Dosimetric evaluation of MRI-to-ultrasound automated image registration algorithms for prostate brachytherapy

PurposeIdentifying dominant intraprostatic lesions (DILs) on transrectal ultrasound (TRUS) images during prostate high-dose-rate brachytherapy treatment planning remains a significant challenge. Multiparametric MRI (mpMRI) is the tool of choice for DIL identification; however, the geometry of the prostate on mpMRI and on the TRUS may differ significantly, requiring image registration. This study assesses the dosimetric impact attributed to differences in DIL contours generated using commonly available MRI to TRUS automated registration: rigid, semi-rigid, and deformable image registration, respectively.Methods and MaterialsTen patients, each with mpMRI and TRUS data sets, were included in this study. Five radiation oncologists with expertise in TRUS-based high-dose-rate brachytherapy were asked cognitively to transfer the DIL from the mpMRI images of each patient to the TRUS image. The contours were analyzed for concordance using simultaneous truth and performance level estimation (STAPLE) algorithm. The impact of DIL contour differences due to registration variability was evaluated by comparing the STAPLE-DIL dosimetry from the reference (STAPLE) plan with that from the evaluation plans (manual and automated registration) for each patient. The dosimetric impact of the automatic registration approach was also validated using a margin expansion that normalizes the volume of the autoregistered DILs to the volumes of the STAPLE-DILs. Dose metrics including D₉₀, D_mean, V₁₅₀, and V₂₀₀ to the prostate and DIL were reported. For urethra and rectum, D₁₀ and V₈₀ were reported.ResultsSignificant differences in DIL coverage between reference and evaluation plans were found regardless of the algorithm methodology. No statistical difference was reported in STAPLE-DIL dosimetry when manual registration was used. A margin of 1.5 ± 0.8 mm, 1.1 ± 0.8 mm, and 2.5 ± 1.6 mm was required to be added for rigid, semi-rigid, and deformable registration, respectively, to mitigate the difference in STAPLE-DIL coverage between the evaluation and reference plans.ConclusionThe dosimetric impact of integrating an MRI-delineated DIL into a TRUS-based brachytherapy workflow has been validated in this study. The results show that rigid, semi-rigid, and deformable registration algorithms lead to a significant undercoverage of the DIL D₉₀ and D_mean. A margin of at least 1.5 ± 0.8 mm, 1.1 ± 0.8 mm, and 2.5 ± 1.6 mm is required to be added to the rigid, semi-rigid, and deformable DIL registration to be suitable for DIL-boosting during prostate brachytherapy.     

Evaluation of the effect of user-guided deformable image registration of thoracic images on registration accuracy among users

User-guided deformable image registration (DIR) has allowed users to actively participate in the DIR process and is expected to improve DIR accuracy. The purpose of this study was to evaluate the time required for and effect of user-guided DIR on registration accuracy for thoracic images among users. In this study, 4-dimensional computed tomographic images of 10 thoracic cancer patients were used. The dataset for these patients was provided by DIR-Lab (www.dir-lab.com) and included a coordinate list of anatomical landmarks (300 bronchial bifurcations). Four medical physicists from different institutions performed DIR between peak-inhale and peak-exhale images with/without the user-guided DIR tool, Reg Refine, implemented in MIM Maestro (MIM software, Cleveland, OH). DIR accuracy was quantified by using target registration errors (TREs) for 300 anatomical landmarks in each patient. The average TREs with user-guided DIR in the 10 images by the 4 medical physicists were 1.48, 1.80, 3.46, and 3.55 mm, respectively, whereas the TREs without user-guided DIR were 3.28, 3.45, 3.56, and 3.28 mm, respectively. The average times taken by the 4 physicists to use the user-guided DIR were 10.0, 6.7, 7.1, and 8.0 min, respectively. This study demonstrated that user-guided DIR can improve DIR accuracy and requires only a moderate amount of time (<10 min). However, 2 of the 4 users did not show much improvement in DIR accuracy, which indicated the necessity of training prior to use of user-guided DIR.     

