Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2011 Mar 1.
Published in final edited form as: Osteoarthritis Cartilage. 2009 Nov 5;18(3):344–353. doi: 10.1016/j.joca.2009.10.004

Semi-automated Segmentation to Assess the Lateral Meniscus in Normal and Osteoarthritic Knees

MS Swanson †, JW Prescott †,‡, TM Best §, K Powell ‡, RD Jackson , F Haq §, MN Gurcan ‡,*
PMCID: PMC2826568  NIHMSID: NIHMS153653  PMID: 19857510

Abstract

Objective

The goal of this study was to develop an algorithm to semi-automatically segment the meniscus in a series of magnetic resonance (MR) images to use for normal knees and those with moderate osteoarthritis (OA).

Method

The segmentation method was developed then evaluated on 10 baseline magnetic resonance images obtained from subjects with no evidence, symptoms, or risk factors of knee (OA), and 14 from subjects with established knee OA enrolled in the Osteoarthritis Initiative (OAI). After manually choosing a seed point within the meniscus, a threshold level was calculated through a Gaussian fit model. Under anatomical, intensity, and range constraints, a threshold operation was completed followed by conditional dilation and post-processing. The post-processing operation reevaluates the pixels included and excluded in the area surrounding the meniscus to improve accuracy. The developed method was evaluated for both normal and degenerative menisci by comparing the segmentation algorithm results with manual segmentations from five human readers.

Results

The semi-automated segmentation method produces results similar to those of trained observers, with an average similarity index over 0.80 for normal participants and 0.75, 0.67, and 0.64 for participants with established knee osteoarthritis with Osteoarthritis Research International Society International (OARSI) joint space narrowing scores of 0, 1, and 2 respectively.

Conclusion

The semi-automatic segmentation method produced accurate and consistent segmentations of the meniscus when compared to manual segmentations in the assessment of normal menisci in mild to moderate OA. Future studies will examine the change in volume, thickness, and intensity characteristics at different stages of OA.

Keywords: meniscus, segmentation, osteoarthritis, magnetic resonance imaging

Introduction

Osteoarthritis (OA) is the most common form of arthritis and approximately 40% of persons age 55 and older have frequent knee pain or radiographic evidence of knee OA1–3. OA of the knee involves progressive degeneration of both the bone and soft tissues, including joint space narrowing, sclerosis, and osteophyte formation. Up to 6% of patients have at least a one point increase in Kellgren-Lawrence (K-L) grade annually, but it is difficult to predict which patients will progress and at what rate4. Thus, there is a clear need to delineate the factors that put patients at the highest risk for progression to identify targets that may be amenable for intervention.

The health of the articular cartilage and its effects on the extent of knee OA has received some attention5, however the meniscus may also modify the development and progression of OA. The meniscus reduces stress on the cartilage in the tibiofemoral compartment by absorbing and distributing force under increasing loads, lubricating the joint, and contributing to joint stability6–9. During standing alone, the meniscus distributes 30 – 55% of the total body weight across the articular cartilage. Degeneration of the meniscus decreases tibio-femoral contact area by 50 – 70% and accelerates cartilage degradation leading to joint degeneration6–7, 10–19. Although the widely accepted definitions of OA do not directly address the health of the meniscus4, the increased incidence of structural joint degeneration after partial and total meniscectomy16 raises questions about the effects of this structure in the development and progression of OA.

Magnetic Resonance Imaging (MRI) is ideal for quantitative image analysis of the meniscus as it permits detailed viewing of the tissue not accessible to radiographs6, 20–21. Detection of meniscal injury by MRI has an accuracy of up to 85% compared to arthroscopic and clinical examination, thereby supporting its utility in the diagnostic and treatment algorithm of the acutely injured knee20, 22–23. Segmenting the meniscus from MR images could be the foundation of the analysis for changes associated with this tissue and its potential role in OA. Segmentation of the meniscus will permit quantitative measures such as volume, mean intensity, intensity distribution, texture, and thickness. Any number of quantitative measures can be compared with OA progression with the intent to find more reliable biomarkers or yield insight into the natural history of this disease.

Manual segmentation is time consuming, making meaningful analysis on large data sets both labor and time intensive. In addition, manual segmentations are prone to intra and inter-reader variability, both of which can be largely overcome with computer-assisted segmentation. The purpose of this study was to develop a semi-automated segmentation technique to characterize the meniscus in order to reduce segmentation time, inter, and intra-reader variability. The technique was designed to allow for flexibility between patients and across a spectrum of joint degenerative states, while maintaining constraints based on known anatomical information. This tool was not designed to completely eliminate the need for manual segmentation since is cannot perform well on subjects with severe OA or acute meniscal injury, but is rather to be used as a tool to speed up quantitative analysis on large data sets. Indeed, there is a large degree of variation among manual readers in such cases, which is why further manual scrutiny is required for reliable analysis.

