Shearlet Transformation of Multiparametric MRI for Prostate Cancer Diagnosis

Shearlet Transformation of Multiparametric MRI for Prostate Cancer Diagnosis

In this presentation I will describe a new way of image texture representation on magnetic resonance imaging. The objectivity is by this new technique, that we can increase the diagnostic accuracy of these images to diagnose prostate cancer.

The presentation outline is this. I will explain what multiparametric MRI is and then move on to the shearlet transformation. Then explain how the algorithm works. MRI depends on the water molecules, the hydrogen atoms in the water molecules, for getting an image and using an oscillating magnetic field.

The magnetic field could be 1.5 or 3-Tesla. At the moment you can use endorectal coil or body surface coil for diagnosis of prostate cancer. The three sequences that we are interested in are transverse relaxation time weighted, the T2W sequence. Then the longitudinal relaxation time T1 is not all that remarkable, but if you use a contrast agent that has a chelate, it becomes quite useful diagnosing cancer. The third one is the diffusion weighted imaging. We use the diffusion process of the water molecules for this. We can actually use this image to generate what you call the apparent diffusion coefficient map, which allows you to identify high-grade disease as much.

There are several limitations in the T2W sequence. The cancer is usually represented as a dark, low signal area. Unfortunately, there are other types of tissue that also can appear as dark regions, such as prostatitis, benign hyperplasia, etc. Also more difficult to detect cancer in the transition and central zones because of the bladder neck involvement.

The dynamic contrast enhanced imaging. Again, once you inject this contrast agent, you can see a rapid enhancement due to the angiogenesis. You can see higher wash-in, wash-out rates and large tumors can be easily detectable, 1C or larger. You might have problems detecting low-grade and low-volume tumors with this sequence. The apparent deficient coefficient can be used for identifying high-risk patients of Gleason 7 and above. Again, the accuracy may depend on the size and location of the tumors as well.

Here are the typical three sequences. I used a large computer model. This patient had prostatectomy surgery following his MRI. You can see it’s a fairly large lesion. You can see on the dynamic contrast enhance you can basically kind of trace out the shape of the tumor. The problem is because of the limitations the sensitivity and specificity is all over the map. The objectivity of this experiment of our study is to see whether we can come up with an image texture representation method that can give us high accuracy.

What we are using is something called shearlet transformation. This is a new type of transformation that is coming along with Fourier transform, Wavelet transform, Curvelet transform. What you do is all of these have their benefits and also have limitation. This is a new type of transformation.

I’ll give you an example. If you use the shearlet transformation, you can see if you have this red leaf, it can be represented by this blue and larger by a combination of reflection, rotation, scaling, and translation. If you want to do this kind of work, you wan to preserve the original information in the MRI image, the shearlets use parabolic scaling and shearling and translation. Then you can take that image into a different domain and then do the analysis there. Then translate the information back to the original image.

It is really suitable to identify structures like tumors. Here is an actual tumor, a wax replica of it, and then here’s a computer reconstruction of the thing. You can very clearly see these have in the MRI images these voxels and you can see these ridges on these structures, which the parabolic scaling law allows you to give is an accurate representation.

I’ll keep the equations out of the presentation. We use a set of equations. You can basically automate this whole thing in the computer. There’s not a whole lot of overhead involved, this post processing of the imaging. Essentially it takes the image to go through a low pass filter and then a series of high pass filters. Then you can use that information for classification.

Here is a picture of it. If you look at the original image, set of vegetables, here is the low pass image. Then you take the high frequency information and go through a set of series of high pass filters in order to get to the original image.

We go through four different steps preprocessing with multiparametric imaging, and that includes the segmentation and all those things. Then we cleared the histogram with shearlet coefficients. Then use the fissure extraction from this histogram and then do the classification.

What we have done is we have taken the original end rectal coil MRI. Those are the sizes of those images in pixels. Then our radiologist will separate the image using Photoshop software. Once you extract the prostate information, the revised sizes are given here. These are too small to run a shearlet transformation. What we do is we do an up sampling of these images using bicubic interpolation. Then we select regions of interest of 78 x 78 voxels for analysis purposes.

Here is a representation of it. Here is the original image. Then you separate benign and malignant regions on this image. Then you run through the shearlets. Then you clear this histogram. You can very clearly see the red ones indicate the benign regions and the blue ones indicate the malignant regions. You can do a feature extraction histogram and then do the classification.

To do the classification, there are a number of methods to do. We can use neural networks. We in this case use what is called Support Vector Machine. This is a binary classification, benign versus malignant. We repeated the experiment 50 times and then we tried to estimate the sensitivity, specificity, and the classification rate.

The results, we use, as I said, 3 Tesla, endorectal coil MR images. We used four patients. Here are the sizes of the tumors, if you are curious, of the four patients. The first patient had a large tumor. The rest the Gleason volumes and all these are given here. We took 10 benign regions of interest and then 10 malignant from each patient, a total of 80 images.

Then we selected the regions and then went through the classification. These are the results. At the patient level and also the overall results have been given. You can see the sensitivity and specificity is really good.

When we were selecting the regions, we also make sure that we will pick up not the entire tumor as we were going through the ages, the borders of the tumor. That’s the region that’s actually challenging for most of the radiologist. They can sometimes see the tumor, but they don’t know exactly where it ends.

The objectivity of this whole thing is to augment this software versus the radiology so that he can draw a little box on an image and then ask the algorithm is that tumor or not? Then if it is tumor, do the segmentation.

The sensitivity for apparent deficient coefficient is about 90%. The contrast enhance is almost like 100%. You can pick up the regions and the T2W is also very high. Then of course we compared this to what is currently available in a state-of-the-art methods. The Gabor filters, which are the wavelet transformation and then the histogram-oriented gradients, clearly these results show the current diagnostic methods, the shearlet methods, are far superior.

Of course this is just classifying the region, so this is like a rectangular area. Is this cancer now? The next question is obviously: can you identify? Do like a little segmentation and say where the tumors are. We have done that part also. Localization of the tumor boundary within the region of region of interest.

The first column shows the radiologist’s estimate of what the tumor border looks like. Then we have used two separate methods, graph cut and snake algorithm. You can see now these images are not actually the original images. These are all done in the shearlet domain. It was done in the shearlet domain and then transferred, mapped back to the original images. It can be done. We use the first level of shearlet coefficients for the segmentation. Then combine several who may come up with this composite image. Then it has a very small root mean square error.

In conclusion, it has improved accuracy, especially for detecting tumor edges and orientation. It’s a well-organized multi-scale structure so you can actually automate this whole thing in the computer. This actually out-performs the state-of-the-art methods that are out there at the moment. This technique has potential to significantly improve the diagnostic accuracy of multiparametric imaging.

We need to do additional studies to prove that this will work for different-size lesions, less than 1 cc, less than 0.5 cc, etc. Also depending on the anatomical location, peripheral transient zone, and central zone.

Acknowledgement. This is collaborative study. My colleagues at the University of Denver Canvas, Dr. Mohammad Mahoor and Jason Zhang and his graduate student, Hadi. Then of course Dr. La Rosa and Dr. Chang are the radiologists. We had a small grant to do this work from the National Science Foundation.

Flores-Tapia D, Venugopal N, Thomas G, et al. Real time MRI prostate segmentation based on wavelet multiscale products flow tracking. Conf Proc IEEE Eng Med Biol Soc. 2010;2010:5034-7.
http://www.ncbi.nlm.nih.gov/pubmed/21096687

Peng Y, Jiang Y, Yang C, et al. Quantitative analysis of multiparametric prostate MR images: differentiation between prostate cancer and normal tissue and correlation with Gleason score-a computer-aided diagnosis development study. Radiology. 2013 Jun;267(3):787-96.
http://www.ncbi.nlm.nih.gov/pubmed/23392430

Turkbey B, Choyke PL. Multiparametric MRI and prostate cancer diagnosis and risk stratification. Curr Opin Urol. 2012 Jul;22(4):310-5. http://www.ncbi.nlm.nih.gov/pubmed/22617060