Quantification of the accuracy limits of image registration using peak signal-to-noise ratio

A new method was developed for quantifying the accuracy limits of image registration devices and the distortion of anatomical structures in verification images without image registration. A correlation was found between peak signal-to-noise ratio (PSNR) and the amount of parallel movement (1–10 mm at 1-mm intervals) of a rectangular parallelepiped phantom [correlation coefficient (CC) ?0.91, contribution ratio (CR) 0.83]. Rotating the phantom from 1° to 10° at 1° intervals produced a similar correlation with PSNR (CC ?0.91, CR 0.83). To allow for manual registration, the grid pattern of the Mylar top plate was extracted from 455 pelvic portal images of 21 patients using a band-pass filtering technique. This revealed a different correlation between the original data (CC ?0.62, CR 0.38) and averaged data (CC ?0.96, CR 0.92), but this is considered to have been caused by structural distortion and manual matching errors. Thus, PSNR can be used to evaluate the accuracy limits of image registration and provide a judgment index that can be used in re-planning or re-setup in adaptive radiotherapy.     

Simulated required accuracy of image registration tools for targeting high-grade cancer components with prostate biopsies

Objectives

To estimate the required spatial alignment accuracy for correctly grading 95 % of peripheral zone (PZ) prostate cancers using a system for multiparametric magnetic resonance (MR)-guided ultrasound (US) biopsies.

Methods

PZ prostate tumours were retrospectively annotated on multiparametric MR series using prostatectomy specimens as reference standard. Tumours were grouped based on homogeneous and heterogeneous apparent diffusion coefficient (ADC) values using an automated ADC texture analysis method. The proportion of heterogeneous tumours containing a distinct, high Gleason grade tumour focus yielding low ADC values was determined. Both overall tumour and high-grade focal volumes were calculated. All high-grade target volumes were then used in a simulated US biopsy system with adjustable accuracy to determine the hit rate.

Results

An ADC-determined high-grade tumour focus was found in 63 % of the PZ prostate tumours. The focal volumes were significantly smaller than the total tumour volumes (median volume of 0.3 ml and 1.1 ml respectively). To correctly grade 95 % of the aggressive tumour components the target registration error (TRE) should be smaller than 1.9 mm.

Conclusions

To enable finding the high Gleason grade component in 95 % of PZ prostate tumours with MR-guided US biopsies, a technical registration accuracy of 1.9 mm is required.

Key Points

? MRI can identify foci of prostatic cancer with reduced apparent diffusion coefficients ? Sixty-three per cent of prostatic peripheral zone tumours contain high-grade tumour low ADC foci ? The median volume of such foci is 0.3 ml ? Biopsy targets are significantly smaller than whole tumour volumes ? Simulated registration accuracy is 1.9 mm for correctly grading 95 % of tumours     

Factors associated with deformation accuracy and modes of failure for MRI-optimized cervical brachytherapy using deformable image registration

PurposeTo identify factors associated with MRI-to-CT image deformation accuracy and modes of failure for MRI-optimized intracavitary high-dose-rate treatment of locally advanced cervical cancer.Methods and MaterialsTwenty-six patients with locally advanced cervical cancer had preimplantation MRI registered and deformed to postimplantation CT images using anatomically constrained and biomechanical model–based deformable image registration (DIR) algorithms. Cervix (primary) and cervix plus 10-mm margin (secondary) were used as controlling regions of interest for deformation. High-risk clinical target volume defined on pre-MRI was propagated to CT and evaluated for clinical utility in optimizing target volumes using scores 0 (low performing) to 4 (high performing). Quantitative evaluation of deformation performance included Dice index, distance to agreement, center of mass (COM) differences, cervical/uterus volume, and geometric change in organ position for MR-projected structures. Statistical analysis was performed to identify predictors of clinical utility and modes of failure.ResultsAnatomically constrained and biomechanical model–based deformable image registration algorithms achieved clinical utility >3 in 65% and 81% of patients, respectively. This improved to 81% and 85%, respectively, if cervix plus margin was used to drive deformations. Total COM displacement (cervix plus uterus) had the highest sensitivity in predicting low from high clinical utility in optimizing target volumes. Deformation failure (low clinical utility) resulted from high COM displacement, high cervical volume change, and retroverted uterine anatomy.ConclusionsMRI-to-CT deformable image registration using a cervix-controlling region of interest can aid clinical target delineation in cervical brachytherapy and potentially improve brachytherapy implant quality and clinical workflow. Deformation failures warrant further study and prospective deformation validation.     

