Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Biomedical Engineering

Supervisor

Dr. Aaron Fenster

2nd Supervisor

Dr. Aaron Ward

Joint Supervisor

Abstract

Multi-parametric magnetic resonance imaging (mpMRI) of localized prostate cancer has the potential to support detection, staging and localization of tumors, as well as selection, delivery and monitoring of treatments. Delineating prostate cancer tumors on imaging could potentially further support the clinical workflow by enabling precise monitoring of tumor burden in active-surveillance patients, optimized targeting of image-guided biopsies, and targeted delivery of treatments to decrease morbidity and improve outcomes. Evaluating the performance of mpMRI for prostate cancer imaging and delineation ideally includes comparison to an accurately registered reference standard, such as prostatectomy histology, for the locations of tumor boundaries on mpMRI. There are key gaps in knowledge regarding how to accurately register histological reference standards to imaging, and consequently further gaps in knowledge regarding the suitability of mpMRI for tasks, such as tumor delineation, that require such reference standards for evaluation.

To obtain an understanding of the magnitude of the mpMRI-histology registration problem, we quantified the position, orientation and deformation of whole-mount histology sections relative to the formalin-fixed tissue slices from which they were cut. We found that (1) modeling isotropic scaling accounted for the majority of the deformation with a further small but statistically significant improvement from modeling affine transformation, and (2) due to the depth (mean±standard deviation (SD) 1.1±0.4 mm) and orientation (mean±SD 1.5±0.9°) of the sectioning, the assumption that histology sections are cut from the front faces of tissue slices, common in previous approaches, introduced a mean error of 0.7 mm.

To determine the potential consequences of seemingly small registration errors such as described above, we investigated the impact of registration accuracy on the statistical power of imaging validation studies using a co-registered spatial reference standard (e.g. histology images) by deriving novel statistical power formulae that incorporate registration error. We illustrated, through a case study modeled on a prostate cancer imaging trial at our centre, that submillimeter differences in registration error can have a substantial impact on the required sample sizes (and therefore also the study cost) for studies aiming to detect mpMRI signal differences due to 0.5 – 2.0 cm3 prostate tumors.

With the aim of achieving highly accurate mpMRI-histology registrations without disrupting the clinical pathology workflow, we developed a three-stage method for accurately registering 2D whole-mount histology images to pre-prostatectomy mpMRI that allowed flexible placement of cuts during slicing for pathology and avoided the assumption that histology sections are cut from the front faces of tissue slices. The method comprised a 3D reconstruction of histology images, followed by 3D–3D ex vivoin vivo and in vivo–in vivo image transformations. The 3D reconstruction method minimized fiducial registration error between cross-sections of non-disruptive histology- and ex-vivo-MRI-visible strand-shaped fiducials to reconstruct histology images into the coordinate system of an ex vivo MR image. We quantified the mean±standard deviation target registration error of the reconstruction to be 0.7±0.4 mm, based on the post-reconstruction misalignment of intrinsic landmark pairs. We also compared our fiducial-based reconstruction to an alternative reconstruction based on mutual-information-based registration, an established method for multi-modality registration. We found that the mean target registration error for the fiducial-based method (0.7 mm) was lower than that for the mutual-information-based method (1.2 mm), and that the mutual-information-based method was less robust to initialization error due to multiple sources of error, including the optimizer and the mutual information similarity metric. The second stage of the histology–mpMRI registration used interactively defined 3D–3D deformable thin-plate-spline transformations to align ex vivo to in vivo MR images to compensate for deformation due to endorectal MR coil positioning, surgical resection and formalin fixation. The third stage used interactively defined 3D–3D rigid or thin-plate-spline transformations to co-register in vivo mpMRI images to compensate for patient motion and image distortion. The combined mean registration error of the histology–mpMRI registration was quantified to be 2 mm using manually identified intrinsic landmark pairs.

Our data set, comprising mpMRI, target volumes contoured by four observers and co-registered contoured and graded histology images, was used to quantify the positive predictive values and variability of observer scoring of lesions following the Prostate Imaging Reporting and Data System (PI-RADS) guidelines, the variability of target volume contouring, and appropriate expansion margins from target volumes to achieve coverage of histologically defined cancer. The analysis of lesion scoring showed that a PI-RADS overall cancer likelihood of 5, denoting “highly likely cancer”, had a positive predictive value of 85% for Gleason 7 cancer (and 93% for lesions with volumes >0.5 cm3 measured on mpMRI) and that PI-RADS scores were positively correlated with histological grade (ρ=0.6). However, the analysis also showed interobserver differences in PI-RADS score of 0.6 to 1.2 (on a 5-point scale) and an agreement kappa value of only 0.30. The analysis of target volume contouring showed that target volume contours with suitable margins can achieve near-complete histological coverage for detected lesions, despite the presence of high interobserver spatial variability in target volumes.

Prostate cancer imaging and delineation have the potential to support multiple stages in the management of localized prostate cancer. Targeted biopsy procedures with optimized targeting based on tumor delineation may help distinguish patients who need treatment from those who need active surveillance. Ongoing monitoring of tumor burden based on delineation in patients undergoing active surveillance may help identify those who need to progress to therapy early while the cancer is still curable. Preferentially targeting therapies at delineated target volumes may lower the morbidity associated with aggressive cancer treatment and improve outcomes in low-intermediate-risk patients. Measurements of the accuracy and variability of lesion scoring and target volume contouring on mpMRI will clarify its value in supporting these roles.

Share

COinS