Medical Image Analysis
Alzheimer's Disease
Prostate RadiotherapyIntroduction
Prostate cancer is the second most common cancer diagnosis in males (after non-melanocytic skin cancer) and the second most common cause of male cancer death (6). Radiotherapy provides a non-invasive therapy for prostate carcinoma which has been shown to provide an effective treatment for patients at a broad range of risk levels. However due to uncertainty in the delineation of the prostate from CT scans the radiation dose can be applied to areas of healthy tissue (such as the bladder and rectum).
The success of image guided radiotherapy depends on the accurate localisation of organs of interest. During prostate cancer radiotherapy treatment there is a need to minimize the dosage applied to the bladder and rectum (to reduce post-treatment complications). Recent advances in prostate radiotherapy have led to improvements in the amount of dose delivered to the target organ while reducing the amount to organs at risk, however side effects can still include inflammation of the anus, rectal bleeding and haematuria (1). Although Computed Tomography (CT) is usually used for prostate radiotherapy treatment planning, Magnetic Resonance Imaging (MRI) has a number of advantages including improved soft tissue contrast and better definition of tumour margins (2). In addition, prostate borders delineated on MR scans by radiation oncologists have been shown to have lower inter-observer variability and are smaller than on CT (3-5).
The aim of this project is to develop methods to use high contrast MRI scans directly for prostate cancer radiotherapy treatment planning. This would enable the prostate to be automatically delineated and then electron densities assigned from the MR scan. The main benefit of this work will be to improve treatment outcomes by reducing the dosage to normal tissues, and increasing dosage to the prostate.
Original axial CT image of the pelvis (left) and original axial MR image (right) from the same patient.
This project is a collaboration between the CSIRO Australian e-Health Research Centre , the Department of Radiation Oncology, Calvary Mater Newcastle Hospital the University of Newcastle and is partially funded by the Cancer Council NSW through Project Grant RG 07-06.
Aims
The project involves two main aims: firstly to investigate methods to automatically detect and segment the pelvic organs of interest (prostate, bladder, rectum, bones) on MR images and to secondly develop algorithms to automatically assign electron density information to MRI scans for radiotherapy dose calculations for treatment planning (currently this is not possible as MRI scans lack the electron density information required to calculate radiation dose).
Newcastle Mater Hospital collaboration
The study involves MR and CT scans from forty patients undergoing prostate cancer radiotherapy treatment planning at Calvary Mater Newcastle Hospital, Australia. Each patient received an axial proton density-weighted (PDw) whole body pelvic MRI acquired with a GE Signa 1.5 Tesla scanner using a phased array surface coil with an echo time (TE) of 44.2 ms, repetition time (TR) of ~5000ms, voxel size 1.56x1.56 mm, and slice thickness 3 mm. In addition planning CT volumes (helical 2.5 mm supine) were acquired for each patient. Manual segmentations of the prostate, bladder, bones and rectum were made on the MR and CT volumes by radiation oncology staff at the Newcastle Mater Hospital.
MR Image pre-processing
Interleaved images (or multipacket) acquisition are frequently used during full pelvic MR scans to reduce cross-talk and scanning time. One significant issue with the use of interleaved MR acquisition is that patient motion, including breathing motion, can result in non-linear 'staircase' imaging artifacts which are most visible on sagittal and coronal reconstructions. To allow these volumes to be processed more accurately with 3D image processing, registration or segmentation algorithms it is essential to correct this interleave artifact.
Previous methods (such as (7,8)) have assumed that only rigid transformations occur, and are unable to correct non-linear deformation. We have developed a novel non-rigid registration based method to reduce the effects of interleaving motion artifacts in single-plane MR scanning of the pelvic region that does not require k-space information. As most of the motion is planar, it is possible to correct each slice in 2D. Two main steps are involved: Firstly a new volume is created where every artifact affected axial slice is removed and replaced with an interpolated slice. Secondly, for each of these slices, 2D non-rigid registration is applied to register each original axial slice back to its matching interpolated slice.
Our results show visible improvements in artifacts particularly in sagittal and coronal image reconstructions, have resulted in improved smoothness of manual organ segmentations, and the use of the proposed method as a preprocessing step results in improvements in MRI to MRI (and multi-modal MRI to CT) affine and non-rigid registration [9].
