|Home | About | Journals | Submit | Contact Us | Français|
A growing number of clinical applications using 2D/3D registration have been presented recently. Usually, a digitally reconstructed radiograph is compared iteratively to an x-ray image of the known projection geometry until a match is achieved, thus providing six degrees of freedom of rigid motion which can be used for patient setup in image-guided radiation therapy or computer-assisted interventions. Recently, stochastic rank correlation, a merit function based on Spearman's rank correlation coefficient, was presented as a merit function especially suitable for 2D/3D registration. The advantage of this measure is its robustness against variations in image histogram content and its wide convergence range. The considerable computational expense of computing an ordered rank list is avoided here by comparing randomly chosen subsets of the DRR and reference x-ray. In this work, we show that it is possible to omit the sorting step and to compute the rank correlation coefficient of the full image content as fast as conventional merit functions. Our evaluation of a well-calibrated cadaver phantom also confirms that rank correlation-type merit functions give the most accurate results if large differences in the histogram content for the DRR and the x-ray image are present.
2D/3D registration, the computation of six degrees of freedom (dof) of rigid body motion from an iterative comparison of projective images such as radiographs or portal images, has gained clinical acceptance as a non-invasive tool for patient setup and motion control in radiotherapy (Bednarz et al 2009; Cho et al 2008) and image-guided interventions in general. Real-time tumor motion control based on intra-fractional x-ray imaging (Cho et al 2008; Sharp et al 2004) is also a mid-term research objective in image-based 2D/3D registration; for intensity-based registration not requiring further invasive measures such as implantation of markers (Sharp et al 2004) or transponders (Sawant et al 2009), the robustness and convergence range of the merit function used for registration is a prerequisite since 2D/3D registration based on digitally rendered radiographs (DRRs) is indeed a complex registration task. On the other hand, it avoids the drawbacks of invasive methods, namely the trauma associated with such a procedure and further implications on follow-up imaging using magnetic resonance imaging (MRI) (Zhu et al 2009).
Numerous merit functions for 2D/3D registration were presented (Penney et al 1998); specific problems related to 2D/3D registration include the fact that mutual information—a universal and well-known merit function for multi-modal 3D/3D registration—may prove unreliable in the case of 2D/3D registration (Penney et al 1998; Birkfellner et al 2009) due to the sparse population of the joint histogram used for the computation of Shannon's entropy. Furthermore, some merit functions known to give very exact results may show poor success rates due to a poor convergence behavior; most methods based on difference images and gradient images fall in this category. On the other hand, merit functions based on classical correlation measures show a good convergence behavior by avoiding strong local minima and featuring a wide convergence range. Among these are cross correlation, which by definition requires a strict linear relationship between image intensities in the DRR and the reference image (Birkfellner et al 2007) and local normalized cross correlation (Khamene et al 2006), the latter showing good performance but requiring a careful radiometric calibration of the imaging device. From our experience, one cannot easily favor one of these methods above the other since the problem of 2D/3D registration is not well defined, and therefore an arsenal of possible registration measures is required to solve a specific task.
In Birkfellner et al (2009), a merit function called stochastic rank correlation (SRC) based on Spearman's rank correlation coefficient was presented. Rank correlation requires a monotonic relationship between gray values and is, by definition, robust against outliers. It is therefore not suitable for multimodal image registration, but these properties nevertheless render this measure as an attractive alternative for 2D/3D registration.
The drawback of rank correlation is the fact that all gray values encountered in the reference radiograph and the DRR have to be sorted, which has a deleterious effect on the iterative registration process. SRC compensates this by computing a static random mask valid for both images which selects a small percentage of image content. From this subsample of the images, the rank correlation coefficient is computed. In Birkfellner et al (2009), it is shown that a rank correlation coefficient computed from 3 to 5% of image content can provide significantly better results compared to classical cross correlation on image pairs with equalized histograms while the computational effort is similar. In this note, we present a variation of the SRC algorithm which allows for using full image content without significantly corrupting registration speed.
The basic concept of SRC is described in Birkfellner et al (2009); in order to avoid the time-consuming sorting of pixel gray values using the introsort algorithm while avoiding random subsampling of the image, we utilize the fact that the number of pixels in a medical image is usually much higher than the intensity depth of the image. In 2D/3D registration, it is for instance common to select an appropriate window for the source computed tomography (CT) scan used for generating the DRR and to scale the volume down to 8 bit, resulting in 256 levels of gray. Given this circumstance, the sorting problem with a performance of order N log(N) on average can be replaced by a simplified bucket sort. N is the number of pixels used. Here, a histogram of the DRR is computed; since all pixels belonging to the same histogram class have the so-called tied rank since they are equivalent, an average rank computed from the boundaries of the histogram bin is assigned to each bin and further sorting of gray values is not necessary. The average rank i for each histogram bin i—that is, the mean of ranks belonging to the same histogram class in the ordered list of pixel gray values—is computed by
where ρimin and ρimax are the minimum and maximum rank for the pixels belonging to one histogram bin. ρimin is determined by the number of pixels that are in bins 0,…,(i−), whereas ρimax is determined by adding the number of pixels in the histogram bin i to ρimin. If the number of histogram bins is equal to the intensity depth of the images used for registration, the process of sorting a hash table of gray values and associated pixel coordinates can be replaced by a histogram computation of order N. The computation of Spearman's rank correlation coefficient RC (Birkfellner et al 2009)—which is used as the merit function for optimization—from the tied ranks is straightforward:
with I being the number of histogram bins, equivalent to image intensity depth, the (i(IDRR) − i(Ix−ray))2, N the number of pixels used. Equation (2) is defined in such a manner that it yields a minimum for an optimal match of the DRR IDRR and the radiograph Ix–ray.
The algorithm was implemented on a standard Dell Precision 490 PC with 4 GB of RAM and an nVidia Quadro 7500 graphics card. DRRs were generated using a GPU-based splat rendering routine (Birkfellner et al 2005) described in Spoerk et al (2007). A simplex optimization algorithm from the GSL (Free Software Foundation, Boston/MA) was used for minimization. The implementation of the RC-based merit function RC was compared to merit functions based on cross correlation (CC), histogram-based normalized mutual information (MI) (Penney et al 1998), correlation ratio (CR) (Skerl et al 2008) and SRC using 5% image content (Birkfellner et al 2009). The image content percentage for SRC was chosen in accordance to the results from Birkfellner et al (2009), where it was shown that a higher percentage does not improve registration outcome.
The results were validated on a novel porcine reference dataset developed by our group. A fresh cadaver head of a minipig was implanted with seven spherical polytetrafluoroethylene (PTFE) markers of 10 mm diameter; the markers were attached to the skull of the pig by cutting a thread using standard Schanz-screws from a orthopaedic fixateur externe set. A radiolucent screw holding the spheres was inserted. The skull was scanned using a Philips Brilliance 64 multislice CT with a resolution of 0.566 × 0.566 0.4 mm3 and a slice thickness of 0.8 mm. The volume was interpolated to a cubic voxel size of 1 mm3. An x-ray of the phantom was taken using the kV-imaging unit of an Elekta Synergy linear accelerator (see figure 1) with 10 mm aluminum spheres attached to the radiolucent screws. The x-ray detector was a Perkin–Elmer XRD amorphous silicon detector driven in a 2 × 2 binning mode with a resolution of 0.8mm2. Since we are mainly interested in high-performance 2D/3D registration for organ motion compensation in radiotherapy and since most clinical scans are not acquired with the high resolution of the original CT dataset, we settled for 1 mm3 voxels; this also increases the performance of the algorithm due to the smaller data size. The gold standard registration was determined by a direct linear transform. The markers were not removed from the images; in additional experiments where the markers were excluded by masking the 2D image, we did not encounter a significant change in the results. Details on the phantom can be found in Pawiro et al (2010). An intrinsic advantage of this phantom is the large amount of soft tissue connected to the cadaver, which gives a realistic clinical situation even for registration of organs other than the head.
For evaluation of the merit functions, known gold standard registration parameters were randomly changed in such a manner that the initial target registration error (iTRE) as computed from 400 randomly chosen positions on the cadaver specimen ranged from 0 to 15 mm. For the monitoring of intra-fraction tumor motion, this is a reasonable range if the patient is precisely aligned and if the tumor target volume is selected as the region-of-interest for registration. The protocol is in accordance with other standard evaluations found in the literature (van de Kraats et al 2005). One hundred and fifty registrations were carried out for each merit function. We have recorded both the resulting projected TRE (mean projection distance mPD) and the average Euclidean distance of the resulting 3D positions for the reference locations in 3D, which gives the mean target registration error (mTRE). The minimum rendering threshold applied to the CT dataset varied from −868 to −243 HU.
Table 1 gives the results for all registration runs as mPD and mTRE, both as mean and standard deviation, and median. In general, it shows that RC, SRC computed from 5% image content and CR give the best results, which are usually very close. However, RC tends to outperform the CR in cases where the image histogram content shows strong differences due to the rendering threshold used for DRR generation. The time required for the evaluation of a single merit function is dependent on the rendering threshold used. This is a property of the volume-driven splat rendering method (Birkfellner et al 2005) used for DRR generation. However, the average times necessary are very similar (0.12 s for CC, RC, CR and MI, and 0.07 s for SRC 5%—the latter being an exception since the runtime heavily depends on the percentage of the image content used for evaluation). Our optimization setup—a repeated optimization by using a simplex algorithm—requires a very similar number of iterations for convergence; therefore, the runtime for all merit functions excluding SRC 5% was very similar. From the data on time requirements for SRC presented in Birkfellner et al (2009), it can be extrapolated that SRC with 12% image data would on average require the same amount of time as CC, RC, CR and MI.
The proposed modification of the SRC algorithm renders rank correlation as an attractive alternative to common merit functions in 2D/3D registration, provided that the intensity depth of the images used is considerably smaller than the number of pixels used—an image of 256 × 256 pixels image size that uses the full range of 16 bit intensity depth would not benefit from the proposed algorithm in its current form. However, this scenario is rather academic since diagnostic CT scans used as source volumes for DRR rendering usually feature 12 bit resolution, and it is questionable whether full exploitation of 16 bit depth provides any advantage for the purpose of registration after appropriate windowing. Even the original image data from commercial flat panel detectors, which are provided in 16 bit resolution, only use a small part of the intensity range possible in our experience.
Our evaluation reveals that the theoretic advantage of using rank correlation-type measures—the independence of histogram content in the DRR and the reference x-ray as long as a monotonic relationship of corresponding gray values is provided—is also visible in practice. This is especially true for large histogram discrepancy, where the CR is inferior to RC and SRC according to our results. The performance of MI—a measure known for its outstanding performance in multimodal 3D/3D registration—is somewhat disappointing. This is, however, a well-known fact from the literature (Penney et al 1998; Birkfellner et al 2009) as pointed out in the introduction.
Furthermore, one may ask whether the computation of RC results in an advantage over SRC, namely since the latter is even more performant by means of computational load in the present evaluation. The data on algorithm stability and performance do, however, indicate that RC gives more accurate results with a lower variance compared to SRC. We conclude that both methods—RC and SRC—have the potential to solve a given clinical task in 2D/3D registration. As pointed out in the introduction, a wide range of methods is desirable in this case since there is no golden rule.
The limitation of rank correlation lies in the fact that it is, by definition, a merit function for intramodal registration; a registration of x-ray image data and DRRs rendered from MRI is not feasible. The advantages of rank correlation—namely the robustness against outliers and the invariance to variations in image intensity and detector characteristics—do however outweigh this drawback for the specific problem of x-ray and CT data. A future task for further research is a throughout evaluation of the applicability of rank correlation to 2D/3D registration using electronic portal images. It is especially interesting to see whether a robust merit function such as RC can also cope with pairs of simultaneously acquired kV and MV images, which may be retrieved from a number of modern image-guided radiation therapy units.
We wish to thank D Georg, M Stock (Department of Radiotherapy, MUW), F Kainberger, I Nöbauer (Department of Diagnostic Radiology, MUW) and H Bergmeister (Center for Biomedical Research, MUW) for their help in the acquisition of the porcine reference image data. This work was supported by the Austrian Science Foundation FWF under projects P19931 and L503. JH was supported by FWF project L625.