| THE APPLICATION OF IMAGE PROCESSING
AND COMPARISON OF 3-D TRANSRECTAL ULTRASOUND
(TRUS) AND ENHANCED 3T MRI (E-MRI) IN STAGING
CLINICALLY CONFINED PROSTATE CANCER
Mr Santbir S Mehta, D C Barber, E van Beek, J Goepel, F C Hamdy Academic Unit of Urology & Medical Physics, University of Sheffield |
 |
Prostate cancer is the second leading cause of male cancer death both in the USA and in Europe. Radical prostatectomy is a recognised and well-established treatment option for localised disease. Accurate staging is critical to the management of patients with prostate cancer. While prostatectomy is appropriate for patients in whom the disease is organ confined, it is ineffective for patients where disease has penetrated the capsule. Undetected extra-capsular tumour extension may result in incomplete tumour excision and an increased risk of treatment failure.
Current methods used for staging lack sensitivity and under stage up to 50% of cancers. Two-dimensional TRUS is currently a standard procedure within the urology clinic, although accuracy can range from 50-70% due to various factors. MRI has also been investigated for staging and several reports show a potential for high accuracy
3D-TRUS has many advantages, it can be used in the urology clinic, it has no contraindications and is inexpensive compared to MRI. There is often a certain amount of disagreement between clinicians when interpreting images, this inter-observer variation remains an issue whether using TRUS or MRI, but any method which may introduce a more objective interpretation of the prostatic capsule can only serve to improve our general diagnostic abilities.
The critical sign for managing prostate cancer is whether disease has breached the prostatic capsule or boundary. Although the capsule cannot be visualised directly, in most subjects the adjacent layer of fat can. As prostate cancer invades locally, it breaks down this layer of fat and this can be seen on ultrasound as a defect. We have used a method called image registration to automatically delineate the prostate boundary with a view to detecting areas of disease penetration.
We examined a series of men with proven prostate cancer, thought to be organ confined. We were planning radical prostatectomy on all these men, so we would have the excised prostate to compare with after the operation. Most of this study was aimed at determining if we could develop novel image registration techniques for 3D TRUS and whether we could apply this to 3T MRI.
We wrote our own software to access the ultrasound and MRI images so that we could manipulate them with computer processing programmes. The first step was to prepare a reference prostate image. This was done by overlaying several images on top of each other, in effect producing a ‘perfect’ prostate to compare with the other scans. We then marked out the true boundary or capsule ourselves so now we have a perfect prostate with a definite boundary marked out.
The full report illustrates the technique in detail but in principle we were able to mould the reference image to take on the shape of any scan. In doing so the predetermined boundary is moulded too. This then gives us the automatic boundary on any given patient scan. We validated this and found it to be more accurate than any other similar published work. At this point we have been able to automatically identify the prostate boundary, using our own self designed software. The next step was to look closely at the boundary of fat surrounding the prostate to see if our programme could identify any defects, which may represent cancer penetrating outside the gland.
Normally the fat layer appears bright, however, if disease penetrates this layer it becomes dull or dark in appearance. In looking closer at the now automatically detected boundary, we determined exactly how bright the boundary was. This is called the ‘local image intensity’. We determined this for the whole surface and in many prostates found areas of defects, which we correlated to areas of real cancer spread.
We devised a way of first slicing and then reconstructing the removed prostate (with cancer within), the full report describes this in detail. We used the same technique of image registration to do this. Once we reconstructed the pictures of the sliced prostate we could define the capsule on that excised prostate image as well. Then in the same way, we registered the reconstructed excised real prostate to the scan. This registered the disease to the scan also.
When the true pathology disease (as marked by our pathologist) is transferred this way, we can determine its overall distribution and when looking at several prostates, we can also determine an average distribution of disease, which may help target prostate biopsies. We found that our method was able to reconstruct any form of disease distribution, no matter how complex. Furthermore, we were able to isolate the prostate cancer accurately from the surrounding normal prostate. Once this was done, we could measure the volumes of each. At present we use mathematical equations to predict cancer volumes, our technique would appear more accurate and versatile.
We were able to accurately transfer all the prostate cancer disease information to the individual scans. Despite this and with prior knowledge of areas of disease penetration we could not find any correlation between predicted areas of disease penetration and true pathologically proven regions.
On going work focused on applying the technology we have developed to magnetic resonance imaging (MRI). Initially results for 3D TRUS were promising. We would conclude however, that TRUS images lack sufficient resolution for effective computer analysis. Nevertheless, we have shown that we can produce novel techniques as outlined in our aims and these methods can potentially produce the desired outcome. Despite the difficulties when applied to TRUS, we felt that our technique was sufficiently robust to apply to MRI.
We undertook preliminary work on the then 1.5 Tesla (standard strength) equipment to determine to feasibility of image registration on MRI images. Despite some limitations the technique appeared to be sound, we therefore applied this to the higher field strength 3T MRI. Pilot work evaluating seven different MRI sequences has been applied to this cohort. All scans were viewed by two radiologists. In this small cohort, consensus was achieved in only 60% and interpretation was accurate in 54%. As yet our computer analysis data of the MRI data is only preliminary. We found that even in this small group, there is potential for considerable inter-observer variability, and we suspect that even high field strength 3T MRI lacks the ability to accurately stage clinically localized prostate cancer.
We have applied the same registration principles as used on 3D TRUS to the MRI data. We found that using these principles it is possible to register the cancer histology to the MR image also. We have begun work collecting a true 3D MRI image, albeit preliminary. Of all the MRI sequences examined here, the 3D image appears to be the most versatile for image registration even though it would not necessarily be the radiologists’ choice for raw image interpretation. We intend to further analyse the MRI subset and produce an ideal prostate protocol, to which image registration techniques can be better applied. Once perfected, we can continue recruitment to our full cohort.
Summary of final research report, 30 June 2007
Project G2003/13.