Methods

IMAGE ACQUISITION

Data used in this study were obtained from the Osteoarthritis Initiative (OAI) database, which is available for public access at http://www.oai.ucsf.edu/. Images from the 0.B.2 release were used with IRB approval. The algorithm was designed using Sagittal T2 Map 120 mm FOV MRI’s obtained from 3T Siemens machines. The images were 384 × 384 pixels (pixel spacing = 0.3125/0.3125 mm) with a slice thickness of 3 mm. The imaging protocol and limb positioning was uniform, with the lateral meniscus in the right leg being examined in all cases24–25. The right knee was chosen since most work on cartilage quantification available to the OAI has been performed on the right knee. Also, the OAI image database only has data for the right knee available for the image sequence used.

PATIENT SELECTION

The segmentation method was tested on baseline magnetic resonance images obtained on 24 randomly selected participants enrolled in the OAI24; 10 with no evidence of OA, osteophytes, risk factors, or knee symptoms of pain and stiffness, and 14 with evidence of right knee OA defined as frequent knee symptoms in past 12 months and K-L score of ≥ 2 on fixed flexion radiograph (Progression subcohort). Participants with established knee OA were further stratified in this study based on the OARSI joint space narrowing (JSN) score of 0 – 3 based on paired x-ray reading of the lateral compartment of the right knee (Figure 1). Four participants were selected at random from each of JSN = 0, JSN = 1, JSN = 2, and 2 with JSN = 3 totaling 14 patients from the Progression subcohort. Only 2 participants were selected for JSN = 3 because no other participants met this criteria. All participants had JSN grades of 0 or 1 for the medial compartment, indicating that disease was located mainly in the lateral compartment. All analysis was done on the lateral meniscus of the right knee, with future work to examine the medial meniscus as well. Semi-automatic segmentations of 10 lateral menisci from non-exposed control participants were quantitatively compared with manual segmentations, two of which were compared to segmentations of five human raters for intra and inter-reader variability. The 14 lateral menisci from progression patients were quantitatively examined by five manual raters for inter-reader variability, and three of these were segmented twice by each rater for determination of intra-reader variability.

Fig. 1.

Fig. 1

Open in a new tab

An example of a Sagittal view of a (a) normal knee, (b) one with OA OARSI 1, (c) OARSI 2, and (d) OARSI 3. Notice that as the OARSI score increases, meniscus degeneration becomes apparent along with osteophytes and cartilage damage.

SEGMENTATION METHOD

The semi-automated segmentation consisted of five phases: initialization, threshold determination, segmentation, conditional dilation, and morphological post-processing26 (Figure 2).

Fig. 2.

Fig. 2

Open in a new tab

Flowchart of the segmentation process.

Phase I: Initialization/Manual Inputs

The segmentation process starts with two manual inputs: 1) a seed point within the lateral meniscus and 2) the selection of the last slice to be segmented. A seed point was chosen by viewing the sagittal sequence from lateral to medial, and choosing a pixel within the lateral meniscus on the first slice that it appeared. In some participants, an image containing merged anterior and posterior horns was not present, so one seed pixel was selected for each horn of the meniscus at their most lateral extent. The final slice that showed the meniscus was also determined manually, providing the algorithm an end point. This endpoint was necessary to manually determine the boundary between the medial extent of the meniscus horns and the meniscofemoral and cruciate ligaments (Figure 3). The remaining segmentation steps were performed automatically.

Fig. 3.

Fig. 3

Open in a new tab

Here are three consecutive images showing the lateral meniscus horns arranged lateral to medial. In this case, the middle slice would be chosen as the end point because the meiscofemoral ligaments predominate in the next image.

After determining a seed point, each image underwent a preliminary thresholding operation to identify regions that should be excluded from the segmentation. First, a mask of pixels with intensities higher than 450 and lower than 120 was generated based on reliable exclusion of non-meniscus pixels determined based on a training set. Then, segmented regions whose areas were less than five pixels were removed, which allowed this mask to serve as a map of pixels that were not representative of meniscus (Figure 4).

Fig. 4.

Fig. 4

Open in a new tab

Image (a) is the original image, and (b) is a mask of high and low intensity pixels. This mask was used to determine pixels restricted from being included in the segmentation.

A second mask was created to identify bone, which was then dilated as an indirect method of roughly excluding many areas of articular cartilage from the segmentation (Figure 5). The tibia mask was dilated with a larger structuring element than the femur mask because of the larger separation between the tibia and the tibial cartilage than between the femur and its distal cartilage. Additionally, a range filter with a 3×3 neighborhood was used as an additional method of edge detection. A mask of pixels whose intensity values were greater than 120 was created from the filtered image (Figure 6). Similar to the intensity and bone constraints, these pixels were excluded from the segmentation because they represent boundaries between meniscus and cartilage.

Fig. 5.

