DR. REEVES: Thank you, Matthew, and thank you, Ed, for inviting me. I'm honored to be one of the first speakers here, being that I am in fact an electrical engineer, without a strong medical background.

I am from the School of Electrical and Computer Engineering Department at Cornell in Ithaca, New York. But I'm also a member of the ELCAP team, the technical part of the ELCAP team, which is with the Weill Medical College of Cornell University in New York City. And some of the methods I'm going to be talking about have been licensed by Cornell to GE in the computer analysis. The title here, "Characterization of Small Pulmonary Nodules and Measurements of Growth with Computer Analysis." That's the last two words of what I'm going to focus on, is how the technology, how the computer analysis can assist the radiologist, and in the treatment of this lung cancer.

Prior to this meeting we had some discussions as to what are the issues that we would like to focus on here. And so I'm just going to go through those quickly.
For the detection of small lung nodules, one of the primary things we want to know is the size. And the reason we want to know the size is so that we can estimate growth rate. And from growth rate, since cancer is growth, this is one of the most direct methods of indicating cancerous nodules.

Secondly, we are interested in the shape of the nodule. Determining growth rates takes time, with repeated measures. Perhaps just by observing the shape and the appearance of the nodule, we can determine whether or not it is malignant or benign.

Third, now that we are doing screening, now that we have these sensitive CT scanners, we are going to see some new findings, specific the ground glass opacities that are turning up in CT scans, new entities to be dealt with, and they are going to require some new investigation to exactly determine what they are, and how they progress, and then how to deal with them.

The fourth issue is establishing a CT nodule database, and how to handle that issue in the face of evolving technology. As you start building a database, technology changes. Now is your database out of date? This has been taken up by an NCI initiative, and a group of institutions are getting together to try and define a reference database here. I should stress that the work in computer image analysis requires such a database. This is what is needed, is a database of cases, plus known outcomes, and then the work is to go over that database time and time again, until we refine methods that very reliably can perform computer-aided diagnosis.

And finally, something that we haven't explored too much to date is how to evaluate treatments. Also in terms of chemoprevention, and in treatments of observing what happens to these nodules under various different effects.

Okay, so how can the computers assist in CT scan analysis? Let me very quickly review the concepts here. First, there is visualization. All CT scanning uses computers to interpret the data for the radiologist, and present it in some imaging form. That is the notion of reconstruction in the CT scanner. But there are many other opportunities here for ways to match the imaging data, the sense data to enable the human to perceive what is in turn a three dimensional density distribution that humans can actually perceive.

Second, detection. Large lung scans, we need to go over them with a fine toothed comb to see if there are any tiny nodules. And certainly this is an area where the computer is particularly adept at the laborious task of very finely searching through large data scans.

Diagnosis once these nodules have been located, and this will be the main issue that I will be focusing on, will help to determine whether or not that nodule is in fact a malignant cancer.

And as I also mentioned, evaluation. If one begins a treatment, does that nodule respond? And again, the sooner that one can give that information, the more useful is the result from imaging.

I split this up into two groups as an engineer, the first being visualization. How do we match the data to the human who is going to make the decision? And this is a qualitative thing. Some people will like some things, others will not, and beauty is in the eye of the beholder. This is not of great interest to me, but I will be showing you a lot of visualizations, because I think they get the point across. For me, the interest is in detection, diagnosis, evaluation, because all of these are quantified, all of these can be applied to extensive computer algorithms, and give you very rigorous outcomes in terms of probability of malignancy, and confidence of that probability.

And this is what I really call computer-aided diagnosis.

Next there is the impact of technology. As we have heard, the scanners are now getting thinner and thinner slices. Here is the impact of that. Five to ten years ago one would be 10 millimeter slices to capture the whole lung in one breath hold. And that image on the left is what you might expect from a small nodule attached to the pleural wall. On the right-hand side is what that would look like with a 1 millimeter slice. Now with today's technology, we can routinely do 1 millimeter slice thickness scanning of the whole lung, and within a few years we will be able to go finer still.

But the advantage, not just for the detection purpose, but of using these thinner slices is that we get actually the third dimensional structure of the nodule now, and now we can start using three dimensional imaging methods where we actually think of that nodule as a three dimensional entity, rather than a projection in a two dimensional entity, which is more the classical way of looking at two dimensional images, and I'll try and give an example of that.

But the motivation for computer-aided diagnosis from my perspective, is quantitation, accuracy, repeatability. This is where the computer is good, and this is where the human is not so good, perhaps. But this is not the motivation.

Access to an ever increasing knowledge database. If we can build these databases, then the computer can be aware of thousands or millions of previous nodules. And every time it makes a decision, based on this humongous database. That is today's computer technology. What we don't have is the database to refer to.

But the real kicker is the information explosion. With chest x-ray we are looking at one or two images to make a diagnosis. With the screening CT a few years ago we booked 30-80 images of 10 millimeter or 5 millimeter sections. Now we are going down, and currently we are looking at 160 to perhaps 320 images per scan. And that is a tremendous amount of work, looking for a needle in a haystack. The advantage of CT is we get very high contrast images. We can very clearly see these nodules. The disadvantage is we have to look through hundreds of images to find them. The next generation CT may have within a few years many hundreds of images per study.

And our computer-aided diagnosis, I'm going to talk about two dimensional and three dimensional methods. Basically, the conventional approach is to look at an image, make a decision, and that's the conventional two dimensional scheme of seeing what it is. The three dimensional scheme is look at the set of images, and integrate that data into a unified three dimensional object form. And our research is really focused on that, because we anticipate that with the CT technology, this is just going to become a routine practice.

The advantage of going to that third dimension, although it's a little difficult to visualize is that now you are using all the information of that nodule, not just a sample slice through the middle of it. And when there may be uneven growth in the nodule, then you can capture that by looking at all three dimensions simultaneously. And the other third advantage is we can use geometric filtering, which I will get to.

So here is the first look now at a CT scan. How we would take the different slices from a CT scan of a nodule, and in three dimensions, represent -- that is a light shaded representation, although that's still a two dimensional shape. But since it's now a three dimensional nodule, I can rotate that around into some particular orientation that you hadn't seen before. And also since it is isotropic, I can sort of flip it around and view that nodule from different perspectives.

Further, we can enhance and use volume rendering, which is one of the more recent techniques. Here is a solid nodule, quite visible in a two dimensional sense. If we do volume rendering, we could present it this way. And since we can now do geometric differentiation, you can see that that's a nodule, I hope, and you can see that there is a structure around it. These structures are defined geometrically. There is no density difference visible. It's all just white material. We can use that to actually identify the nodule. This is all done completely automatically to say this is a shape that doesn't belong in here.

And we can expand on that, and again, once we have that, we can see this from all different perspectives, and see how it's attached, the vessels, and compare from one time to another time, exactly what is going on with this type of nodule. So the visualization can be very helpful in just the standard context before we go into our computer-aided diagnosis.

With computer-aided diagnosis, the first thing we need to do is extract the nodule from the CT images as I showed in that last image. This is really the hard problem, finding out which voxels belong to the nodules, or which parts of which voxels belong to the nodule. Once we have accomplished that task accurately, then the two things we look at, growth or shape, is fairly straightforward. Once we have accurately got which voxels belong to the nodule, or which partial voxels do, then just counting those voxels gives us the volume.

Once we have got an accurate representation, then measuring shape parameters is also greatly simplified.

And here are some of the problems we encounter. Here is a nodule. Here it is one slice. This is a two dimensional thresholding technique. The issue here is whether this part here is a vessel, or is it part of the nodule? And from a two dimensional image, it is very hard to tell in these cases. Obviously, these are vessels. But is this also a vessel?

If we now look at that from a three dimensional perspective, it becomes fairly apparent that this is a cylindrical structure attached to the nodule. And now you can use simple reasoning to figure out which is which, and the computer can too. And this is what we call geometric filtering, and the automatic step is to go from this, to say -- actually, this is the nodule, and this is a connecting vessel. This can all be done with simple mathematical algorithms.

Another example of a problem is a nodule that is attached to the pleural surface. And again, automatic methods can go in and define that it is attached, and define the banding surface, and simply extract that nodule. Once it is extracted, now we can measure its volume, and we can also measure its shape.

Now the first issue I discussed is nodule size measurement, because we want to get growth rate. And the issues we have related to that is "when do we have to repeat this measurement?" "When do we tell the patient to come back?" This is critical, as any delay makes the patient very anxious. To address these, we need to know what growth rate precision we need, what growth rate represents malignancy, and how accurately can we actually do the measurements?

To do the growth rate measurement once we address those issues, this is the procedure. We segment it, as was just mentioned before. We have a two dimensional method where we just look at a single slice, figure out the area of the nodule of that slice, and a three dimensional method where we look at the whole volume.

Here is an example of the two dimensional method. We see in time one and time two there is a separation here of about a month. And the area which is given here is essentially identical. The two dimensional method says there is no appreciable growth in this nodule.

We went to the three dimensional method and determined that there is dramatic growth. Here is the rendering in three dimension. It does not appear that much is going on. This is in a 10 millimeter boundary, so this is about a 7 or 8 millimeter nodule.

When we rotated this nodule and viewed it from a coronal viewpoint, we noticed that there was a big change over that month from time one to time two. This was happening in the actual dimension. And this wasn't apparently visible when we just looked at the two dimensional slices. So we learned from this that nodules do not necessarily grow uniformly in all directions, and this was in fact an aggressive cancer.

Some of our preliminary results are shown here. Growth rate, instead of doing doubling times, we did percent increase in volume, compared it annually, as many people understand this from their life experiences. By doing this we found a dramatic difference in our small database, 5 malignant cases and 11 non-malignant cases, that the growth rate -- and think in doubling terms that a growth doubling time of 365 days is 100% interest. On that scale, the growth rate for malignant cancers ranged from 17% to 141% for our 5 cancers. For our 11 non-malignant lesions, the growth rate was maxxed at 2.2% per year. So there was a very different quality in the growth rates that can be determined from these measurements.