An original sagittal reconstruction from a patient is shown in (a), (b) shows the reconstructed volume after axial slice interpolation, and (c) displays the result of interleave correction (non-rigid registration of original to interpolated axial slices).
Contour Processing
The manual treatment planning images and planning information (eg. Organ structures ) have been exported from two different treatment planning software platforms in RTOG and DICOM-RT. Custom software, utilizing the Insight Toolkit (ITK) and the Grassroots DICOM toolkit (GDCM2) has been written to parse these files and generate 3D binary segmentation volumes for each organ of interest from these formats.
Wireframe surface rendering of DICOM-RT structures (not to scale) from a CT (left) and MR (right) scan of the same patient.
Automatic prostate segmentation from MRI
We are using atlas methods to automatically segment the organs of interest. This involves constructing a labeled atlas from a training set of MR images (using a number of iterations of affine and non-rigid registration). The amount of overlap between labels in the atlas is used to generate probabilistic labels for each organ. To segment a new image, we warp the atlas (again using affine and non-rigid registration) to the image to obtain good correspondence between structurally equivalent regions on the two images and then transfer the labels from the atlas to the new image.
(left) Axial, coronal and sagittal views of the pelvic MRI average shape atlas, with (right) associated probabilistic maps of the bladder, bones, prostate and rectum.
A surface rendering from the atlas shown in previous figure, where the probabilistic atlas has been thresholded at 0.5 (which includes voxels where half or more of segmentations are overlapping).
The atlas was used in the segmentation scheme to constrain organs of interest from surrounding tissue. The probabilistic labels where thresholded to provide a general segmentation for each individual subject. The automatic segmentations were compared against expert manual segmentations using the Dice Similarity Coefficient (DSC) , where a DSC above 0.7 is usually considered a very good match.
(a) Shows the axial, coronal and sagittal views of a test MR (H003). (b) shows the expert manual segmentations for the prostate, rectum, bladder and bones for this test MR. The automatic segmentations from registration of the probabilistic atlas are shown on the right (c). The resulting DSC scores for this subject were Rectum=0.753; Bladder = 0.783; Prostate = 0.822; and Bone = 0.830.
The main cause of error in the automatic segmentation results are related to organ variation, which would be alleviated with the contribution of additional subjects to the atlas. Obesity and age may also be sources of error and BMI and age may be useful method to stratify subjects into multiple atlases.
References
1. Swallow T, Kirby R. Cancer of the prostate gland. Surgery 2008;26(5):213-217.
2. Prabhakar R, Julka PK, et al., Feasibility of using MRI alone for 3D radiation treatment planning in brain tumors. Jpn J Clin Oncol 2007;37(6):405-411.
3. Roach M, Faillace-Akazawa P, et al., Prostate volumes defined by Magnetic Resonance Imaging and computerized tomographic scans for three-dimensional conformal radiotherapy. Int J Radiat Oncol Biol Phys 1996;35(5):1011-1018.
4. Debois M, Oyen R, et al., The contribution of magnetic resonance imaging to the three-dimensional treatment planning of localized prostate cancer. Int J Radiat Oncol Biol Phys 1999;45(4):857-865.
5. Rasch C, Barillot I, et al. Definition of the prostate in CT and MRI: a multi-observer study. Int J Radiat Oncol Biol Phys 1999;43(1):57-66.
6. AIHW, 2007. Cancer in australia: an overview. Australian Institute of Health and Welfare (AIHW) & Australasian Association of Cancer Registries (AACR) Cancer series no. 37.
7. Gedamu A, Arnold D, et al., MRI inter-packet movement correction for images acquired with non-complementary data. Proc Int Symp Biomed Imaging (ISBI) 2008:pp. 416-419.
8. Rohlfing T, Rademacher MH, et al., Volume reconstruction by inverse interpolation: application to interleaved MR motion correction. Proc Med Image Comput Comput Assis Interv (MICCAI) 2008;11(Pt 1):798-806.
9. Dowling, J., Bourgeat, P., et al., Nonrigid correction of interleaving artefacts in pelvic MRI, J.P.W. Pluim and B.M. Dawant, Eds. 2009, SPIE MI. p. 72592P.