Fig. 5

Open in a new tab

Image (a) shows a mask of high intensity areas. Image (b) isolates the femur, which is dilated in image (c) to include cartilage in the mask as an indirect way to exclude it from meniscus segmentation. The same process is done for the tibia.

Fig. 6.

Fig. 6

Open in a new tab

Image (a) shows the original image before filtering. Image (b) is the result of a range filter with a 3×3 neighborhood. Image (c) is a mask of pixels having an intensity value greater than 120 made from the range filtered image. The masked pixels on image (c) are excluded from segmentation.

Since the meniscus maintains a similar size and shape from slice to slice, size and shape constraints based on the previous meniscus segment were also imposed. The meniscus segmentation was prevented from getting more than 30% larger than the previous segmentation because the meniscus tended to decrease in area from superficial to deep, and was limited to a dilated mask of the previous slice.

Phase II: Threshold Determination

The algorithm automatically determines a threshold value by interrogating the intensity distribution surrounding the seed point within a region of interest (ROI) with a 20×20 pixel area. Four different regions with Gaussian intensity distribution within these areas were observed: 1) dark areas representing meniscus, 2) high intensity areas within the meniscus, 3) cartilage areas, and 4) bone. This task was complicated by the fact that cartilage and high intensity areas within the meniscus often had similar intensity values.

Four Gaussian curves were fitted to the histogram of the ROI around the seed point using the least squares method, with each fitted curve representative of a single pixel type (Figure 7). The fitted Gaussian distributions were used for determining an adaptive threshold level, allowing for variation between patients and images. Using the fitted histogram, two threshold candidates, Threshold1 and, Threshold2 were experimentally determined based on the means (μ)and standard deviations (σ) of the second and third curves, with the smaller of the two threshold candidates being chosen (Equations 1–2).

Fig. 7.

Fig. 7

Open in a new tab

(a) A dilation of a sample segmentation result, (b) with corresponding histogram and fitted Gaussian curves of the region used for threshold determination.

Threshold1=μ3−(3∗σ3) (1)
Threshold2=μ2−(1.25∗σ2) (2)

Threshold2, was only calculated if the difference between Threshold1 and μ2 was less 100, indicating that there was significant overlap of intensity values between the high intensity meniscus and cartilage curves. Due to natural variations in meniscus tissue, a wide range of threshold values may be calculated that would lead to inaccurate segmentations. To reduce these effects, the threshold values were further restricted to the range [310 – 400], which effectively excluded the cartilage regions from the segmentation.

Phase III: Initial Segmentation

Next, a basic thresholding operation was applied to the slice under consideration, where pixels above the calculated threshold were excluded. The area of thresholding was limited to the area encompassed by a dilation of the previous segmented slice using a 13×19 rectangular structuring element. This structuring element limited the segmentation area to only those areas where the meniscus was expected. Initial segmentation was followed by a 1 pixel morphological dilation of the generated mask, filling of holes, and removal of small areas of unconnected pixels. These morphological operations were done to connect fragmented areas of the meniscus for use in the next phase of segmentation (Figure 8). Notice that the thin layer of meniscus between the anterior and posterior horns is excluded in Figure 8(b) due to the morphological operations. This is an inevitable consequence of having a conservative segmentation approach to reduce the risk of segmenting cartilage.

Fig. 8.

Fig. 8

Open in a new tab

Image (a) shows a mask after thresholding and (b) is the mask after morphological operations. Note the high intensity areas within the meniscus (a) which were not included in the thresholding operation. OA participants tended to have a higher proportion of these areas than those of non-exposed participants.

Phase IV: Conditional Morphological Dilation

The initial threshold segmentation acquired from phase III was used as a starting point for the conditional morphological dilation. The initial segmented area was dilated by one pixel with a 3×3 square structuring element, and the intensity of each new pixel was tested. The tested pixel was included in the final segmentation if its intensity was below the threshold calculated in Phase II. After each new pixel was tested, the new segmented area was dilated again by one pixel to repeat the cycle, and was repeated until no new pixels were added (Figure 9). After every fourth iteration, the region was morphologically closed to fill gaps. The operation was also stopped if certain anatomical restrictions were met, namely the cross-sectional area becoming larger than the previous segment by 30%, the area reaching a pre-set maximum, or the operation exceeding 20 iterations. The pre-set maximum area was determined experimentally based on manually segmented regions, and was set at 1100 pixels with an exception for the first and second slices which was set at 1500 pixels since the meniscus tended to be larger in those slices.

Fig. 9.

Fig. 9

Open in a new tab

The same mask shown in Fig. 7.b after the growth phase was completed. The edges of the meniscus are now better defined.

Due to overlap of the Gaussian curves between high intensity values within the meniscus, and high intensity values representative of cartilage, a conservative threshold value was calculated which prevented some meniscus pixels from being included in the segmentation. This situation most often occurred when the region growing phase ended prematurely because of its small size and presence of high intensity areas within the meniscus. To overcome this problem, the calculated threshold was flexible during the conditional dilation phase. Using anatomic information gathered from manually segmented cases, it was found that the meniscus was never smaller than 105 pixels and that the horns of each meniscus rarely reduced in area by more than 25% from one slice to the next. Using this information, if the segmentation was stopped before these criteria were met, the threshold was temporarily increased by 10% until the bordering pixels were included to continue the region growing phase of the segmentation. This variable threshold ensured that the segmentation was not stopped prematurely, and that high intensity areas within the meniscus were included.

Phase V: Post-Processing

Two morphological post-processing steps were used to further improve the segmentation accuracy. First, all pixels that fell within the intensity and range constraints outlined in Phase I were removed from the segmentation. Second, pixels that were not on the previous segmented slice but were included on the current slice had their intensities tested again with a decreased threshold to exclude pixels that were more likely to be included incorrectly. Similarly, pixels that were included on the previous segment but had not been included on the current segmentation had their intensities retested with a higher threshold to increase the likelihood of acceptance. This second part of these post-processing steps was only performed on slices where the meniscus horns were separated since from that point on, the location of the meniscus from slice-to-slice was less variable. The final segmentation was completed after one additional growth iteration to include low intensity pixels that had not been included before, possibly due to incomplete conditional dilation or imposed range constraints.

After the segmentation was completed for the image, the next image in the sequence underwent the same segmentation phases, beginning at the constraint determination imposed in Phase I. The histogram used in Phase II for threshold calculation was generated based on a dilation of the completed segment of the previous slice, and a new threshold was calculated for the anterior and posterior horns separately. Throughout each phase of the segmentation, the anterior and posterior horns were treated separately due to variations in pathology and vascularization9, 27, allowing for the threshold and conditional dilation restrictions to be unique for each horn. An example of the final meniscus segmentation is shown in Figure 10, and the ZSIcan be compared to the mean manual intra-reader variability seen in normal knees of 0.86 (Table I), indicating a similar degree of discrepancy between manual raters.

Fig. 10.

Fig. 10

Open in a new tab

The green outline is a segment completed by a human reader, and the yellow outline is the segmentation completed by the semi-automated algorithm. Note the high degree of similarity (0.91), even when the posterior horn of the meniscus has high intensity areas throughout.

Table I.

Intra-Reader Variability

OARSI (OAI Subcohort)
Reader 0 (Control) 0 (Control) 0 (Progression) 1 (Progression) 2 (Progression) Mean
Semi-Automated Segmentation .98 1.0 .99 1.0 .99 .99
Manual 1 0.84 0.85 0.90 0.88 0.86 0.87
Manual 2 0.85 0.88 0.69 0.82 0.71 0.79
Manual 3 0.88 0.88 0.86 0.84 0.88 0.87
Manual 4 0.83 0.86 0.78 0.84 0.76 0.81
Manual 5 0.88 0.83 0.84 0.85 0.81 0.84
Open in a new tab

VALIDATION

Semi-automated segmentations were quantitatively compared to those of manual raters according to the similarity index described by Zijdenbos (ZSI)28–29.

S=2[A1∩A2][A1+A2] (3)

A1 and A2 represented all pixels included in a segmented region, and S was defined as twice the area of overlapping pixels divided by the added area of both segmented regions. S varies between 0 and 1, where 1 indicates complete agreement, and 0 indicates regions that do not overlap. This measure of similarity was dependent on degree of overlap between the two regions and location of the two regions, giving a higher score if two regions shared similar centers27. According to Zijdenbos, a similarity index above 0.7 provided “excellent agreement”, and 0.6 indicated “substantial agreement” between two regions29.

In order to quantitatively measure the accuracy of the semi-automated segmentation results, five readers were trained to manually segment the meniscus using “Image-J”, an image viewing and segmenting program from the NIH. Each reader outlined the lateral meniscus of two healthy and twelve progression OA cases of varying severity. These segmentations were then compared to those of each other rater using the ZSI in Equation 3. Each of five raters’ segmentations had ZSI calculated in comparison to the rest of the raters and the semi-automated segmentations. Due to varying interpretations between raters about the lateral and medial extent of the meniscus, only slices where all raters agreed that meniscus was present were compared. The intent of this study was to determine the validity and viability of semi-automated segmentations, so the analysis was focused on the performance compared to manual raters. Slices where manual raters disagreed would factor in a ZSI of 0, which would inappropriately skew the comparison results.

Results

Images from 10 normal knees and 14 knees with OA were analyzed in this study. The average age of participants was 54 years; approximately 38% were female and 80% were white (Table II).

Table II.

