A fluoroscopy-based planning and guidance software tool for minimally invasive hip refixation by cement injection

The long-term survival of hip prostheses is primarily limited by the occurrence of aseptic loosening [1]. This pathological process involves extensive resorption of bone adjacent to the prosthesis and its replacement by fibrous tissue that offers little mechanical stability. Minimally invasive cement injection, already an established technique in the orthopaedic field of vertebroplasty [22], has in recent years been used experimentally to treat hip prosthesis loosening [8, 24].

Injected cement stabilizes a loosened prosthesis by re-establishing rigid mechanical contact between bone and the existing peri-prosthetic cement mantle [2]. Knowing the amount of cement that reaches its intended target may therefore be important in judging the success of a cement injection procedure.

Cement is injected through hollow needles that access the target lesions. Needle insertion may be guided by two-dimensional (2D) X-ray fluoroscopic supervision. Using intraoperative computed tomography (CT), more accurate guidance of needles’ out-of-plane displacement and rotation may be performed [23], but this presupposes the availability of CT hardware in the treatment room. More recently, systems have become available that use a combination of pre-operative CT and intraoperative fluoroscopy. One such example is XperGuide (Philips, Best, the Netherlands) [16, 25], where the accurate alignment of the patient, fluoroscopy images and the CT image volume depends on specialized hardware.

Once needles have been placed, cement is injected. During cement injection, the amount of deposited cement can directly be monitored by the position of the syringe’s plunger which is marked in millilitres, as shown in Fig. 1a. The spatial distribution of the injected cement can be estimated by looking at 2D X-ray fluoroscope projections as shown in Fig. 1b, where it appears as darkened image areas. The distribution of injected cement may also be monitored using three-dimensional (3D) intraoperative CT as shown in Fig. 1c. Currently, relying on 2D fluoroscopy for monitoring the flow of injected cement during minimally invasive prosthesis stabilization is the clinical standard [8, 24].

Open image in new window

Cement injection is typically performed over the course of fewer than ten minutes due to polymethylmethacrylate (PMMA) bone cement hardening within this time span. Of the three intraoperative monitoring methods shown in Fig. 1, only CT provides true 3D information. CT cannot be continuously active throughout the procedure due to its higher radiation, as well as the need for the surgeon to have access to the patient. This makes 3D flow monitoring with CT scan difficult. Lastly, in practice, CT hardware is seldomly available in the treatment room. The patient needs to be slid into and out of the CT tunnel, and medical staff need to stand clear while the X-ray source is active to reduce radiation exposure. Three-dimensional CT can be performed intermittently, at best [7].

We present a proof-of-concept software tool that enables the volume of injected cement to be quantitatively estimated, using a single pre-operative CT image volume, a pre-operatively defined cement target, and one or more intraoperatively acquired 2D fluoroscopic images. This estimate may be performed for each image as it is read from the fluoroscope. The software is designed to augment specific aspects of a minimally invasive cement injection procedure including planning, execution and analysis.

In this paper, we describe our proof-of-concept planning software and then proceed to analyse its use in a pre-clinical cadaver experiment that was performed at our institution. In this experiment, percutaneous cement injection was planned in HipRFX and subsequently performed as for real human patients by an experienced orthopaedic surgeon.

Substituting real fluoroscopic images with simulated DRRs allow for perfect image registration, as all the required images could be generated using the exact same projection parameters. However, to test our software in a physically representative workflow using realistic clinical hardware, we enacted minimally invasive hip refixation procedures on five ex-articulated cadaver legs. We followed the workflow shown in Fig. 2. An experienced orthopaedic surgeon was asked to perform the minimally invasive cement injection—a procedure with which he was familiar.

Cement injection

An experienced orthopaedic surgeon, who has previously performed percutaneous cement injection in more than thirty patients, performed the procedure on each of the prepared cadaver legs. Standard clinical grade needles (Biomet VerteShark Access, 11Gx15cm), vertebroplasty cement (Biomet Bone Cement V) and mixing sets (Optivac Procedure Set) were used—all supplied by Biomet Europe BV (Dordrecht, the Netherlands). Intraoperative fluoroscopy was performed using a standard clinical C-arm fluoroscope (Philips BV Pulsera , Best, the Netherlands).

As opposed to a traditional percutaneous procedure where no pre-operative guidance was available, an assistant held a 10.1-inch tablet computer (Asus TF700T, Taipei, Taiwan) in view of the surgeon on which the pre-computed DRR snapshots were displayed. A photograph of one of our experiments is shown in Fig. 15.

Open image in new window

For each cadaver, the C-arm was moved to the appropriate position as indicated by the planning snapshot. Vertebroplasty needles were subsequently inserted as per pre-operative plan. As opposed to patients, the bones of the cadavers were not affected by osteolysis and necessitated insertion of the needles by first drilling into the bone using a thin drill bit. Even with pre-drilling, inserting the needles into the bone required a surgical hammer to supply sufficient force. In four of the five legs, both planned needles could be inserted, while needle breakage limited use to inserting a single needle in the remaining leg.

Polymethylmethacrylate (PMMA) bone cement was then mixed and injected under fluoroscopic C-arm guidance. The cement flow was fluoroscopically monitored in real time, as is usually done in this kind of procedure [8, 24]. Cement injection was continued until the periprosthetic space appeared filled or until the cement started leaking into the surrounding tissue.

Figure 16 shows a side-by-side comparison between a fluoroscopy image obtained during the cadaver experiment and a corresponding HipRFX-generated DRR. Both of these images were generated using the actual experimental attained cement distribution—the fluoroscopic image created at the end of the cement injection, and the DRR generated using the post-operative CT volume. When the fluoroscopic image was made, the cement had not yet hardened. Some changes to the cement distribution are visible between the images, especially where cement leaked outside the femoral shaft. The physical incision that resulted from our experimental preparation method can be seen as the radiolucent band into which the cement leaked.

Open image in new window

We observed that much of the leaked cement did so into the incisions created by our experimental method. This is clearly visible in Fig. 16. We compared all computed values with ground truth values measured in post-operative CT. With a median of 81 % (range 55–84 %), the cement leakage in our experiment was higher than observed in real patients [18]. The filling percentage (median 52 %, range 18–55 %) was comparable to that observed in real patients [18]. The properties of the experimentally created cadaver leg lesions and the subsequent injected cement volumes are shown in Table 1.

For each real fluoroscopic image, HipRFX was used to estimate the amount of cement that was injected into the target region. Two of the five legs were each imaged from two different angles. The median estimation error, expressed as a percentage of the cement target volumes, was 5.2 % (range 1.6–14.9 %). These results are also summarized in Table 1.

To examine the role that experimental inaccuracies and assumptions had on our results, we further differentiated the absolute per-pixel errors made in the 2D “filling images” like those in Fig. 11. We compared results when using either (a) our real experimentally obtained fluoroscopic images that include unaccounted-for cement leakage, image noise and possible geometric-and-image-intensity registration errors, (b) simulated fluoroscopic images that include unaccounted-for cement leakage but with perfect registration and (c) simulated fluoroscopic images with no cement leakage. For each of these scenarios, we furthermore distinguished between the estimation error made in the “computable” and the “non-computable” interpolated regions that are described and illustrated in Fig. 14. Results are shown in Fig. 17.

Open image in new window

The simulated fluoroscopic images for (b) were DRRs that used the complete post-operative 3D cement distribution, segmented in the post-operative CTs. The simulated fluoroscopic images for (c) were identical to (b) except that it excluded all leaked cement. As our cement filling algorithm does not model cement leakage, this modification meets the assumption inherent in our cement estimation algorithm, which is that all injected cement projecting onto the image of the cement target must also be located within the cement target.

After performing the experiment, we discussed our proof-of-concept software with the orthopaedic surgeon who participated. He valued the pre-operative planning capabilities that the software offers without necessitating additional or new clinical hardware. He stated that HipRFX represented a promising development in image-based guidance and also represents a step towards reducing the need for intraoperative CT imaging.

Minimally invasive hip refixation by cement injection is a novel and experimental technique for refixing loosened orthopaedic implants [8, 24]. To our knowledge, there currently exists no task-specific tool for planning or performing this procedure. Our experimental HipRFX system has been purpose-designed for this task and uses existing fluoroscopic hardware to provide intraoperative guidance and to analyse the distribution of injected cement. HipRFX can simulate fluoroscopic images and computes cement volume estimates and a combined filling-and-uncertainty distribution. Should it be used intraoperatively, our system could provide a surgeon with valuable extra information that currently is either absent, or can only be obtained after the fact. HipRFX focuses on simulating and analysing the fluoroscopic images a surgeon would see—including the realistic fluoroscopic appearance of injected bone cement. This allows us to progress beyond needle guidance and to also simulate and assess the cement filling.

By visually judging the agreement between a simulated desired outcome and real intraoperative images, a surgeon could be aided in judging whether sufficient cement penetration has been achieved. An obvious deviation between simulated and observed fluoroscopic images may alert him to unwanted outcomes, such as cement leakage. In addition, we experimentally showed that HipRFX can compute numerical estimates of the injected cement volume.

Our pre-operative software tool is built around the assumption that C-arm fluoroscopy is the dominant intraoperative technology for monitoring percutaneous cement injection and will most likely remain so for the foreseeable future. Additionally, we assume that at least a single pre-operative CT image volume will be available before commencement of surgical planning. These two assumptions are compatible with existing radiological protocol for minimally invasive cement injections [8, 24].

The Philips XperGuide [10, 16] system shows some similarities with our approach to intraoperative guidance as it also bases its guidance on a prior CT volume and also combines this with live single-plane C-arm fluoroscopy. However, XperGuide is focused on guiding needle biopsies only and achieves this by overlaying glyphs onto radiological images. Unlike XperGuide, our system is, in principle, compatible with all existing CT imagers and fluoroscopes. This was demonstrated in our cadaver experiment where we used a commercial C-arm fluoroscope that precluded any direct calibration between our software and the imaging hardware. Here we showed that HipRFX is capable of estimating the volume of cement to have reached pre-operatively defined target lesions. While we used post-operative CT as ground truth to compare our results to, post-operative CT does not form any part of our workflow as shown in Fig. 2.

Where they project to the same pixel positions, our algorithm cannot distinguish leaked cement from cement inside the target region. A large amount of leaked cement may therefore negatively affect the accuracy of the cement filling estimation, whereby the algorithm interprets leaked cement to be located in the target lesion. This view is supported by Fig. 17 that shows much reduced estimation errors when we analysed DRRs where leaked cement was digitally removed.

An independent analysis of real patient data showed that cement leakage is prevalent in clinical practice [18]. Given this reality of cement leakage, HipRFX plays the important role of specifically estimating the injected cement fraction that reaches a target lesion. This fraction is presumed to be the one that directly contributes to the stability of a hip prosthesis [2]. Cement leakage was exacerbated in our experiments by the cadaver legs having been sawed through, thereby providing an easy escape route for injected cement. Even in this challenging scenario, the median cement volume estimate error was 5.2 % (range 1.6–14.9 %). We are encouraged by this result, especially considering the limitations of our experimental set-up.

Residual image registration mismatches limited the accuracy with which our algorithm could deduce cement filling. From Fig. 17, we made the paradoxical observation that, for the fluoroscopy images, the estimation error was higher in computable than in interpolated regions. We explain this by random image noise being present in fluoroscopic images that cause per-pixel fluctuations in filling estimation. By contrast, the interpolated regions are smooth and show a lower per-pixel absolute error. This explanation is supported by the observation that the errors in the computable regions decreased when we substituted the fluoroscopic images with DRRs, while the interpolated regions’ error decreased by less. We observed that per-pixel estimation errors tended to cancel out when they were summed, resulting in a reduced estimation error for the total cement volume.

A limitation of HipRFX in its current form is its inability to directly record the position and projection parameters of the relevant fluoroscope. In fact, in our experiments, these parameters were not numerically recorded at all. The simulated C-arm was visually and manually aligned to match the DRRs with the recorded fluoroscopic image. The inevitable discrepancies that resulted from such qualitative alignment required us to perform image registration, as schematically illustrated in Fig. 2. Having to deal with fewer unknowns and the use of iterative techniques that better match radiographs with CT-derived DRRs [12, 26] may, in future, improve registration accuracy. Calibrated systems using accurate angle encoders such as the Philips XperGuide may, in future, be constructed for our application and make the system easier to use and more robust.

In clinical practice, needle placement is mainly performed under the supervision of single-plane fluoroscopy [8, 24]. Intraoperative CT may also be used for more accurate 3D guidance. Similar to Philips XperGuide, our software has the potential to be developed to guide a surgeon during needle placement. One possibility would be to simulate a correctly placed needle and overlay it in live fluoroscopy. We leave the exploration of this topic to future work.

Our system may, in future, be used to compute optimal C-arm orientations for performing an intervention. This can be done by maximizing the image difference between DRRs of the expected pre-operative and post-operative situation. Large image differences indicate that the changes caused by the planned procedure are clearly visible—this then represents an informative viewing angle. Differences may enable discernment between correct and incorrect needle placement, or between sufficient and insufficient cement filling. The parameter space that needs to be traversed in this case consists only of the C-arm’s elevation and azimuth—i.e. a two- dimensional optimization problem. When a C-arm fluoroscope needs to be positioned accurately to within \(5{\circ }\), all possible orientations may be examined in fewer than 800 iterations. Given the speed with which DRRs and difference images can be computed, this brute force approach would be feasible for optimizing the view for each desired step in the cement injection procedure, while being guaranteed to find the global optimum.

Another direction for future work would be to combine the estimates obtained from several distinctly oriented fluoroscopy images. A stereo fluoroscopy system [3] may be used for this purpose, or several single-plane fluoroscopic images may be taken successively from different angles. The area occluded by a large metal prosthesis would be different for each image and the combined estimate may be considerably more accurate than either estimate on its own.

We are encouraged by the feedback received from the orthopaedic surgeon who used HipRFX in planning and executing the cadaver experiment. We foresee that, in future, the techniques described in this paper could provide surgeons with sufficient insight and feedback on percutaneous cement injection procedures to avoid the use of explicit 3D imaging tools in the operating room. We believe that the ability to glean quantitative estimates of the cement volume that is deposited in a desired target region, without having to resort to 3D imaging modalities, is novel. This approach may be applied to any application where radiopaque cement in injected, including vertebroplasty.