read all the spine oncology studies, not because I was the best reader or anything like that, but because I would do it this way and they loved it. They would just read that little box. That's what they needed for their algorithmic treatment of these patients. All right, so I mentioned accommodative elements. I'm not gonna go into that too much now, but on the ESNR website, you can find a description. There are things, there are macros for stroke, for MS, for head and neck, other ENT things, everything you can imagine, spine, brain, everything. So very useful if you wanna learn more about that. And you can imagine the implications. If we all use these terms, automated critical results, communication can be taken from our reports, data mining, machine learning, algorithm training, a bunch of things. So we'll learn more about that. All right, so let's move on to the other arm of our algorithm here. Core compression, pain, paralysis, death, epidural extension of disease, for which we have the epidural spinal cord compression scale. So this was another renal cell case, T1 hypo-intense, enhancing, ster hyper-intense. Here on our axial, we see a little bit of epidural disease. What would you call this? How would you describe this? Mild cord compression, moderate, severe, or just cord compression? How would you describe it? And how would the other people in your practice describe it? And would you all describe it the same way? And if you wouldn't, you might say, well, I'm just gonna let my surgeons, whoever, describe it. And in that case, how are we adding value? So I would argue that if we all describe this the same way, that would be a lot better. How about this case? Angioplastic angiosarcoma of the breast. Four bodies, a lot of epidural disease, unusual loculations of the CSF on sagittal T2. Here on the axial post-contrast, circumferential epidural extension of disease. How would you describe this one? Moderate, severe, I don't know. What would you say about this? The epidural spinal cord compression scale was also published in 2010. And this is a six-point scale for surgical and radiation oncology planning. And again, for uniform reporting and standardization. People don't like that word, but you can see how this could be helpful. So the reason, now, in this case, I'm gonna show you why it's useful for these patients. When you used to have, in old days, when you had osseous metastatic disease, then you would have large on-block resections of the bone. But now, with stereotactic radiosurgery or stereotactic body radiotherapy, instead of having this conventional external beam radiation where you had to keep your doses low because you didn't want to fry your spinal cord, these large boxes, now we've got all these highly conformal beams, which can all focus here on the bone, and they're in highly ablative doses. A lot of things that were radio-resistant are now sensitive to SRS. But if you have a case like this, here's a treatment plan, typical treatment plan. So sagittal and axial. This is a typical horseshoe-shaped plan. You could also have donuts, but basically, they can just contour and just get the bone and spare the spinal cord. But if you have a case like this with renal cell carcinoma on the sagittal T2, large mass coming out into the epidural space, compressing the cord. Here's our axial. A lot of epidural disease. If you put on your horseshoe-shaped or donut-shaped plan on this, you're going to leave that epidural disease, and then what good have you done? You didn't do any good at all. So what they need to do, if you can tell what my little graphic here is, body, spinous process, transverse processes, this is spinal cord. Surgeons will do a minimal surgery. They'll go and just take out the epidural disease very quickly, and then the person can go to radiation oncology and get their horseshoe-shaped or donut-shaped plan. So that's the utility of this system. So we're going to tell them whether it's low-grade or high-grade, because high-grade is going to need that separation surgery, that minimal surgery to take out the epidural disease before they go to radiation. The low-grade can go straight to radiation. Let me just show you some examples here. So grade zero, like this melanoma met here, this is osseous disease. It's confined to the bone. This one, this is breast cancer. You see a little bit of epidural disease here. It's coming into the frame a little bit. It's not really deforming the fecal sac, so that's grade 1A. And this is a renal cell. It is kind of deforming the fecal sac, but not contacting the cord, so that's 1B. This system used to be just 0, 1, 2, 3, but the radiation oncologist added this ABC because of their ability to just have such fine degrees of control. It's like a millimeter or two. Sometimes I think it's still hard to tell the difference between these three, but we try. And then you have the high-grade. This is the case I just showed you. So this one was actually grade two because there was a little bit of CSF left. All right, so like I said, if we think about the bigger picture for the patient, the way healthcare is going, the way surgical treatment is going, data-driven algorithmic management of these patients. And we need to think about what is our role in this too. We need to help out and play our part. So the NOMS criteria out of Sloan Kettering is a very good example of the way spine oncology patients are treated. This considers four aspects of disease status, neurologic, which is the amount of cord compression, oncologic, which is the histology or radiosensitivity, mechanical stability, which I said is the most important, and systemic status, what else is going on with the patient? So these four things that determine how the patient is treated, radiation surgery, what order, we contribute half of this data through our images and our reports. So the epidural spinal cord compression scale addresses cord compression. SIN score addresses mechanical stability. So you can see the importance of our images, our reports in this. We just need to do it the right way. So this is my last slide. So how can we improve quality and add value? I would argue that this type of reporting, using these macros, using these common data elements, using these scales of our surgeons, we want to work with the team to have the best outcomes for our patients. So in this particular case, we're making the stability assessment and the management recommendation. You can see what a huge difference we're making. If we get one of those patients who could have spine failure at any minute, rapidly to that surgeon before that happens, how much difference that's going to make for them. So use the spinal instability and plastic score, epidural spinal cord compression scale, and all of the common data elements, if you're interested. Again, go to the ASNR website for more efficient, effective communication. And so we can be better involved, better integrated into the larger healthcare team.
Video Summary
The video transcript provides an in-depth discussion of the application and potential of artificial intelligence (AI) in spine imaging, highlighting current advancements and the anticipated trajectory of AI technologies in this domain. Initially, the concept of AI is introduced as a branch of computer science that enables machines to emulate human intelligence, focusing on reasoning, learning, and self-improvement. The use of AI in radiology, particularly in spine imaging, is explored, emphasizing its capacities from order-weighted acquisition through segmentation to computer-assisted diagnosis.<br /><br />One of the highlighted applications is automated scan acquisition, which enhances efficiency by auto-aligning to the spine's angle, adjusting coverage, labeling spine levels, and performing tasks like curve planar reformatting. These innovations lead to fewer repositionings and reduced user interaction, making the process more efficient. Specific tools like Siemens' Dot Engines are cited as examples having significant positive impacts on reducing scan times.<br /><br />Moreover, the transcript elaborates on the use of AI in automatic scanning and reconstruction in CT, including case studies that show AI's potential to auto-estimate angles and align scans, providing higher quality readouts. The speaker underscores the importance of iterative reconstruction and compressed sensing in accelerating scanning processes while preserving image quality. Deep learning-based reconstruction, with its enhanced noise reduction capabilities and higher resolution maintenance, is also discussed for both CT and MR.<br /><br />The future direction AI might take in aid of quantifying spine assessments, segmentation tasks, and managing common data elements is also touched upon, with AI expected to handle preliminary measurements, leading to more structured reporting and data mining. The speaker closes with reflections on how AI tools are poised to redefine roles in spine imaging, aiming to optimize clinical decisions and improve patient outcomes through enhanced diagnostic confidence and efficiency.
Keywords
artificial intelligence
spine imaging
radiology
automated scan acquisition
computer-assisted diagnosis
Siemens Dot Engines
iterative reconstruction
deep learning
compressed sensing
segmentation tasks
structured reporting
diagnostic confidence
RSNA.org
|
RSNA EdCentral
|
CME Repository
|
CME Gateway
Copyright © 2025 Radiological Society of North America
Terms of Use
|
Privacy Policy
|
Cookie Policy
×
Please select your language
1
English