Improving accuracy and efficiency of mutual information for multi-modal retinal image registration using adaptive probability density estimation

Mutual information (MI) is a popular similarity measure for performing image registration between different modalities. MI makes a statistical comparison between two images by computing the entropy from the probability distribution of the data. Therefore, to obtain an accurate registration it is important to have an accurate estimation of the true underlying probability distribution. Within the statistics literature, many methods have been proposed for finding the ‘optimal’ probability density, with the aim of improving the estimation by means of optimal histogram bin size selection. This provokes the common question of how many bins should actually be used when constructing a histogram. There is no definitive answer to this. This question itself has received little attention in the MI literature, and yet this issue is critical to the effectiveness of the algorithm. The purpose of this paper is to highlight this fundamental element of the MI algorithm. We present a comprehensive study that introduces methods from statistics literature and incorporates these for image registration. We demonstrate this work for registration of multi-modal retinal images: colour fundus photographs and scanning laser ophthalmoscope images. The registration of these modalities offers significant enhancement to early glaucoma detection, however traditional registration techniques fail to perform sufficiently well. We find that adaptive probability density estimation heavily impacts on registration accuracy and runtime, improving over traditional binning techniques.     

Multimodality image registration with software: state-of-the-art

Introduction  

Multimodality image integration of functional and anatomical data can be performed by means of dedicated hybrid imaging systems or by software image co-registration techniques. Hybrid positron emission tomography (PET)/computed tomography (CT) systems have found wide acceptance in oncological imaging, while software registration techniques have a significant role in patient-specific, cost-effective, and radiation dose-effective application of integrated imaging.     

The influence of MRI scan position on image registration accuracy,target delineation and calculated dose in prostatic radiotherapy

Objective

To investigate the necessity of performing MRI in the radiotherapy position when using MRI for prostatic radiotherapy.

Methods

20 prostate patients received a CT, diagnostic MRI and an MRI scan in the radiotherapy position. The quality of registration between CT and MRI was compared between the two MRI set-ups. The prostate and seminal vesicles were contoured using all scans and intensity modulated radiotherapy (IMRT) plans were generated. Changes in the target volume and IMRT plans were investigated. Two-tailed paired Student''s t-tests determined the statistical significance.

Results

There was a decrease in the mean distance from the centre of the bony anatomy between CT and MRI (from 3.9 to 1.9 mm, p-value<0.0001) when the MRI scan was acquired in the radiotherapy position. Assuming that registering CT with an MRI scan in the radiotherapy position is the gold standard for delineating the prostate and seminal vesicles, using a planning target volume delineated on the CT with a diagnostic MRI scan viewed separately, resulted in a mean conformation number of 0.80 instead of the expected 0.98 (p<0.0001).

Conclusion

By registering CT with an MRI scan in the radiotherapy position, there is a statistically significant improvement in the registration and IMRT quality.

