SCreg: a registration-based platform to compare unicondylar knee arthroplasty SPECT/CT scans

Despite the numerous advantages of unicondylar knee arthroplasty (UKA) compared to total knee arthroplasty [1‐5] and its increasing use in clinical practice (+ 32.5% per year in US [6]), this procedure has still a lower survivorship at long term (70% vs 89% at 15 years [7]). The main causes explaining this low survivorship are aseptic loosening and osteoarthritis progression [8] which can both be linked to bone remodeling [9‐13]. Therefore, an indicator based on bone remodeling [9, 10, 14‐16] could be greatly beneficial to improve UKA survivorship. In order to characterize bone remodeling, spatial patterns and temporal evolutions of osteoblastic activity, which is representative of bone remodeling following arthroplasty, can be observed.

SPECT/CT is the modality typically used to measure those spatiotemporal changes in osteoblastic activity. SPECT, or single photon emission computed tomography, can quantify the intensity of osteoblastic activity [17] and its combination with CT, or computed tomography, provides the documentation of the anatomic location of the activity.

However, in order to observe changes in osteoblastic differences through time or between different surgical parameters, SPECT/CT scans must be compared both longitudinally, i.e. at different timepoints for a same patient, and/or transversely, i.e. between different patients.

Nevertheless, due to differences in patient positioning within the scanner or anthropometric differences between patients, SPECT/CT scans are usually not mutually aligned and thus cannot be directly compared. To overcome this challenge, regional classification systems, in which several regions are manually defined based on anatomic landmarks and/or predefined dimensions, have been developed for evaluating unicondylar knee [14, 18], total knee [9, 10, 19], total hip [20, 21] and shoulder [22] arthroplasty. Despite prior success with this approach, it is limited in that it significantly reduces data resolution as thousands of SPECT voxels are reduced into just a few regions for analysis (around 200 voxels per region) leading to the loss of local information. Therefore, full resolution comparisons are required to avoid any loss of information with the potential to give better insight in the evolution of osteoblastic activity following UKA.

By consequence, an image registration problem [23‐25] must be solved in order to spatially align SPECT/CT scans and enable full resolution comparisons. Considering the differences in patient positioning during SPECT/CT acquisitions, several approaches have been proposed in the literature to solve this problem – for instance, standalone rigid transformations [26], global scaling [27], and the use of external markers to combine these approaches [28, 29]. Despite their success in intra-subject registration of hard tissues, such as bone, these procedures cannot tackle inter-subject alignment, since they do not consider anthropometric variability. Non-rigid transformations, which could overcome this issue, have only been used to further optimize intra-subject alignment of SPECT/CT scans [30‐32]. However, to enable an accurate inter-subject alignment necessary to perform transverse analysis of SPECT/CT scans, variability in the joint condition, either due to anthropometry or due to surgical intervention such as arthroplasty must be taken into account. Indeed, this variability causes the registration problem to be more challenging to solve and, to our knowledge, there is no reported study in the literature discussing registration procedures to align hybrid SPECT/CT scans both within and between UKA subjects.

Therefore, the aim of this study is to develop and validate a platform, further referred as SCreg (standing for SPECT/CT registration platform), to spatially align hybrid SPECT/CT scans of UKA patients with the potential to allow full-resolution intra-subject and inter-subject comparisons of osteoblastic activity patterns using a combination of rigid, affine and non-rigid transformations.

Qualitative validation showcased successful registrations for all subjects with the absence of image blur, irregularities and loss of bone contours.

Registration resulted in significantly improved DSC values greater than 80% for all VOIs defined previously (Table 2). Different resolutions of pre-operative and post-operative CT images revealed a significant impact on DSC following intra-subject registration. Average distances between the coordinates of selected landmarks following intra-subject and inter-subjects registrations were within the satisfactory limit (Table 3).

The areas of increased osteoblastic activity on the aggregate map were located below the tibial insert in the posterolateral position (Fig. 2) and on the medial patella facet. A decrease of the activity was observed in the rest of the bone.

The in-house SCreg successfully performed intra-subject and inter-subject registration of CT images from subjects undergoing UKA, thereby allowing full-resolution comparisons, as well as identification of post-surgical rise in SPECT activity. To our knowledge, this is the first study presenting a complete pipeline for registration and analysis of UKA SPECT/CT scans intra- and inter-subjects including the additional steps necessary to go from raw images to the SPECT analysis of a population. Indeed, in addition to the tuning of the parameters of the Elastix toolbox, pre-processing and analysis tools as well as the complete workflow were described. The pre-processing step allowing a normalization and optimization of the scans prior the registration while limiting the required user inputs and the post-processing providing analysis tools to summarize an entire population data into one single representation. However, as it will be discussed hereunder, each part of the platform is a compromise between quality, time and user input in order to have the most suitable application for the clinic environment.

