2.1.

Knee MRI Collections

We used three collections of knee MRI from the Center for Clinical and Basic Research (CCBR), the OAI study, and the SKI10 challenge. The collections included scans with manual segmentations that we used for training and validation. The relevant ethical review boards approved the studies.

The collections substantially differed in scanner type, scan quality, and disease severity as described below. The scanner specifications and the analyzed anatomical structures are described in Table 1 and the populations are summarized in Table 2.

Table 1

Study descriptions in terms of scanners, sequences and available manual annotations.

Table 2

Cohort population characteristics for training and validation subpopulations from CCBR and OAI. The SKI10 collection includes 60 training, 40 test scans, and 50 validation scans, but the population characteristics are unavailable.

2.2.

Automatic Segmentation Framework

The proposed segmentation framework combines rigid multiatlas registration with voxel classification in a multistructure setting (see Fig. 2 for illustration of the workflow). The voxel classification step includes an ROI analysis and a feature selection for each structure. This is a generalization of the method proposed by Folkesson et al.¹⁴ that was implemented for the two medial cartilage compartments. Here, we generalized to arbitrary compartments and added the registration step to improve the performance.

Fig. 2

Flowchart of the quantification steps. The multiatlas registration, region of interest (ROI), and feature computation steps are performed both during training (top) and scoring (bottom). The training phase results in the transformations for the training scans, the ROI definitions, and the selected feature sets—these parameters are used during scoring of “unseen” scans.

2.3.

Training Data

We assumed the existence of $T$ training scans ${\mathrm{Im}}^t$, for which $N$ structures have been (manually) outlined, forming a set of ground-truth segmentations. An example scan with manual segmentation is illustrated in Fig. 3. These were represented by $TxN$ training volumes $T{V}_n^t$ with $t\in [1,T]$, $n\in [1,N]$, where each $\text{pixel}/\text{voxel}$ gave the presence of structure $n$. This allowed for binary segmentations with values $\{0,1\}$ or for partial volumes with values [0,1]. Overlapping structures were allowed. The scans were assumed to be of dimension $D$, where $D\in \{1,2,3\}$.

Fig. 3

Scan, manual segmentation, and automatic segmentation for a validation scan from the OAI collection. Three sagittal slices are shown going through the centers of the (a) tibial medial cartilage, (b) patellar cartilage, (c) tibial lateral cartilage. The scan slice (left) is given with the manual (center) and the automatic segmentations (right). The Dice volume overlaps for the example were between 0.84 and 0.91 for the tibial and femoral cartilages, 0.76 for the patellar cartilage, and 0.78 and 0.80 for the menisci.

2.4.

Multiatlas Rigid Registration

The aim of the registration step was to produce a transformation from a given scan to a training space center. As explained below, this allowed determination of the ROI for each anatomical structure and transformation of scan features to a common feature space. The effect of the inter-subject alignment from the registration step is illustrated in Fig. 4.

Fig. 4

The effect of multiatlas registration illustrated for scans in the CCBR training collection. The manual segmentations for the three structures (Tibia, medial tibial cartilage, and medial femoral cartilage) are accumulated in the three color channels of an RGB picture and shown in a slice in the middle of the medial compartment. (a) This is done before and (b) after registration. The intensity in each pixel can be interpreted as a spatial prior for each structure. (a) Before registration, these values ranged from 0 to 1.00, 0.34, and 0.28 for the three structures. (b) After registration, they ranged to 1.00, 0.86, and 0.73.

We represented a transformation as a $2xD+1$ vector including $D$ elements for translations, $D$ for Euler angle rotations, and 1 for isotropic scaling. By allowing scaling, the registration is not mathematically a rigid transformation, but a similarity transformation.

We registered two given scans by maximization of normalized mutual information²⁵ (NMI) using L-BFGS optimization.²⁶ The optimization was performed in two steps with Gaussian blurring of the scans at a coarse scale and then at a fine scale.