Advances in knowledge

To achieve an acceptable registration and IMRT quality in prostatic radiotherapy, neither CT with a separate diagnostic MRI nor CT registered to a diagnostic MRI will suffice. Instead, a CT registered with an MRI in the radiotherapy position should be used.It is of the utmost importance that the radiotherapy (RT) planning process accurately defines the gross tumour volume (GTV) and organs at risk for successful patient management. Many centres routinely register MRI to CT data sets to take advantage of the superior soft-tissue contrast of MRI and the electron density information from CT.The use of dose escalation techniques, such as intensity modulated radiotherapy (IMRT), enables the delivery of high radiation dose to irregular target volumes with increased sparing of healthy tissue over conventional static field RT. Additionally, the more conformal delivery of radiation dose places greater importance on the localisation of the GTV, given that IMRT can generate high doses with steep dose gradients.Typically, external beam RT planning uses CT axial images to define the tumour volume. Because CT offers excellent visualisation of bony anatomy and electron density information, it has largely been the favoured modality for RT planning. However, CT is poor at distinguishing between structures with similar electron densities. An example of this is the prostate and seminal vesicles, where it is difficult to distinguish the extent of the prostate from the surrounding soft tissue.The improved soft-tissue contrast of MRI enables more accurate localisation of adjacent critical soft-tissue over CT [1]. MRI has reduced interference from metal implants such as prosthetic hips [2] and gold seed implants and therefore has the potential to provide improved target volume delineation over CT. Changes in anatomical and tumour definition as a result of using MRI data compared with CT have been reported for prostate [2-7], head and neck [8-11] and brain cancer [12,13] patients.Registration of MRI to CT is an effective method of gaining both the improved target definition of MRI and the geometric accuracy and electron density of CT. However, although MRI provides improved soft-tissue contrast, there are several difficulties with integrating MRI into the RT planning process. MRI is known to suffer from geometric distortion owing to the non-linearity of the imaging gradients over large fields of view [14]. Today, most vendors provide in-plane distortion correction to deal with this [5]. Owing to the lengthy scan times of MRI, motion artefacts can diminish the image quality and alter the accuracy of target localisation. Magnetic susceptibility should also be considered, as this can affect the homogeneity of the main magnetic field, leading to further distortion [15,16]. It is also important that patients are positioned the same way in MRI as they are for CT. Changes in patient position in MRI can lead to misalignments when registering with CT. The difference in the shape of the MRI table from the CT table is one reason for the misalignment in registration.There have been a number of studies published investigating rigid registration accuracy [17-20]. Moore et al [17] assessed the registration accuracy of their treatment planning system by outlining structures on a registered phantom. Using the position, area and perimeter of each structure and with a coefficient of agreement analysis, the accuracy of the image registration algorithm was determined. Another group [18] used anatomical landmarks to determine the registration accuracy of CT and MRI data sets.The aim of this study was to determine how different patient positions in MRI affected the registration quality with CT. A further goal of this research was to examine the changes in the target volume and how this affects RT planning. It was the overall intention of this article to determine if it was necessary to obtain an MRI scan in the RT position for RT planning or whether a diagnostic MRI scan would suffice.An attempt was made to register the CT and MRI data sets using the prostate seeds as registration points to then transform the MRI coordinate system to the CT. Our treatment planning system (Eclipse™, Varian Medical Systems, Inc., Palo Alto, CA, v. 8.6.15) could not perform the registration because of the close proximity of the seeds to one another. Changing the volume of interest to include only the prostate so that the registration was performed using the pixel data within the volume of interest resulted in poor registration. There was also no change to the registration results by prioritising the prostate volume. Although deformable registration is available for Eclipse, it does not allow multimodality registration. As with most UK RT centres, deformable CT–MRI registration is not used clinically at our centre and so was not investigated.Although it could be anticipated that imaging the patient in the same orientation as they receive their treatment would result in an improvement in the registration, many RT centres neither register the CT with MRI nor ensure the MRI positioning is consistent with treatment. Furthermore, the authors were unable to find a publication which rigorously tests the theory presented in this study. It is anticipated that by addressing the problems related to changes in the tumour volume, registration and, importantly, the subsequent effect on prostatic RT, this will facilitate an adjustment in current practice.     