Participant demographics

Lateral JSN OARSI Grade Cohort Sex Race Age BMI
0 Normal F African American 47 22.3
0 Normal F Caucasian 53 20.6
0 Normal F Caucasian 54 20.1
0 Normal M Asian 55 24.6
0 Normal M Caucasian 55 27.0
0 Normal M Caucasian 55 28.3
0 Normal M African American 56 25.7
0 Normal M Caucasian 57 24.1
0 Normal F Caucasian 70 21.4
0 Normal M Caucasian 72 20.8
0 Progression M African American 46 27.2
0 Progression M Caucasian 48 25.9
0 Progression M Caucasian 52 33.6
0 Progression F 55 27.8
1 Progression F Caucasian 51 35.9
1 Progression F African American 51 43.0
1 Progression F 59 30.9
1 Progression M Caucasian 62 34.9
2 Progression M Caucasian 46 24.8
2 Progression M Caucasian 47 28.8
2 Progression M 51 26.7
2 Progression F Caucasian 52 20.9
3 Progression M Caucasian 48 26.5
3 Progression M Caucasian 55 27.3
Open in a new tab

Similarity data of the manual and automatic segmentations of the lateral meniscus showed that the semi-automated segmentation method performs similarly to manual raters for normal and moderately degenerated menisci (Table III). Rater 1 marked eight additional non-exposed control cases, all with a ZSI that yielded excellent correlation (similarity index > 0.70) with manual raters. A visual example of average results is seen in Figure 11. All five manual raters were unable to identify a meniscus in OARSI grade 3 participants, therefore these images were excluded.

Table III.

Overall performance of semi-automatic segmentations on non-exposed and OA progression cases compared to a manual rater. No manual rater was able to identify meniscus in the OARSI 3 participants

Mean Similarity Index (Standard deviation) Standard Deviation
Non-Exposed Control (n=10) .80 .06
Progression Subcohort(n=12) .69 .12
OARSI JSN grade = 0 (n=4) .75 .04
OARSI JSN grade =1 (n=4) .67 .12
OARSI JSN grade = 2 (n=4) .64 .16
OARSI JSN grade = 3 (n=2) N/A N/A
Open in a new tab

Fig. 11.

Fig. 11

Open in a new tab

This is an automatic segmentation (yellow) with a similarity index of .79 compared to a manual rater (green) showing the average level of agreement for normal knees(see Table III).

The automatic segmentation had an average similarity index 0.80 when used on control participants, and had an average similarity of 0.75, 0.67, and 0.64 when looking at progression patients with OARSI JSN of 0, 1, and 2, respectively. The algorithm, including manual input, was able to produce segmentation results in 2–4 minutes per case for both normal and progression subjects, while manual segmentations consistently took 7–20 minutes per case, depending on the speed of the reader and the familiarity with the segmentation program. The running time for the software was not optimized for speed, which can bring further reductions in time without sacrificing accuracy.

Intra-reader variability was assessed by examining the segmentations of each manual rater to previous segmentations of the same images from the same rater. The semi-automated method was assessed for intra-reader variability by choosing 100 random seed points within the meniscus, and comparing each segmentation result to the previous result. Semi-automated intra-reader segmentations yielded an average ZSI of 0.99, while the manual segmentations averaging 0.84 (Table I).

Areas with low cartilage intensity posed a challenge for the segmentation method and reduced the accuracy in these instances (Figure 12). An example of an entire lateral meniscus series is shown in Figure 13, with corresponding data on how the segments improved based on comparisons with and without Phase V post segmentation processing compared to a single manual rater (Table IV). The average improvement was 1.1% per segmentation, with an overall improvement of 0.94%.

Fig. 12.

Fig. 12

Open in a new tab

This is a more extreme example of a segmentation error where incorrect areas were segmented due to a low intensity area in the cartilage. The yellow outline is the automatic segmentation, and the green outline is a segmentation performed manually. The error is limited due to area and intensity constraints, and has a similarity index of 0.66.

Fig. 13.

Fig. 13

Open in a new tab

A series showing the manual segmentations (green) and automatic segmentations (yellow) of the lateral meniscus from a patient with OA progression.

Table IV.

Similarity index improvement due to Phase V

Overall Similarity Index Improvement
Case 1 (OARSI 0) −0.01
Case 2 (OARSI 0) 0.01
Case 4 (OARSI 0) 0.01
Case 5 (OARSI 1) 0.02
Case 3 (OARSI 2) 0.02
Average 0.01
Open in a new tab

Discussion

A more complete understanding of important questions such as natural history and risk factors for OA should result from additional knowledge about the meniscus and its role in this disease. A study performed by Link et al. compared findings on medical images to K-L score, and noted that meniscal lesions were present in 46 out of 50 OA patients, and that 4 patients without lesions had low K-L scores (1 or 2). Knees with a K-L score of 4 had severe meniscal lesions characterized as complex tears with deformity or severe destruction22. In addition, the tibiofemoral contact area is correlated with the severity of OA, and the meniscus plays a large role in joint mechanics30. Thus, it is inferred that the health of the meniscus may play a role in the progression of OA, however definitive studies addressing this important issue are limited. Quantifying the changes of the meniscus as OA progresses (as seen in Figure 1) has the potential to establish disease severity and predict disease trend. Herein, we report on a semi-automated segmentation method that demonstrated acceptable accuracy, intra-reader variability, and efficiency in characterizing the meniscus in healthy and moderately osteoarthritic knees, which should prove helpful in addressing these and other important questions from longitudinal studies such as the OAI (Figure 13). Although the sample size is limited, this study determined that meniscus segmentation is feasible and indeed can be very accurate.

The usefulness of computer assisted segmentation methods can be appreciated when examining their use in cartilage segmentation. Several semi-automated and automated methods have been developed to enhance quantitative analysis of the knee articular cartilage because of their speed, reliability, accuracy, and decreased inter-reader variability31. One study reported that automated segmentation results in improved inter-reader variability compared with two manual readers when measuring number of segmented pixels, maximal thickness, and volume31. These improvements over manual segmentations offer promise for enhanced objective measurement of natural history, osteoarthritic cartilage changes, and the effects of therapeutic interventions.

The same advantages achieved by computer-assisted segmentation of the cartilage were addressed for the meniscus in the current study. Our semi-automated approach overcomes many segmentation challenges through the use of carefully designed image intensity and anatomical constraints, shown by its consistently high similarity index for normal and pathologic knees compared against five manual raters. Therefore, our method shows particular promise with assessment of normal menisci and those with mild to moderate OA, making it ideal for use in assessing OA development and progression. The use of a single semi-automated approach to a large dataset should greatly reduce effects of intra-reader variability, and eliminate inter-reader variability compared to manual segmentations. As Table I indicates, manual segmentations are subject to inevitable intra-reader variation, a limitation which is largely overcome with semi-automated segmentation. An important advantage of a semi-automated segmentation method is that the determination of borders is unbiased and consistent. Through observation of manual segmentations in the current study, it was noted that a manual rater can have significantly different determinations of a meniscus border when given the same segmentation case at a different time. This intra-reader variability is virtually eliminated with our semi-automated segmentation method, as demonstrated by the average similarity index score of 0.99 (Table I). The key to reliable image data analysis is a fast, consistent, and reliable method of segmentation, which is best achieved through automated segmentation methods. The degree of similarity and the speed offered by this segmentation method allows for rapid detection and quantification of meniscus properties without the burden of excessive manual time and intra-reader variability, which should provide the ability to further our understanding of the meniscus and its role in OA.

Semi-Automated segmentation is not without its limitations. Our algorithm does not perform well in patients with severe OA or and likely in those with acute meniscal tears. This method is best used as a tool to facilitate analysis of the meniscus in large studies since it performs well in most cases and works faster and without intra-reader variability when compared to manual segmentation. We recommend that in cases where it does not perform well that manually segmentation by experienced raters be employed since these cases are often prone to the greatest degree of manual inter-rater variability.

One way to enhance our understanding of the relationship between meniscus degradation and OA is to examine the tissue’s volume change as OA progresses. Recently, the volume of the meniscus has been calculated using cross-sectional area from each slice in the MR image sequence and the distance between slices32, an approach that can be easily applied using segmentations. The automatic segmentations produced in the present study demonstrated a cross-sectional area similar to that of manual raters, suggesting the potential of this approach in calculating tissue volumes. The automated segmentations consistently produced a high degree of similarity to manual segmentations, making our method suitable for quantitative analysis of the meniscus. As the algorithm improves to incorporate more complex menisci and more degenerated knees, it may be necessary to employ the segmentation expertise of MSK radiologists for the evaluation of accuracy. A focus of our group is the implementation of these methods to the OAI dataset to determine the role of the meniscus as an indicator of OA progression or predictor of disease course, making it a potential biomarker for disease progression. Future efforts will expand the algorithm to include the medial meniscus in order to provide a more comprehensive examination of the knee. Given the more complicated structure of the lateral meniscus, segmentation should also be feasible for the medial meniscus using similar methodology.

Acknowledgments

Sources of support: This project was funded in part through the NIH Roadmap Training Program in Clinical Research. Also, the OAI is a public-private partnership comprised of five contracts funded by the National Institutes of Health. Private funding partners include Merck Research Laboratories; Novartis Pharmaceuticals Corporation, GlaxoSmithKline; and Pfizer, Inc.

This project was supported in part by Award Number UL1RR025755 and TL1RR025753(MSS) from the National Center For Research Resources and the Award Number R01LM010119 from the National Library of Medicine. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Library of Medicine, or the National Institutes of Health. The OAI is a public-private partnership comprised of five contracts (N01-AR-2-2258; N01-AR-2-2259; N01-AR-2-2260; N01-AR-2-2261; N01-AR-2-2262) funded by the National Institutes of Health, a branch of the Department of Health and Human Services, and conducted by the OAI Study Investigators. Private funding partners include Merck Research Laboratories; Novartis Pharmaceuticals Corporation, GlaxoSmithKline; and Pfizer, Inc. Private sector funding for the OAI is managed by the Foundation for the National Institutes of Health. This manuscript has received the approval of the OAI Publications Committee based on a review of its scientific content and data interpretation.

Footnotes

Conflict of interest statement

The authors have no conflicts of interest to report.

References

  • 1.Urwin M, Symmons D, Allison T, Brammah T, Busby H, Roxby M, et al. Estimating the burden of musculoskeletal disorders in the community: the comparative prevalence of symptoms at different anatomical sites and the relation to social deprivation. Annals of Rheumatic Diseases. 1998;57:649–55. doi: 10.1136/ard.57.11.649. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Peat G, McCarney R, Croft P. Knee pain and osteoarthritis in older adults: a review of community burden and current use of primary health care. Annals of Rheumatic Diseases. 2001;60:91–97. doi: 10.1136/ard.60.2.91. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Felson DT, Zhang Y. An update on the epidemiology of knee and hip osteoarthritis with a view to prevention. Arthritis Rheum. 1998;41(8):1343–55. doi: 10.1002/1529-0131(199808)41:8<1343::AID-ART3>3.0.CO;2-9. [DOI] [PubMed] [Google Scholar]
  • 4.Emrani PS, Katz JN, Kessler CL, Reichmann WM, Wright EA, McAlindon TE, et al. Joint space narrowing and Kellgren-Lawrence progression in knee osteoarthritis: an analytic literature synthesis. Osteoarthritis Cartilage. 2008 August;16:873–82. doi: 10.1016/j.joca.2007.12.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Cicuttini FM, Wluka AE, Stuckey SL. Tibial and femoral cartilage changes in knee osteoarthritis. Annals of the Rheumatic Diseases. 2001 Oct;60:977–80. doi: 10.1136/ard.60.10.977. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Davis KW, Tuite MJ. MR Imaging of the Postoperative Meniscus of the Knee. Seminars in Musculoskeletal Radiology. 2002;6:35–45. doi: 10.1055/s-2002-23162. [DOI] [PubMed] [Google Scholar]
  • 7.Krause WR, Pope MH, Johnson RJ, Wilder DG. Mechanical changes in the knee after meniscectomy. J Bone Joint Surg. 1976;58-A:599–604. [PubMed] [Google Scholar]
  • 8.Newman AP, Daniels AU, Burks RT. Principles and decision making in meniscal surgery. Arthroscopy. 1993;9:33–51. doi: 10.1016/s0749-8063(05)80342-1. [DOI] [PubMed] [Google Scholar]
  • 9.Grood ES. Meniscal function. Adv Orthop Surg. 1984;12:193–197. [Google Scholar]
  • 10.Ahmed AM, Burke DL. In-vitro measurement of static pressure distribution in synovial joints – Part I: Tibial surface of the knee. Journal of Biomechanical Engineering. 1983;105:216–25. doi: 10.1115/1.3138409. [DOI] [PubMed] [Google Scholar]
  • 11.Donahue TLH, Hull ML, Rashid MM, Jacobs CR. A nite element model of the human knee joint for the study of tibio–femoral contact. Journal of Biomechanical Engineering. 2002;124:273–280. doi: 10.1115/1.1470171. [DOI] [PubMed] [Google Scholar]
  • 12.Donahue TLH, Hull ML, Rashid MM, Jacobs CR. The sensitivity of tibiofemoral contact pressure to the size and shape of the lateral and medial menisci. Journal of Orthopaedic Research. 2004;22:807–14. doi: 10.1016/j.orthres.2003.12.010. [DOI] [PubMed] [Google Scholar]
  • 13.Donahue TLH, Hull ML, Rashid MM, Jacobs CR. How the stiffness of meniscal attachments and meniscal material properties affect tibio-femoral contact pressure computed using a validated finite element model of the human knee joint. Journal of Biomechanics. 2003;36:19–34. doi: 10.1016/s0021-9290(02)00305-6. [DOI] [PubMed] [Google Scholar]
  • 14.Huang A, Hull ML, Howell SM. The level of compressive load affects conclusions from statistical analyses to determine whether lateral meniscal autograft restores tibial contact pressure to normal: a study in human cadaveric knees. Journal of Orthopaedic Research. 2003:21. doi: 10.1016/S0736-0266(02)00201-2. [DOI] [PubMed] [Google Scholar]
  • 15.Huang A, Hull ML, Howell SM, Donahue TH. Identification of cross-sectional parameters of lateral meniscal allografts that predict tibial contact pressure in human cadaveric knees. Journal of Biomechanical Engineering. 2002;124:459–64. doi: 10.1115/1.1503061. [DOI] [PubMed] [Google Scholar]