During training, each training scan ${\mathrm{Im}}^t$ was registered to all training scans ${\mathrm{Im}}^r$, $r\in [1,T]$, where the optimization was initialized by the translation implied by the center of masses (CoMs) of the manual segmentations of a specified structure. This initialization ensured a robust estimate for all training scans. From the resulting $T$ similarity transformations $R^r$, the transformation to a training space center was defined as the element-wise median. The median transformation for training scan ${\mathrm{Im}}^t$ was denoted $R^t$.

When segmenting a (new) scan $S$, the scan was registered to all training scans, but with no initialization, resulting in similarity transformations $S^t$. The compositions $S^{t}x{R}^t$ provide an estimate of the transformation to the training space center via a training scan. Due to inter-subject variation, scanner noise, and optimization failures, the $T$ estimates will not be identical; however, the element-wise median of the compositions $S^{t}x{R}^t$ defined a robust estimate of the final multiatlas transformation for the given scan to the training space center.

We experimented with more sophisticated definitions of the typical transformation than the element-wise median—specifically atlas selection based on the NMI score and with a geometric transformation based on Fletcher’s principal geodesic analysis.²⁷ However, none of these experiments improved the performance, therefore, we used the simplest approach. Note that since the angle coordinates represent rotations; they are typically close to zero (and definitely between $-\pi /2$ and $+\pi /2$) and, therefore, have no angle wrap-around challenges.

2.5.

Structure-Wise Regions of Interest

Both in terms of computational efficiency and classifier specificity, it is advantageous to only train and apply a classifier on the region surrounding the structure of interest. We learned a simple rectangular ROI in the registered scans for each structure. Figure 5 illustrates the resulting ROIs learned from the OAI training data.

Fig. 5

Visualization of the learned regions of interest (ROI) for the OAI training collection. For each structure, the coordinate extrema after multiatlas registration are added a margin to cover feature filter support. These resulting bounding boxes are shown from sagittal/coronal/axial views. Each structure is only trained and segmented inside the ROI.

Specifically, we defined a structured ROI as the coordinate extrema encountered in the registered training scans with an added margin of 5% of the scan’s size in each direction. Thus, the structures with margin for feature filter support were inside the ROI.

2.6.

Voxel Classification Features

The features used for the voxel classifier were very similar to the set previously used by Folkesson et al.¹⁴ As potential features, we used the N-jet²⁸ including Gaussian derivatives up to order 3, and nonlinear combinations of these including the Hessian and structure tensor eigenvectors and values. Also, intensity and position were included. All nonposition features were computed at scales of 0.6, 1.2, and 2.4 mm. This gave a 178-dimensional space of potential features.

The position and Gaussian derivative features for a given scan were transformed using the similarity transform from the multiatlas registration step. Each individual feature was translated and scaled such that the values for the training collection had mean 0 and standard deviation 1.

2.7.

Training Voxel Sampling

A large set of training voxels are desired to ensure classifier generalizability. However, for practical computations, the computer memory is limiting the training size. Also, it is desirable to sample the training voxels to be specific to the classification task at hand. Particularly with small/thin structures like cartilages, it is desirable to include many structure voxels, relatively many near-structure background voxels, and fewer voxels from the background far away from the structure in the training scans.

We applied a simple sampling scheme with a constant sampling density inside each structure, another density in the background and a linear transition in density in a narrow margin around each structure. The exact sampling densities were calculated such that the training data for each structure would be suitable for execution on a standard laptop computer.

Each voxel $x$ from the training scan ${\mathrm{Im}}^t$ included in the training data was attributed with the local sampling density $\mathrm{samp}(x,{\mathrm{Im}}^t)\in [\mathrm{0,1}]$.

2.8.

Feature Selection

We used floating forward feature selection²⁹ optimized on the sampled training voxels for defining the actual set of features used for each structure. The Dice overlap on the sampled training voxels was used as an optimization criterion. With the sampled training voxels, the overlap contribution must be inversely weighted with the local sampling density. For a given feature set and a given hard classifier $\text{class}(x,{\mathrm{Im}}^t,n)$ returning 0 or 1 indicating membership for structure $n$, with sampled training voxels $x_i$:

The overlap was evaluated using leave-one-scan-out such that the evaluation of sampled feature voxels from a given scan only included training voxels from the other scans. To encourage sparse feature sets, a penalty of 0.001 per feature was added to the Dice overlap in the feature selection objective function. This avoids the addition of several additional features with only marginal improvement on the training collection—therefore, the generalization to the validation set was potentially improved. Typically, the feature set for each knee structure classifier became 10 to 30 dimensional.

2.9.

Structure-Wise Classification

The above feature selection and sampling strategies can be used for any classifier. Like Folkesson et al.,¹⁴ we applied a $k$-NN classifier to allow for multimodal distributions with no assumptions on the feature distributions that will arise from the different structures. Other nonlinear classifiers such as support vector machines, random forest, or convolutional neural networks could also be appropriate and are likely computationally more efficient.

For $k$-NN to become a hard classifier, a threshold on the number of neighbors is needed. For each structure, this threshold ($k_n$) was determined during training to optimize the Dice overlap score for a given feature set. The classification strength was defined as the $k$-NN count minus $k_n$ such that

2.10.

Multistructure Segmentation of a New Scan

Above, we trained $n$ independent structure-wise classifiers to allow dedicated feature selection for each structure and simple generalization to multiple structures. When segmenting an unseen scan, we first performed the multiatlas registration step resulting in the atlas transformation (element-wise median of $S^{t}x{R}^t$). This transformation was applied to each trained structure ROI, defining a bounding box for each structure in the scan. The trained, structure-specific features were computed within the ROIs and we applied each structure-classifier independently.

For computational efficiency, we used the sparse sample-expand and sample-surround sparse classification methods.³⁰ These methods essentially aim to only classify voxels that are relevant for the segmentation in order to improve computational efficiency. The sample-expand algorithm is a region-growing algorithm that is initiated in many randomly defined seed voxels in the scan. For each seed, the structure classifier is used and the region is grown if the classifier determines the voxel to be inside the structure. This growing continues to neighbor voxels in a breadth-first manner until a connected component is formed. The number of seed voxels is defined to ensure that all components of a suitable size are hit by a seed point (i.e., the probability of missing a component is below ${10}^{-10}$). This sample-expand region growing is very effective for small/thin structures since the majority of the scan voxels are not classified. For large, compact structures, the sample-surround algorithm does not expand in all directions, but rather makes a line search to the object boundary and then only classifies voxels necessary to outline the structure boundary. We apply sample-expand sparse classification for cartilages/menisci and sample surround for the bones.

For each structure, the one-versus-all $k$-NN classifier resulted in classification strengths in the classified voxels. The voxels that were not visited by the sparse classification methods were assumed to have strength -1. These structure-wise strength maps were combined to form a single map of class labels by assigning each voxel to the structure with the highest strength. This was followed by a largest connected components step including the largest components to exclude small, spurious components.

2.11.

Experimental Setup

Given the annotated training data, the above framework is fully automatic. The following set of fixed parameters was used for all experiments.

The rigid registration was optimized at blurring scales 5 mm and then 1 mm using 64 histograms bins in the NMI computation. NMI was computed using a regular grid of evaluation points with 2-mm spacing, except the 15% outer margin of the scans was ignored to avoid nonoverlap artifacts. The voxel classification features were computed at blurring scales 0.6, 1.2, and 2.4 mm.

During feature selection, the training voxel sampling was performed with densities (as outlined above) such that the training data for each structure approximated 250 MB assuming a resulting 30-dimensional feature vector. The narrow margin around each structure was set to 4 mm (approximately double the thickness of a cartilage sheet).

The $k$-NN classifier was used with $k=100$ and approximation parameter $ϵ=2$ in the approximate nearest neighbour implementation.³¹

The largest connected components step was performed to include components with a volume above 15% of the volume of the largest component.

To investigate the effect of the multiatlas registration step, the entire framework was evaluated without this step. Thereby, the framework becomes essentially equivalent to the original Folkesson framework.¹⁴ Unlike Folkesson, the nonregistration framework still allows an arbitrary set of structures and there are implementation differences, but the basic methodology is very similar. We denote this variant Folkesson07.

2.12.

Statistical Analysis

Segmentation accuracy was evaluated by the Dice volume overlap score where segmentations were available. Where only volume scores were available, accuracy was evaluated by the Pearson linear correlation coefficient $r$, the absolute agreement intra-class correlation coefficient, and by the median of signed, relative differences (in %).

Segmentation precision was evaluated by the root-mean-squared coefficient of variation (RMS CV) of the cartilage volume pairs. This was done both on repeated segmentations of the same scan (intra-scan) and on segmentations of repeated scans from the same subject acquired within a short period (inter-scan).

For the SKI10 collection, we used the evaluation metrics defined by SKI10 that combines surface distances measures into a bone score and volume overlap/difference measures into a cartilage score.¹⁶ The bone and cartilage scores are both normalized by comparison to inter-reader variation such that a score of 75 matches the performance of the SKI10 expert manual readers.

3.1.

Multiatlas Registration Effects

As illustrated in Fig. 4, the registration step makes the positioning of the structures more consistent. For each structure, this is quantified by the CoM position variation (PV) defined as the mean distance from each CoM to the training collection mean CoM. For the CCBR collection, the cartilage compartment PVs were 8.4 to 9.2 mm before registration and 1.8 to 4.0 mm after registration. For the OAI collection, the cartilage PVs were 8.0 to 9.3 mm before and 2.7 to 5.1 mm after. However, here the tibio-femoral compartments were all close to 3 mm, whereas the patellar was higher at 5.1 mm. Finally, for the SKI10 collection, the cartilage PVs were 7.3 to 10.9 mm before and 3.5 to 4.4 mm after.

The ROIs illustrated in Fig. 5 also illustrate a registration effect. For the CCBR collection, the volumes of the tibial and femoral cartilage ROIs were 9%, and 32% of the full scan corresponding to 61 and 79 times the mean volume for each structure. In comparison, for the Folkesson variant without registration, the cartilages ROIs were 23% and 38% of the scan and 147 and 87 times the structure volume. For the OAI collection, the ROI relative volumes were 5% to 6% for tibial/patellar and 14% to 15% for femoral cartilages—corresponding to 45–56 times the structure volumes.

3.2.

Segmentation Accuracy

The training and validation segmentations accuracies are shown in Table 3 for the CCBR and OAI collections. For medial/lateral tibial/femoral cartilages, the validation volume overlaps were between 0.804 and 0.866. For patellar cartilage and menisci, they were a bit lower.

Table 3

Segmentation accuracy for training and validation in the Center for Clinical and Basic Research (CCBR) and osteoarthritis initiative (OAI) collections. Accuracy is given as the Dice volume overlap showing mean±std over minimum/maximum values. The number of knees (n) is given for each set.

For comparison, for the CCBR study the radiologist repeated the manual segmentations for the 38 scans also used for the scan-rescan precision analysis. The intra-operator, intra-scan mean Dice volume overlaps for the medial tibial and femoral cartilages were 0.858 and 0.860.

3.3.

Cartilage Volume Precision

The intra-scan and inter-scan precision for the CCBR collection is given in Table 4. The radiologist had an intra-scan precision 6% to 7% and a can-rescan precision at 9%. The automatic method had an intra-scan precision of 0%, demonstrating that repeated runs on the same scan reproduced the volume scores, and a scan-rescan precision at 5%. When merging the segmentations for the two medial compartments, the precision was considerably improved for the radiologist, but only marginally for the automatic method.

Table 4

Precision of cartilage volumes for the 38 scan-rescan knees from the CCBR collection. The intra-scan results are for repeated segmentations on the same scan. The scan-rescan results are for segmentations of different scans acquired approximately 1 week apart. Precision is given as root-mean-squared coefficient of variation (RMS CV).

3.4.

Cartilage Volume Comparisons on OAI

Both VirtualScopics and Chondrometrics have cartilage volumes available via the OAI, quantified for 150 and 1436 knees, respectively. For the interested reader, the Biomediq cartilage volume scores are available for the same knees at the Biomediq website.³²

Table 5 shows the level of agreement between the KneeIQ quantifications and the alternative methods. Due to differences in subcompartment definitions for the femoral cartilage, only the medial/lateral tibial compartment volumes are included.

Table 5

Agreement between the automatic knee image quantification (KneeIQ) volume scores and the scores from the independent groups on different subset from the OAI cohort. The table includes number of knees available (n), Pearson linear correlation coefficient (r), absolute agreement intra-class correlation coefficient (ICC) and median signed relative difference (%).

The agreement between KneeIQ and Chondrometrics cartilage volumes is further illustrated by the Bland–Altman plots in Fig. 6.

Fig. 6

Bland–Altman plots of the agreement between manual Chondrometrics and automatic knee image quantification (KneeIQ) segmentations for the tibial compartments [(a) lateral, (b) medial]. The agreements were evaluated by linear correlations of 0.90 and 0.91 and intra-class correlation coefficients (ICCs) of 0.90 and 0.81, respectively. The measure difference is shown as KneeIQ-Chondrometrics scores implying that KneeIQ overestimated compared to Chondrometrics. The horizontal, dotted lines show steps of a standard deviation above and below the mean.

For a subset of 58 knees, the volume scores are available for all four methods. The mutual agreements for this subset are shown in Table 6.

Table 6

For 58 OAI knees, volume scores are available for all groups. The inter-agreement is given as Pearson linear correlation coefficient (r) and mean signed relative difference (%). Positive differences mean that the method in the left column has higher volumes, on average.

Generally, the correlations are high between all methods, ranging from 0.90 to 0.97. There is a tendency for the automated methods to estimate the volumes higher than the manual scores from Chondrometrics.

3.5.

SKI10 Evaluation

The results for the validation on the SKI10 collection are available in Ref. 24. The performance on the training set gave Dice overlap scores of 0.97 for the tibia and femur bones and 0.64 to 0.73 for the cartilage compartments.³³

The official scores for the evaluation set include volume overlap errors of 25.9 and 25.7 for the femoral and tibial cartilages, contributing to cartilage scores of 65.4 and 66.7, respectively. The bone scores include RMS surface distances of 1.4 and 1.0 mm for femur and tibia, contributing to bone scores of 63.2 and 61.8.

3.6.

Comparison to Folkesson07

The importance of the multiatlas registration step was evaluated on the CCBR cohort. The Folkesson07 variant had volume overlaps of 0.823 and 0.772 on the medial tibial and femoral compartments, as compared to 0.839 and 0.804 for the full KneeIQ framework. The scan-rescan precision was 5.6% and 7.8% for Folkesson07 against 4.9% for KneeIQ.

4.1.

Previous Results on the Validation Collections

The CCBR, OAI, and SKI10 collections have previously been investigated with respect to segmentation and cartilage quantification. On the CCBR collection, Folkesson et al reported results on a fully automatic segmentation of the tibial and femoral medial cartilage compartments, including volume overlap scores at 0.81 and 0.77, respectively, and scan-rescan precision given by mean absolute volume differences of 5.8% and 7.4%, respectively.¹⁴

For OAI, several investigations were performed on the sagittal DESS scans. A pilot population of 19 with a mix of healthy and OA subjects were segmented twice unpaired and the precision was evaluated by RMS CV. For cartilage volume, RMS CV was 3.9% and 8.2% for medial and lateral tibia, and 6.2% and 4.6% for posterior medial and lateral femoral subcompartments.³⁴ Another OAI study compared different segmentation methods on the 19 subjects.³⁵ The across-methods pooled RMS CV scores for the DESS scans were 7% for medial and lateral tibial compartments and 6% and 7% for the central medial and lateral femoral compartments.

Using OAI scans, a multiobject graph optimization method demonstrated a good segmentation performance on a small subset of 9 scans.³⁶ The volume overlaps were between 0.80 and 0.84 for the cartilage compartments. Recently, an automatic method combining a bone ASM segmentation, a semantic context forest cartilage classifier, and smoothing by a graph cut optimization was presented at a workshop.³⁷ The method was validated on the OAI subcohort with manual segmentations by iMorphics also used in this paper. The Dice volume overlaps for the cartilages were between 0.79 and 0.85.

The semi-automated VirtualScopics segmentation method has also been applied to the OAI knee MRI;²³ however, no papers with segmentation performance metrics are known.

Finally, 19 methods have been evaluated on the SKI10 collection. In the normalized scores where 75 matches an expert, manual reader, the bone scores range from 11 to 77 and the cartilage scores range from 23 to 67.

4.2.

Comparison to Previous Results

The performance of the Folkesson07 variant was close to the results originally reported by Folkesson. This supports the observed performance improvements due to the new multiatlas registration step. On average for the two medial compartments, the volume overlap improved by 0.024 and the precision improved by 1.8%.

Looking at the segmentation accuracy on the OAI cohort, our validation volume overlaps between 0.81 and 0.87 for the medial/lateral tibial/femoral compartments compare well with the best published results and may even be slightly higher. However, for the patellar cartilage, our volume overlap at 0.74 is slightly lower than the best published results.

The studies on the 19 OAI subjects suggest that variations around 5% to 6% in tibial and femoral cartilage volume scores are to be expected for scan-rescan experiments using a specific method and around 7% using similar, but different methods. The agreements between the cartilage volumes from our method and the independent methods (Tables 5 and 6) appear to be within this range except for the comparison to Chondrometrics scores in the lateral tibial compartment. Apparently, even if our overestimations compared to the iMorphics segmentations are acceptable around 4 to 5%, the overestimation of iMorphics compared to Chondrometrics of 8% in this compartment resulted in a larger, total overestimation of 14%. Due to the lack of Chondrometrics segmentation masks, this specific discrepancy is difficult to directly investigate.

Our validation against a large collection of manual segmentations is similar in spirit to the evaluation by Shan et al.²² They also measure the linear correlation between automatic and manual segmentation volumes. For their 700 scans, the linear correlations were 0.87 for the tibial cartilage and 0.77 for the femoral cartilage. For the 1436 scans in our evaluation, we achieved automatic to manual correlations of 0.90 and 0.91 for medial tibial and femoral cartilages.

4.3.

Comparison to SKI10 Results

Our method performed mediocre for bone segmentation scoring 62.49 and second highest for cartilage segmentation scoring 66.10.

4.4.

Segmentation Errors

Visual inspection demonstrated that the voxel classification independently performed in each voxel resulted in focal fine-scale artifacts in the segmentations (see Fig. 3). Therefore, our results could possibly be improved by a regularization postprocessing step that refines the segmentation boundaries using either local or global shape information. This could be using a multiobject shape model or possibly by a more local regularization process like the graph cut optimization used by Wang et al.³⁷ However, markers that are to be quantified from the segmentations may be robust to these fine-scale details. A cartilage volume marker will likely not be significantly affected and markers that are quantified with explicit or implicit regularization will also be robust to this, such as surface smoothness³⁸ and joint congruity.⁶

For the OAI collection, our automatic segmentations overestimated the cartilage volumes by 4% to 5% compared to the iMorphics validation scans (see Table 5). Compared to the VirtualScopics, our estimations were, on average, almost equal. Compared to the Chondrometrics volumes, both the iMorphics and our method estimated the numbers to be higher. The plots in Fig. 6 reveal that this overestimation mainly originated from the knee with relatively little cartilage. The availability of semi-manual segmentations for the iMorphics validation set allowed further inspection of the source of the overestimation (Fig. 7). By applying the transformation from the multiatlas registration step, we transformed all segmentation differences to a common reference space. The figure illustrates that some locations were common for oversegmentation (defined as voxels in the automatic segmentation, but not in the validation segmentation, in more than 10% of the 44 cases). This oversegmentation region generally corresponded to the peripheral rim of the cartilage sheets, as illustrated for the medial tibial compartment in Fig. 7. This observation confirmed the conclusion from Williams that segmentation of the periphery is particularly challenging.¹⁹

Fig. 7

Locations of overestimation in the OAI segmentation. (a) Using the multiatlas registration transformations, the mean shape of the manual segmentations can be visualized, as done for the medial tibial compartment. Likewise, the differences between manual and automatic segmentations can be transformed to the same space. (b) The locations of the volume overestimations are visualized showing the locations that in at least 10% of the OAI validation scans were included by the automatic but not the manual segmentation. This is generally in the periphery of the sheet. These oversegmentation locations are color-coded with the mean distance to the manual segmentation with the range shown in the color bar in mm. The sheet is shown with external left, internal right, anterior up, and posterior down.

For the SKI10 scans, visual inspection revealed bone segmentation errors in the shafts toward the boundary of the scans for the cases with weak image quality. A statistical shape model would be the classical approach to handle this.

4.5.

Segmentation of Pathological Knees

It is often assumed that pathological structures are more challenging to analyze than healthy structures due to larger biological variation. The OAI validation population only included KL 2 and 3 knees; however, the CCBR collection allowed analysis of this effect for both automatic and manual segmentations. Specifically, with stratification according to degree of radiographic OA (KL score 0 or 1 versus $\mathrm{KL}>1$) of the CCBR rescan subgroup, the mean Dice overlap scores for our automatic method were 0.843 versus 0.802 for the medial tibial compartment and 0.814 versus 0.780 for the medial femoral compartment. In comparison, the repeated radiologist segmentations had Dice 0.860 versus 0.854 for tibial and 0.870 versus 0.837 for femoral. Therefore, the segmentation accuracy for Dice may be expected to be 0.03 lower for ROA than for healthy tissues, where the effect is slightly larger for the automatic method than for expert manual segmentations.

4.6.

Methodology Choices

There are several approaches to automatic knee MRI segmentation. Some methods are directly hierarchical starting with a bone segmentation that then aids the (more difficult) cartilage segmentation with features for distance-from-bone or position-relative-to-bone.^{18,36,38–42} Other methods achieve a similar coarse-to-fine effect by applying atlas-based registration before either nonrigid registration^21,40,43 and/or classifier-based segmentation.^10,22,43 In general, it appears that integration of global information and local features is essential for solving the challenging problem of segmentation of cartilage on the background of multiple other tissue types. This explicitly or implicitly allows a zoom to several, simpler, local segmentation tasks such as cartilage versus subchondral bone, cartilage versus meniscus, cartilage versus cartilage, and cartilage versus synovial fluid. Due to differences in populations, scanners, sequences, and particularly disease stage, it is challenging to compare segmentation performances across these publications. However, it would appear that tibial/femoral cartilage Dice volume overlaps around 0.85 are attainable with these automatic methodologies.

In our methodology, the rigid multiatlas registration step provided the position features that explicitly allowed ROI definition and implicitly allowed selection of $k$-NN features specific to each of the subsegmentation tasks. Potentially, adding a nonrigid (articulating) registration step could improve the position specificity even further in future work. We could speculate that this could be particularly useful for the patellar cartilage whose postregistration position variation was considerably higher than for the other cartilage compartments.