Anatomical accuracy of abdominal lesion localization. Retrospective automatic rigid image registration between FDG-PET and MRI

Software-based image registration can improve the diagnostic value of imaging procedures and is an alternative to hybrid scanners. The aim of this study was to evaluate the anatomical accuracy of automatic rigid image registration of independently acquired datasets of positron emission tomography with 18F-deoxyglucose and abdominal magnetic resonance imaging. Patients, methods: Analyses were performed on 28 abdominal lesions from 20 patients. The PET data were obtained using a stand-alone PET camera in 14 cases and a hybrid PET/CT scanner in 9 cases. The abdominal T1- and T2-weighted MRI scans were acquired on 1.5 T MRI scanners. The mean time interval between MRI and PET was 7.3 days (0-28 days). Automatic rigid registration was carried out using a self-developed registration tool integrated into commercial available software (InSpace for Siemens Syngo). Distances between the centres of gravity of 28 manually delineated neoplastic lesions represented in PET and MRI were measured in X-, Y-, and Z-direction. The intra- (intraclass correlation 0.94) and inter- (intraclass correlation 0.86) observer repeatability were high. Results: The average distance in all MRI sequences was 5.2±7.6 mm in X-direction, 4.0±3.7 mm in Y-direction and 6.1±5.1 mm in Z-direction. There was a significantly higher misalignment in Z-direction (p<0.05). The misalignment was not significantly different for the registration of T1- and T2- weighted sequences (p=0.7). Conclusion: The misalignment between FDG-PET and abdominal MRI registered using an automated rigid registration tool was comparable to data reported for software-based fusion between PET and CT. Although this imprecision may not affect diagnostic accuracy, it is not sufficient to allow for pixel-wise integration of MRI and PET information.     

The principal axes transformation--a method for image registration

We have developed a computational technique suitable for registration of sets of image data covering the whole brain volume which are translated and rotated with respect to one another. The same computational method may be used to register pairs of tomographic brain images which are rotated and translated in the transverse section plane. The technique is based on the classical theory of rigid bodies, employing as its basis the principal axes transformation. The performance of the method was studied by simulation and with image data from PET, XCT, and MRI. It was found that random errors in determining the brain contour are well tolerated. Progressively coarser axial sampling of data sets led to some degradation, but acceptable performance was obtained with axial sampling distances up to 10 mm. Given adequate digital sampling of the object volume, we conclude that registration by the principal axes transformation can be accomplished with typical errors in the range of approximately 1 mm. The advantages of the technique are simplicity and speed of computation.     

An image registration strategy for multi-echo fMRI.

Functional magnetic resonance imaging (fMRI) based on multiple-echo T(2)* mapping has attracted much attention recently. The contrasts in the parametric T(2)* maps are usually too low to allow direct image registration. In this study, an image registration strategy has been proposed for single-shot multi-echo data sets acquired for dynamic T(2)* mapping. We performed image registration of the T(2)*-weighted images before the calculation of the T(2)* parameter maps using two different strategies. One is to perform separate image registration on each echo and the other is to use the same motion correction parameters extracted from the second echo for all the data. Both strategies increase the number of activated voxels and reduce the effective noise level. The results also indicate that, for a single-shot dual-echo image data set, it is slightly preferable to use the second echo for direct image registration and then apply the same motion correction parameters to the first echo images. J. Magn. Reson Imaging 1999;10:154-158.     