  • 16.Wilson W, van Rietbergen B, van Donkelaar CC, Huiskes R. Pathways of load-induced cartilage damage causing cartilage degeneration in the knee after meniscectomy. Journal of Biomechanics. 2003;36:845–851. doi: 10.1016/s0021-9290(03)00004-6. [DOI] [PubMed] [Google Scholar]
  • 17.Englund M, Lohmander LS. Risk factors for symptomatic knee osteoarthritis fifteen to twenty-two years after meniscectomy. Arthritis and Rheumatism. 2004;50:2811–19. doi: 10.1002/art.20489. [DOI] [PubMed] [Google Scholar]
  • 18.McNicholas MJ, Rowley DI, McGurty D, Adalberth T, Abdon P, Lindstrand A, et al. Total meniscectomy in adolescence: a thirty year follow-up. Journal of Bone and Joint Surgery 200. British Volume 82:217–221. [PubMed] [Google Scholar]
  • 19.Roos H, Lauren M, Adalberth T, Roos EM, Jonsson K, Lohmander LS. Knee osteoarthritis after meniscectomy: prevalence of radiographic changes after twenty-one years, compared with matched controls. Arthritis and Rheumatism. 1998;41:687–93. doi: 10.1002/1529-0131(199804)41:4<687::AID-ART16>3.0.CO;2-2. [DOI] [PubMed] [Google Scholar]
  • 20.Esmaili Jah AA, Keyhani S, Zarei R, Moghaddam AK. Accuracy of MRI in comparison with clinical and arthroscopic finding in ligamentous and meniscal injuries of the knee. Acta Orthop. 2005;71:189–96. [PubMed] [Google Scholar]
  • 21.Wright DH, De Smet AA, Norris M. Bucket-handle tears of the medial and lateral menisci of the knee: value of MR imaging in detecting displaced fragments. American Journal of Roentgenology. 1995;165:621–25. doi: 10.2214/ajr.165.3.7645481. [DOI] [PubMed] [Google Scholar]
  • 22.Link TM, Steinbach LS, Ghosh S, Ries M, Lu Y, Lane N, et al. Osteoarthritis: MR imaging findings in different stages of disease and correlation with clinical findings. Radiology. 2003;226:373–81. doi: 10.1148/radiol.2262012190. [DOI] [PubMed] [Google Scholar]
  • 23.Nikolaou VS, Chronopoulos E, Savvidou C, Plessas S, Giannoudis P, Efstathopoulos N, et al. MRI efficacy in diagnosing internal lesions of the knee: a retrospective analysis. J Trauma Manag Outcomes. 2008 Jun;2:4. doi: 10.1186/1752-2897-2-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Osteoarthritis initiative a knee health study. San Francisco: University of California; [Internet] [updated 2008 Apr 15; cited 2008 Aug 26]. Available from: http://www.oai.ucsf.edu/datarelease/ [Google Scholar]
  • 25.Peterfy CG, Schneider E, Nevitt M. The osteoarthritis initiative: report on the design rationale for the magnetic resonance imaging protocol for the knee. Osteoarthritis and Cartilage. 2008;16:1433–1441. doi: 10.1016/j.joca.2008.06.016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Gonzalez R, Woods R. Digital Image Processing. Prentice Hall; 2007. [Google Scholar]
  • 27.Arnoczky SP, Warren RF. Microvasculature of the human meniscus. The American Journal of Sports Medicine. 1982;10:90–95. doi: 10.1177/036354658201000205. [DOI] [PubMed] [Google Scholar]
  • 28.Atkins MS, Mackiewich BT. Fully Automatic Segmentation of the Brain in MRI. IEEE Trans Med Imag. 1998 Feb;17:98–107. doi: 10.1109/42.668699. [DOI] [PubMed] [Google Scholar]
  • 29.Zijdenbos AP, Dawant BM, Margolin RA, Palmer AC. Morphometric analysis of white matter lesions in MR images: Method and validation. IEEE Trans Med Imag. 1994 Dec;13:716–24. doi: 10.1109/42.363096. [DOI] [PubMed] [Google Scholar]
  • 30.Scarvell, et al. Magnetic Resonance Imaging Analysis of Kinematics in Osteoarthritic Knees. The Journal of Arthroplasty. 2007;22(3) doi: 10.1016/j.arth.2006.06.006. [DOI] [PubMed] [Google Scholar]
  • 31.Stammberger T, Eckstein F, Michaelis M, Englmeier KH, Reiser M. Interobserver reproducibility of quantitative cartilage measurements: Comparison of B-spline snakes and manual segmentation. Magnetic Resonance Imaging. 1999;17:1033–42. doi: 10.1016/s0730-725x(99)00040-5. [DOI] [PubMed] [Google Scholar]
  • 32.Bowers ME, Tung GA, Fleming BC, Crisco JJ, Rey J. Quantification of meniscal volume by segmentation of 3 T magnetic resonance images. Journal of Biomechanics. 2007;40:2811–15. doi: 10.1016/j.jbiomech.2007.01.016. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES