Orthopaedic device manufacturers face numerous challenges in maximizing their return on investment. Regulatory agencies are demanding more quantitative and comprehensive data. Legislators are beginning to levy device taxes on medical device companies in order to pay for universal healthcare legislation. And finally, upper management continues to demand more from its R&D investment in response to the tumultuous economic climate. Collectively, these challenges are driving the need for orthopaedic device manufacturers to employ uniquely adapted, automated, and unbiased analytical approaches.

Today’s newest automated and customized image analysis techniques can be tailored to specific imaging modalities (i.e., computed tomography, x-ray, magnetic resonance imaging, 2-D histology), anatomical regions-of-interest, and implant material and morphology. Using these techniques, researchers can quickly and quantitatively assess new bone formation and implant integration or resorption rates via multiple independent methods without user intervention. Validated with appropriate controls, these techniques can provide quantitative and reproducible metrics that can be used to accurately assess implant efficacy. Furthermore, these output parameters can be flexibly configured to suit the specific needs of the study. As new parameters are requested, they can be reapplied multiple times using batch mode operations. As a result, researchers can reduce the time, cost, and guesswork associated with preclinical and clinical R&D programs, in addition to building a strong case for their product when communicating with key stakeholders (i.e., regulatory, upper management, shareholders, clinicians, and patients).

Algorithms as a Foundation for Success

A wide range of open-source and commercial image analysis software is used effectively today for extraction of bone integration metrics. Unfortunately, researchers face complex implant designs and geometries, variations in implant composition and anatomic placement, numerous sample prep permutations (i.e. 2-D staining protocols, sectioning techniques, etc.), and significant variability in acquisition protocols and image quality for a given modality (across multiple sites and equipment vendors). With these demands, researchers seek significant flexibility not only in selecting a software package but also in aligning image analysis engineers to tailor and optimize algorithms for each study.

­­­­­­In general, because off-the-shelf software is uniquely adapted to a specific application/purpose, it often fails with deviation. Thus it is imperative to have engineers who are adequately experienced in software development for biomedical imaging applications and to ensure that these engineers can develop and customize algorithms to meet a study's quantitative goals. Furthermore, since algorithms are study specific, they must be thoroughly validated using expert guidance and verification (i.e. radiologist or pathologist), cadaver models, phantoms, multiple modalities where feasible, and preclinical studies. Additionally, every attempt must be made to develop robust algorithms that are applicable across multiple patients or specimens in a given study. This will enable batch processing and analysis of imaging data without user interaction, thus eliminating introduction of observer bias. Lastly, the quantitative data generated by a particular set of analysis algorithms must be formatted in a logical manner to facilitate statistical analysis and interpretation. Segmented components (i.e. new bone, implant material, etc.) generated during algorithm execution should be saved for subsequent generation of pseudo-colored images or volumes as visual evidence of bone in-growth and integration (see Images 1 and 2).

Quantitative Analysis in 2-D

Histological cross sectioning, staining, and imaging is the current status-quo for assessing the formation of new bone in preclinical studies. Postacquisition, a team of trained technicians read and score hundreds, if not thousands, of histology tissue sections in an attempt to qualify metrics—including new bone area and distribution; preexisting bone; scar tissue; implant or scaffold area; original defect area; evidence of inflammation or infection; and infiltration or presence of specific cell or tissue types—and void areas all within a given region-of-interest. This process is by no means simple and efficient, and it almost always returns a qualitative score (i.e., 1–5) based on an observer's subjective opinion and current disposition (i.e. whether they are well-rested, adequately caffeinated, etc.). Furthermore, with hundreds of slides across multiple observers, it is easy to understand how intraobserver variability is additionally confounded by variability among multiple graders. This could potentially destroy any statistical significance that may have been present between experimental groups.

Due to the large number of outcome parameters required for most studies, it often becomes necessary to subsample histological data as it would be time- and cost-prohibitive to prepare, process, and review every tissue section from a given specimen cohort and across an entire anatomic site or implant. Additionally, because subsampling only evaluates a small subset of an already inherently variable dataset, the likelihood of an erroneous score is very real. For example, if an investigator is trying to measure how well bone integrates into scaffold, he or she could implant this scaffold into a rat or rabbit fibula defect, wait three to six months and then sacrifice the animal and process the site for histology. Depending on the dimensions of the scaffold or defect, the number of histology slides comprising the entire region-of-interest could be quite large.

To manage feasibility, the investigator may opt to only review and score 20–30% of the slides, hoping that this will be enough to demonstrate implant efficacy with adequate statistical relevance. Since this approach leaves 70–80% of the data unaddressed, it is quite conceivable that the data output metrics collected could be skewed in favor of a false-negative or false-positive result. These undersampling errors could have major ramifications for preclinical or clinical study conclusions ranging anywhere from down-grading or abandoning an R&D program for a product that actually works to mistakenly advocating efficacy of a product that could be used in clinical trials or for patient treatment.

One potential solution to mitigate feasibility issues regarding histological evaluation while still providing comprehensive analysis of a given implant or scaffold involves using both 2-D and 3-D imaging and quantitative image analysis approaches to assess bone formation. This approach leverages:

• 3-D nature and totality of microcomputed tomography (Micro-CT).
• Tissue delineation and resolution of digital 2-D histomorphometry. 
• Power of custom-tailored image coregistration, correlation, and segmentation algorithms.

The result is the generation of data that is both quantitative and representative of the entire dataset while eliminating the time, cost, and guesswork associated with standard histological scoring.

In this approach, the entire implant or scaffold is imaged using a Micro-CT to produce a 3-D volume at resolutions approaching 1 µm. Subsequently, exploiting the x-ray attenuation characteristics through various tissue types, newly formed bone and scaffold material are extracted and quantified using morphometric and intensity based segmentation routines. The implant or scaffold is then histologically sectioned and stained, imaging the resulting slides with a high resolution, fast-scanning, large field-of-view microscope. The metrics analyzed in the Micro-CT volumes are also extracted in the stitched mosaics using color space segmentation and 2-D morphometric operations. The resulting automated histological outputs can then be used to verify the 3-D Micro-CT based analysis. The investigator only needs to select a small subset of the slides generated to confirm the bone growth, or lack thereof, observed in the comprehensive 3-D data set.

Image 1. Correlation of Histology and Micro-CT data. Bone in- and on-growth was evaluated in a ceramic biomaterial implanted into a femur using micro-CT imaging and analysis (A), and histological large field-of-view microscopy and analysis (B).

Image 1 demonstrates this multimodality approach. A ceramic implant was placed into a sheep femur and both Micro-CT and histomorphometry data were generated and quantitatively assessed as described. In this image, the color green denotes the implant material, while the color red (A: Micro-CT) and blue (B: histomorphometry) represent new bone.

Quantitative Analysis in 3-D

Aside from tissue biopsies, histological analysis is generally not a feasible option in patient studies created to assess the safety and efficacy of a new drug or device. As a result, noninvasive, 3-D volumetric imaging modalities such as computed tomography (CT), magnetic resonance (MR) and positron emission tomography (PET) have become the default tools for implant assessment. These imaging modalities offer an investigator a great deal of information in multiple dimensions, enabling the comprehensive evaluation of a device or implant from all angles. Each of these imaging techniques, however, has its strengths and weaknesses. For example, CT is best used when an investigator is trying to resolve dense materials like bone and ceramics, but is not ideal when trying to visualize and delineate soft tissue. In direct contrast, MR is particularly effective in differentiating softer materials such as fat, cartilage, and fluid, albeit at lower resolution than CT, but with the added advantage of not imparting any radiation dose.

The ability to view a device implanted into a patient during a clinical trial is an exciting proposition, as it affords the opportunity to study its characteristics and effect on surrounding tissues in 3-D. Unfortunately, much of the data available in a CT or MR volume is not adequately surveyed and quantified due to time constraints in manually delineating features of interest. Traditionally, bone in-growth (within the implant), bone on-growth (around the implant), implant resorption, etc. are assessed by a trained medical professional who reads a volumetric scan and offers a score for a particular metric or pathology. For example, bone growth will be scored as either “yes” or “no,” and sometimes “more” or “less” relative to a scan from a previous time point for a given patient.

More sophisticated scoring systems offer a range of scores, such as 1 through 5, based on internally defined standards. However, as these methods of scoring rely on observers to extract the critical information that is buried within these 3-D datasets, they are prone to undersampling errors. Furthermore, scoring systems that rely solely on human observers tend to yield data that is subjective, qualitative (not quantitative), and carries with it intra- and interobserver variability as described for histological scoring. Therefore, novel implant- and application-specific image analysis approaches for segmenting and quantifying bone integration will be essential, particularly as orthopaedic implant designs and functions become more complex.

As with the previous preclinical example, quantitative image analysis can be custom-tailored for a particular implant and study design. By carefully optimizing image analysis algorithms to account for metrics such as rate of bone growth, implant material, implant morphology and design, anatomical region-of-interest, type of imaging modality, and modality scanning protocols, investigators will have access to robust quantitative data that both they and the image reader (i.e. radiologist) can use to quickly and accurately interpret for satisfaction of clinical trial end points (e.g., bone growth). Rather than the standard yes/no criteria currently used, investigators can report quantitative values such as 4.3% versus 18.9% bone growth. With actual comparable, quantitative values for both growth determined at multiple time points, relative rates of bone integration can also be defined.

An additional image analysis feature available for multitime-point implant trials that require longitudinal evaluation of bone growth is the ability to spatially align each temporal volume for a given patient into the same exact coordinate system by applying volumetric registration algorithms. This enables characterization of density changes over time in the exact same region of interest across all time points for a particular patient. Since these algorithms are probabilistic in nature and not dependent strictly on image intensity, patient scans acquired using multiple modalities may also be spatially registered to assess pathological phenomena related a particular region of interest across modalities that offer significantly different tissue characteristics (i.e. fluid presence assessed in MRI correlated to void spaces in a ceramic implant visible in a CT scan).

Temporal patient scans are often acquired with significant variations in patient positioning or resolution. These variables make it almost impossible for an image reader to objectively and accurately assess bone growth changes over multiple time points. However, with an appropriate combination of validated image analysis algorithms, an investigator can overcome this hurdle and produce image datasets that can be easily read by the medical technicians or automatically analyzed by customized analysis software.

Image 2. Bone tumor healing. Graft remodeling and new bone formation metrics were extracted from the site of a femoral ostonecrotic defect in a longitudinal study using CT imaging.

To illustrate these points, Image 2 contains images acquired during a clinical study wherein patients with bone cancer were treated with a bone graft product to induce bone regeneration in an osteonecrotic defect of the femoral head. For this particular study, the investigator used customized image processing and analysis software to measure the amount of bone growth and density within the defect site. The study consisted of two CT time points acquired approximately 12 months apart. As indicated in the upper panel while the CT scans were acquired using similar positioning, the dataset for second time point was reformatted to the sagittal orientation.

As is general practice for many hospitals, the original raw scan data was discarded, preventing reconstruction of a transverse dataset similar to the first time point. Therefore, before any analysis algorithms could be applied for the extraction of primary study metrics, each patient’s longitudinal CT scans had to be spatially coregistered. The result of this process is displayed in the lower left corner of Image 2. Subsequently, customized image analysis algorithms could be applied to identify and segment bone growth at the defect site using morphometric and intensity (Hounsfield) based characteristics. As a result, quantitative measurements for bone growth (i.e., “bone volume”) were extracted at each CT time point for a given patient. A visual representation of this analysis is shown in the lower right corner of Image 2, where the red/green colors denote segmented bone within the defect at each time point. In addition to bone volume, the distribution and density of bone was also assessed.

Image Analysis for Retrospective and Postmarketing Studies

In addition to its applicability in preclinical and clinical trials, custom-tailored image analysis has particular use in retrospective and postmarketing studies. Frequently, an orthopaedic company will get a new product approved by FDA (or other regulatory agency) and then struggle to get physicians and patients to adopt the new device as an accepted replacement for a preexisting therapy. Most often, this hesitance is related to a dearth of clinical data demonstrating and supporting the efficacy claims for the new implant as a result of obtaining regulatory approval via the 510(k) predicate device pathway. This method of approval does not necessarily require clinical data if the company can justify substantial equivalence for the new device. Obtaining regulatory approval without widespread market adoption can significantly affect sales and return on R&D investment. Therefore, designing and executing a postmarketing study to generate solid clinical data with as little time and capital investment as possible can be critical to the use and eventual acceptance of an approved device.