Directed graph based image registration

In this paper, a novel image registration method is proposed to achieve accurate registration between images having large shape differences with the help of a set of appropriate intermediate templates. We first demonstrate that directionality is a key factor in both pairwise image registration and groupwise registration, which is defined in this paper to describe the influence of the registration direction and paths on the registration performance. In our solution, the intermediate template selection and intermediate template guided registration are two coherent steps with directionality being considered. To take advantage of the directionality, a directed graph is built based on the asymmetric distance defined on all ordered image pairs in the image population, which is fundamentally different from the undirected graph with symmetric distance metrics in all previous methods, and the shortest distance between template and subject on the directed graph is calculated. The allocated directed path can be thus utilized to better guide the registration by successively registering the subject through the intermediate templates one by one on the path towards the template. The proposed directed graph based solution can also be used in groupwise registration. Specifically, by building a minimum spanning arborescence (MSA) on the directed graph, the population center, i.e., a selected template, as well as the directed registration paths from all the rest of images to the population center, is determined simultaneously. The performance of directed graph based registration algorithm is demonstrated by the spatial normalization on both synthetic dataset and real brain MR images. It is shown that our method can achieve more accurate registration results than both the undirected graph based solution and the direct pairwise registration.     

Real-time 3D image registration for functional MRI.

Subject head movements are one of the main practical difficulties with brain functional MRI. A fast, accurate method for rotating and shifting a three-dimensional (3D) image using a shear factorization of the rotation matrix is described. Combined with gradient descent (repeated linearization) on a least squares objective function, 3D image realignment for small movements can be computed as rapidly as whole brain images can be acquired on current scanners. Magn Reson Med 42:1014-1018, 1999.     

PET与MRI三维脑图像高精度配准

目的 在PC机上实现高精度的PET与MRI三维脑图像配准。方法 采用最大互信息法对6例患者PET和MRI三维脑图像进行刚体配准。使用归一化互信息作为相似性量度。在互信息计算过程中，使用Powell多参数优化法和Brent一维搜索算法。为加快配准速度，使用了多分辨金字塔方法。采用基于坐标的阈值选取方法对PET图像进行分割预处理，消除星状背景伪影。结果 配准误差平均值为2．6mm，误差中位数平均为2．7mm。结论 配准视觉效果良好，评估证明该算法可达亚体元级配准精度。     

Validation of deep-learning image reconstruction for coronary computed tomography angiography: Impact on noise,image quality and diagnostic accuracy

BackgroundAdvances in image reconstruction are necessary to decrease radiation exposure from coronary CT angiography (CCTA) further, but iterative reconstruction has been shown to degrade image quality at high levels. Deep-learning image reconstruction (DLIR) offers unique opportunities to overcome these limitations. The present study compared the impact of DLIR and adaptive statistical iterative reconstruction-Veo (ASiR-V) on quantitative and qualitative image parameters and the diagnostic accuracy of CCTA using invasive coronary angiography (ICA) as the standard of reference.MethodsThis retrospective study includes 43 patients who underwent clinically indicated CCTA and ICA. Datasets were reconstructed with ASiR-V 70% (using standard [SD] and high-definition [HD] kernels) and with DLIR at different levels (i.e., medium [M] and high [H]). Image noise, image quality, and coronary luminal narrowing were evaluated by three blinded readers. Diagnostic accuracy was compared against ICA.ResultsNoise did not significantly differ between ASiR-V SD and DLIR-M (37 vs. 37 HU, p = 1.000), but was significantly lower in DLIR-H (30 HU, p < 0.001) and higher in ASiR-V HD (53 HU, p < 0.001). Image quality was higher for DLIR-M and DLIR-H (3.4–3.8 and 4.2–4.6) compared to ASiR-V SD and HD (2.1–2.7 and 1.8–2.2; p < 0.001), with DLIR-H yielding the highest image quality. Consistently across readers, no significant differences in sensitivity (88% vs. 92%; p = 0.453), specificity (73% vs. 73%; p = 0.583) and diagnostic accuracy (80% vs. 82%; p = 0.366) were found between ASiR-V HD and DLIR-H.ConclusionDLIR significantly reduces noise in CCTA compared to ASiR-V, while yielding superior image quality at equal diagnostic accuracy.