Our initial results show that in the growth is a very good predictor of malignancy. The scan can be taken after a relatively short amount of time from the first scan, perhaps in the order of a month or less. And three dimensional methods can catch some problems for two dimensional methods.

So if you come back to this issue of how long do we have to wait, what kinds of things affect the accuracy of our measurements, the CT scanner repeat accuracy, the CT scanner calibration (although this isn't really an issue), nodule size -- the smaller the nodule, the less voxels there are to measure and to count, and patient motion. Which hopefully can be taken care of by the operating technician.

Here is the graph of the size increase we would expect for different doubling times. And I guess many people expect that the difference between say a malignant and a non-malignant nodule may be in range of 400-600 days doubling time, with 400 tending to be perhaps malignant, and 600 and above perhaps not. But whatever your threshold, somewhere in this region here is where we would have some kind of decision point, at 100 or 30 days, it's obviously a very aggressive cancer.

If we take that graph and put it into a tabular form here, looking at this critical time of 400 days to 600 days, what precision do we need to measure? If we were to wait say a year, then we would expect to see an 88% increase in the nodule, compared to 52%, which is quite dramatic. If we wait say one month, then the difference in increase is somewhere between 3-5%. So we would like to have a precision in our measurement within a couple or two percentage points to distinguish those two cases. And if we try to do it say in two weeks, then this critical case would be in the order of 1 percentage point of accuracy of our measurements, if we could get that precise. An aggressive cancer, however, would be dramatically 10% larger within that two week time span.

Here are just a couple of quick examples. This is an aggressive cancer scanned 33 days apart. You can hopefully see the growth.

And here is one non-malignant nodule scanned 2,400 days apart, and we do not see any appreciable growth.

Let me try and cover on this one of the issues in accuracy of measurement is how many of these voxels are on the periphery of the nodule, and therefore partial voxels? As you see, you get down in size, and this is with the typical of today's scanner parameters, 1 millimeter slice thickness, a 4 millimeter nodule has about a third of its voxels on the surface, 2 millimeter has all its voxels on the surface. So we believe we can be fairly good in this range, still as to perform accurate volume and shape characterization.

If we go to the next generation scanner, which is down -- I guess anyone else will point out that this 2-4 millimeters is an eight times change in volume. And from 2 millimeter to a centimeter is about 100 times change in size.

If we go now to a next generation scanner, with perhaps 2.2 millimeter slice thickness, then we should be able to move down into the 2 millimeter size range.

The bottom line here is that actually our results do show that we can possibly be within 1 or 2 percentage points.

Let me mention shape characterization. How can we predict this? We have done several measures. For this we look at things' geometric form, density distribution within the nodule, and surface curvature information. And again, our size isn't useful here, because all our nodules are less than a centimeter, and we only have one scan.

Our outcome from that is out of our small database, we have managed to get a very high acceptance rate. Malignancies in all but one of the 9 out of 21 cases were identified by this method.

New radiological findings, ground glass opacities, what do we do with them? What are ground glass opacities? Focal areas of increased attenuation that do not observe the underlying structures. This is a very important area right now. And we defined two types, the non-solid and the part solid. And let me move on to just show you what I mean by that.

This is what we call the non-solid. You can see vessels inside, but they are not part of the nodule, and that becomes obvious when we do a vessel analysis. Then we can just localize the region that is the non-solid ground glass opacity.

A more complex case here also has some solid components in it. And we have found that there is some indication here that monitoring these solid components, perhaps they have different growth rates than the whole region and the opacity itself.

And if we just look at this by thresholding and obtain -- this is the GGO component. These are the solid parts of it. I want to sort of wrap here to show this is what the radiologist's viewpoint of this nodule may be, looking through the scans, and go half way through the sequence again. I'll do that one more time.

And on the right side, what we do when we try to go to a computer analysis of this, and right now we are looking at the solid component, which is the normal way of thresholding it, and I take that round now and view it from different directions, we can now add the non-solid component as a translucent region, and label the vessels from their geometric structure.

And since the vessel is not of primary interest, what we really want to see is the solid component, and its relationship to the non-solid component.

And going beyond that, we can just go to that solid component that does not contain vessels, and bring that back to the system again.

So we have come to the final part here. A reference database is essential to pursue this kind of research. And this again, ongoing work, and I'll come back to those issues that I mentioned initially.

What we need to do in the future with these new findings? We have the tools. How do we deal with the development of technology? I don't think this is a big issue, because I think the methods that we are developing will track technology, and it will just mean a size difference. Whereas, now we can handle 4 millimeter nodules, as new technology evolves, that will shrink down to 2 millimeter nodules, and maybe beyond that.

Thank you very much.

TOP