For intra-subject registration, the use of affine transformation greatly limits the possibility of image blur and holes within the image insuring a conservation of image integrity. Concerning the inter-subject registration, the higher magnitude of deformations inherent to the use of non-linear transformations could have led to invalid results. However, the use of an affine registration prior to the non-rigid one, the correct parametrization of the control grid for the B-spline transformations and the hierarchical approach allowed to preserve image integrity.

DSC values, although satisfactory post-registration (Table 2), have the potential to be further improved. In the case of intra-subject registrations, the use of additional non-linear transformations could improve the results; however, this would increase the processing time and the risk of image degradation. In the case of inter-subject registration, a thinner grid could allow more flexibility to the non-linear transformations, thereby improving DSC values. Although, parameter tuning revealed a higher risk of degrading the images during this implementation. It should also be noted that inter-subject registrations were performed using a reference subject randomly selected and results could have been improved if the selected reference subject was the one with the least differences with respect to the other subjects. However, satisfactory DSC despite the random selection of the reference scan demonstrate the robustness of the process.

Significant differences in DSC values were obtained for the intra-subject registration performed on CT images with different resolutions, but not for the inter-subject registration. This could be explained by the use of non-linear transformations in inter-subject registration, which compensates for these differences. However, since this cannot be adopted in intra-subject registration for the aforementioned reasons, it is advisable to acquire the CT images with identical resolution for comparative analysis.

The distance between the coordinates also indicates a successful registration since all values are below (2;2;0) voxels. By consequence, the SPECT voxel associated to a CT voxel on co-registered scans is the same thereby allowing voxel by voxel comparison of SPECT intensities. In addition, the landmarks selected to perform that comparison have an intra-rater and inter-rater reliability below the CT voxel size [33] attesting the robustness and the validity of that measurement.

Concerning the SPECT analysis, aggregate maps created with SCreg were compared with the clinical interpretation of a nuclear radiologist based on the classification system performed on a population including the patients analyzed in this project [18]. Similarly, a general decrease of activity was observed but no increased normalized activity was noticed neither on the tibia nor the patella region. Nonetheless, the tibial area identified in this study as a common region of increased osteoblastic activity correlates with increased strain area following UKA [46]. Bone strain being directly linked to osteoblastic activity through Wolff’s law, it was expected to observe an increase of activity in that region. By consequences, the differences between the two approaches could be explained by the averaging performed within the classification method. Indeed, the areas located in this study could have been separated in the study using the classification scheme hence the increase would have been masked off by the surrounding decrease of activity. These differences suggest that the use of full resolution could provide more insight into bone remodeling following UKA compared to the classification system.

One limitation of this study is the population size (40 scans in total). However, we found it to be similar to the 45 scans used by Papavasileiou et al. [29] (although obtained through simulation for validation purposes) and much higher to other studies using less than 10 scans [28, 30, 31]. However, as a limitation, it must to be noted that all prostheses were identical, thereby limiting the variability of images during registration. Another limitation is the beam hardening artefact caused by the metallic implant during CT scan. Although, the use of a metal artifact reduction algorithm during CT reconstruction limited these artefacts, some streaking artefacts were still visible in some patients; however, these were mitigated by masking the prosthesis during registrations, thereby affecting results minimally.

The main limitation of the platform is linked to the segmentation of the bones in three parts which is done through cuboidal volumes. In this approach, 2000 points are randomly selected within the volume, such that in a case where the volume is large compared with the bone, a large percent of points selected will be in the background and by consequence the quality of the registration will be affected. Therefore, a segmentation approach that limits the image to the actual bone, could improve the registration process. Nevertheless, the approach we used allows to facilitate the bone segmentation and is most suitable for the surgeons which are considered as the main SCreg users in the future. Indeed, the identification of an accurate segmentation, from a technical point-of-view, allowing a smooth registration process would require significant time and a knowledge surgeons might not especially have in comparison to the location of CT landmarks as in the current approach. Alternatively, by using a large number of samples combined with a high number of resolutions, we can address this challenge. Therefore, the simplified approach used in the SCreg reduces computational time and limits required user inputs, thus reducing time and inter-rater differences. In addition, this method increases the intra- and inter-rater reliability since for specific landmarks, it was shown to be lower than the CT voxel size used in this study [33] and for the other landmarks, although not assessed in details, they are easily identifiable on the CT images.

Concerning the future perspectives, an in-depth analysis of the osteoblastic activity changes following UKA in asymptomatic population with SCreg could allow to identify typical patterns. In addition, comparisons with symptomatic population using statistical maps could allow to identify locations with a significantly different osteoblastic activity change following UKA associated with symptoms.

A registration-based platform was developed and validated to compare pre-operative and post-operative SPECT-CT images within, as well as across, subjects undergoing UKA. SCreg allows full-resolution comparisons without loss of data to the contrary of current analysis methodology. With high efficiency and minimal user input, SCreg could be used by clinicians and/or researchers to improve our knowledge concerning UKA as well as patient’s quality of life. Moreover, with an adaptive algorithm and a generic process pipeline, SCreg is neither limited by the type of surgical intervention, nor by the location in the body.