Preexisting imaging data acquired during a clinical trial for radiologist evaluation can be an ideal source of volumetric data to generate clinically relevant data via customized image analysis for use as marketing collateral. To illustrate this scenario, Image 3 demonstrates the extension of image analysis techniques to a retrospective clinical study for the evaluation of longitudinal bone graft integration or remodeling, new bone growth, and anterior cruciate ligament (ACL) interference screw resorption. This study included 18 patients, each undergoing ACL reconstruction using a resorbable ACL screw implanted into the tibia. For each patient a CT scan was acquired at 1 to 2 weeks, 6 months, 12 months, and 2 years postsurgery.

Image 3. ACL Screw Analysis. ACL fixation screw resorption, tibial tunnel morphometrics, and tunnel bone growth were quantitatively measured and evaluated across multiple time-points using customized CT volumetric registration and analysis.

Since this was a retrospective evaluation, customized image analysis algorithms were developed quickly, and were validated and applied to the data without delays in patient recruitment and compliance. Image analysis for each patient dataset involved spatial coregistration of all available time points to the highest resolution scan available and subsequent application of intensity and morphometric filters for component segmentation (graft, screw, tunnel, newly formed bone) to yield the following metrics: 

• Screw resorption rates and density changes.
• Tibial tunnel morphometrics (i.e., volume and diameter changes).
• Mean Hounsfield unit values.
• Bone graft remodeling.
• New bone volume within the tibial tunnel.

The Power of Automation

Automating an otherwise manual, tedious and time-consuming process reduces not only the cost and time associated with that process, but also alleviates observer-based subjectivity and variability that could corrupt the resulting outcomes. Image processing and analysis automation (i.e. batch analysis) applied to preclinical and clinical implant studies using custom-tailored, robust algorithms can enable investigators to:

• Increase their throughput for large sample sets.
• Account for population variability for their chosen study subject.
• Improve the statistical power of their study.

Additionally, automating the image analysis pipeline can significantly reduce the inter- and intraobserver variability inherent in data generated from manually scored images. As a result, the output data generated is more objective and reproducible. In the event that an image analysis process cannot be fully and reliably automated, a semiautomated approach may be implemented to improve the overall workflow efficiency of the image reader. Employing batch image processing and analysis techniques can provide a significant reduction in the time and cost associated with a preclinical or clinical study, and enable reanalysis of an entire image dataset multiple times if necessary to generate additional parameters or account for adjustments in study aims.

To illustrate the time savings possible with batch image analysis, the histological image shown in Image 1B will take an average observer up to an hour to manually delineate the defect site, implant material, and newly formed bone. For 200 histological sections throughout the implant, this manual approach would require nearly a month for data generation for a single specimen.

Conversely, the segmented output shown in Image 1B was the result of a fully automated algorithm that defined the defect boundary, implant material area, and newly formed bone within the voids of the implant (bone in-growth) and 100 m from the defect boundary (bone on-growth). From image loading to data output to an Excel file, the automated approach required approximately four minutes of computing time without user interaction. Thus for 200 slides, the automated analysis would be complete in two days, a substantial time savings over the manual approach with added precision that is not possible when using multiple observers.

Conclusion

Independent of the study subject (human or animal model), anatomical region-of-interest, implant material/morphology or image acquisition modality, customized image analysis techniques can effectively and quantitatively extract metrics to assess orthopaedic implant osseointegration. These techniques can replace or significantly augment traditional subjective observer evaluation. Automation of analysis routines enables faster throughput for larger sample sizes or patient cohorts, in addition to improving the statistical power of a study to minimize the effects of population variability. By employing novel image analysis algorithms validated with appropriate methods (cadaver and preclinical studies and radiologist/pathologist verification), it is possible to remove time, cost, and guesswork from preclinical and clinical trials, substantially increasing the likelihood that primary and secondary study end points will be met.

Amit Vasanji, PhD, is chief technology officer at ImageIQ (Cleveland), a recent spin-out from the Cleveland Clinic. Brett Hoover is vice president at ImageIQ (Cleveland).

Featured Imagef: