WO2009052562A1 - Automatic segmentation of articular cartilage in mr images - Google Patents

Automatic segmentation of articular cartilage in mr images Download PDF

Info

Publication number
WO2009052562A1
WO2009052562A1 PCT/AU2008/001559 AU2008001559W WO2009052562A1 WO 2009052562 A1 WO2009052562 A1 WO 2009052562A1 AU 2008001559 W AU2008001559 W AU 2008001559W WO 2009052562 A1 WO2009052562 A1 WO 2009052562A1
Authority
WO
WIPO (PCT)
Prior art keywords
cartilage
patient specific
3d model
segmentation
bone
Prior art date
Application number
PCT/AU2008/001559
Other languages
French (fr)
Inventor
Jurgen Fripp
Sebastien Ourselin
Stuart Crozier
Original Assignee
Commonwealth Scientific And Industrial Research Organisation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority to AU2007905805 priority Critical
Priority to AU2007905805A priority patent/AU2007905805A0/en
Application filed by Commonwealth Scientific And Industrial Research Organisation filed Critical Commonwealth Scientific And Industrial Research Organisation
Publication of WO2009052562A1 publication Critical patent/WO2009052562A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00Three dimensional [3D] modelling, e.g. data description of 3D objects
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/12Edge-based segmentation
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/143Segmentation; Edge detection involving probabilistic approaches, e.g. Markov random field [MRF] modelling
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/149Segmentation; Edge detection involving deformable models, e.g. active contour models
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10072Tomographic images
    • G06T2207/10088Magnetic resonance imaging [MRI]
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30008Bone

Abstract

This invention concerns the automatic segmentation of articular cartilage in magnetic resonance (MR) images, especially but not exclusively for knee cartilage. In one aspect the invention is a method, in another it is software. In particular the following steps may be used: Filtering the captured patient specific MR image data to smooth it and extracting gradients from it. Generating a patient specific 3D model of the articular bone surfaces, by fitting an a priori 3D model of the bone surfaces to captured patient specific MR image data. Relaxing the patient specific 3Dmodel of the bone surfaces along a probable bone-cartilage interface (BCI) derived from a priori knowledge of the BCI, and updating the patient specific 3D model to identify the BCI. Estimating the properties of the tissue types in the updated patient specific 3D model. And, applying the estimates of tissue types and the extracted gradients as process drivers to a cost function that iteratively assigns values that define the cartilage in the updated patient specific 3D model using an a priori model of expected cartilage properties, to produce a bone-cartilage segmentation with one common BCI.

Description

Title

Automatic Segmentation of Articular Cartilage in MR Images

Related Application

This application claims priority from Australian Provisional Patent Application No 2007905805 filed on 23 October 2007; the entire contents of this provisional application are incorporated herein by reference. Technical Field

This invention concerns the automatic segmentation of articular cartilage in magnetic resonance (MR) images, especially but not exclusively for knee cartilage. In one aspect the invention is a method, in another it is software.

Background Art

The clinical diagnosis of knee problems is based mostly on clinical symptoms, such as pain, qualitative visual analysis for the presence of osteophytes, and quantitative assessment of X-rays by a radiologist. Changes in cartilage are usually quantitatively assessed indirectly, from measurements of the joint space width (between the bones) as seen in X-rays.

Magnetic resonance (MR) imaging can directly visualise cartilage tissue, and it is currently the most accurate non-invasive technique for assessing the articular cartilage in vivo [13,14]. This is important since up to 13% of cartilage can be lost before it is detectable on X-rays [25]. A large number of different pulse sequences can be used to accurately image articular cartilage[14]. However, MR images are more challenging to interpret than X-rays, and it is more difficult to obtain statistically significant results. Assessment of cartilage tissue is usually performed on each cartilage separately, or in subregions, using morphological measures such as volume, thickness, surface area or curvature. To calculate these measures it is necessary for the cartilage to be segmented separately, or in subregions, a task that can significantly influence the error and reproducibility of the quantitative analysis. To date this task has proved difficult to automate.

MR images of the knee can exhibit significant imaging artifacts that obscure the cartilage, and erroneously appear as defects. The difficulties arising from these artifacts are compounded by poor image resolution, anisotropy, magic angle and partial volume effects. As a result in clinical work, state of the art cartilage segmentation is performed manually [9] or semi-automatically. Both techniques are time consuming, taking several hours for each knee [26], and have low reproducibility. As a result the process is very costly, making it impractical to use MR for clinical assessment and studies. There have been several attempts to devise automated approaches to the problem of knee cartilage segmentation, however, to date none of these approaches have shown sufficient accuracy to be used in clinical studies: Semi-automated approaches are usually performed in 2D on individual slices from the MR image of the knee. Processing each slice takes 30 to 45 seconds giving a total of 30 minutes to an hour per knee, which is similar to the time required for direct manual segmentation. Most semi-automatic approaches proceed on a slice by slice basis, and include:

Region growing [10].

B-spline snakes [1 1] [16].

Signal intensity based thresholding with a seed growing algorithm [15]. Live-wires [12].

Active Shape Models (ASMs) [17].

Watershed algorithms [6] [18].

Edge detection [19]

Active contours [21].

In longitudinal studies, significantly improved follow up segmentation times can be obtained by assuming the bones only undergo rigid changes, and by using registration, propagated regions of interest and intensity thresholding [20]. One approach to fully automatic cartilage segmentation directly segments the cartilage tissue. There are two prominent works: fully automatic Folkesson [5] and partially automated Grau [6]. Folkesson uses a trained tissue classifier approach, with additional features including absolute cartilage position. Grau uses a modified watershed algorithm with prior information. Another approach performs segmentation of the bones first and then uses this as a 'prior' to perform cartilage segmentation. This approach was first presented in [4] and has been used in other works [7].

Pakin [23] uses a region growing scheme that involves two-class clustering to segment the cartilages. This approach requires the bones to be already segmented, has low sensitivity (0.66) and was only evaluated on a single scan.

Tamez-Pena [7] presents an approach that requires two MR image acquisitions to segment the bones followed by interactive correction. This manually corrected bone segmentation is then processed by an active contour algorithm to segment the cartilages. No prior or trained knowledge is required, but the technique does require two scans and significant user interaction and correction (30 + minutes on average) to obtain accurate results.

Li and Millington use a similar idea in [8] to produce an almost fully automatic segmentation of the ankle cartilages, by first segmenting the bones, creating a surface mesh from which a local graph is built and two (bone and cartilage) surfaces are extracted simultaneously using two separate cost functions. This approach only requires a simple manual initialization of seed points for the bone segmentation. Some investigators have used three-dimensional reconstruction of the articular cartilage to perform subsequent volumetric, thickness and surface area quantification of the entire cartilage, each cartilage or sub components surfaces. This type of analysis is the most common form of quantitative analysis in the literature [13].

The inventors' earlier work included bone segmentation and extraction of the bone- cartilage interfaces; previously published in [2]. They have also published a paper about the generation of thickness models and extraction of prior knowledge about the probable location of the bone-cartilage interface; this was presented in [3]. Both of [2] and [3] are incorporated herein by reference. Nevertheless, the current state of the art in automated cartilage segmentation does not provide the accuracy in cartilage segmentation required to perform quantitative analysis. Such analysis is useful for clinical studies, surgical treatments and drug trials.

Disclosure of the Invention

This invention makes use of 3D a priori models of the knee bones, the probable bone- cartilage interface and the expected cartilage properties. These models include properties such as shape, thickness, curvature and spatial relationship variation exhibited in the knee bones and cartilages, and may also include cartilage appearance, as well as other trained information. The model may be generated using subjects from appropriate demographics. In addition the invention makes use of localization and patient specific tissue models generated from magnetic resonance (MR) images.

The process proceeds using a segmentation hierarchy where the bones are segmented as a preliminary step, and then the articular cartilages are segmented implicitly as a thickness map and then explicitly as a surface and voxelisation. Subsequent to this the meniscus cartilage may also be segmented, as well as other objects.

In particular the invention is an automatic process for the segmentation of articular cartilage in magnetic resonance (MR) images, comprising the steps of:

Filtering the captured patient specific MR image data to smooth it, and extracting gradients from it.

Generating a patient specific 3D model of the articular bone surfaces, by fitting an a priori 3D model of the bone surfaces to captured patient specific MR image data. Relaxing the patient specific 3Dmodel of the bone surfaces along a probable bone-cartilage interface (BCI) derived from a priori knowledge of the BCI, and updating the patient specific 3D model to identify the BCI.

Estimating the properties of the tissue types in the updated patient specific 3D model. Applying the estimates of tissue types and the extracted gradients as process drivers to a cost function that iteratively assigns values that define the cartilage in the updated patient specific 3D model using an a priori model of expected cartilage properties, to produce a bone-cartilage segmentation with one common BCI. The process of the invention is able, starting with an MR image, to automatically generate separate accurate segmentation of the femoral, patella and each tibia cartilage. The data may be of sufficient quality to allow the segmented cartilages to be analysed using quantitative measures. This in turn may allow monitoring of cartilage changes, including morphological changes and the detection of lesions and focal defects.

The first step in generating a patient specific 3D model of the articular bone surfaces may involve an affine registration of the a priori 3D model of the bone surfaces to the acquired (patient specific) MR image data. Filtering, may also be applied to the MR image data to smooth the image data; such as median filtering, to generate image gradient data or extract other features. This filtering can be used to improve the fitting of the model of the bone surfaces.

The relaxation of the patient specific 3Dmodel of the bone surfaces may also benefit from the results of the filtering applied directly to the captured patient specific MR image data. This filtering may involve anisotropic filtering. The relaxation process may take place iteratively, and a final 3D model of the patient specific bone surfaces may be generated as an intermediate product before the BCI is finally identified in the updated patient specific model.

The estimation of the tissue types in the updated patient specific 3D model may be represented as probabilities.

The properties in the a priori model used by the cost function may include cartilage thickness, curvature and appearance. The appearance may be represented by probabilities.

The cost function may operate in a similar fashion to parametize all the properties. For instance, taking thickness as an example, in each iteration the cost function assigns the thickness to be the place of maximum cost found in the capture region, irrespective of neighbouring assignments, to produce an estimated cartilage thickness map. The cost function itself is constrained by the values in the model of expected properties. A feedback loop passes the estimated cartilage thickness map back through the a priori model to constrain parametization by negative feedback.

The process is able to implicitly obtain a cartilage thickness map in 3D from the underlying bone. It is also able to allow the calculation of cartilage volume, thickness and surface area, both as a 3D surface and after voxelization of the data.

In a further aspect the invention is a software, in the form of machine readable coded on a machine readable medium, to perform the method and allow submission, processing and return of segmentation results and quantitative analysis.

The software may be used as a standalone application, with an MR scanner or provided in a web style interface where clinicians and researchers can submit MR data and receive the processed and analysed results. Brief Description of the Drawings

An example of the invention will now be described with reference to the knee joint and the accompanying drawings, in which:

Fig. 1 is a flowchart overview of the present invention.

Fig. 2(a) is a flowchart showing the initialization and bone segmentation phases of the present invention in detail.

Fig. 2(b) is a flowchart showing the cartilage segmentation phase of the present invention.

Fig. 3(a) is a pictorial illustration of the initialisation process. Fig. 3(b) is a pictorial illustration of bone segmentation. Fig. 3(c) is a pictorial illustration of cartilage segmentation.

Fig. 4 is a sagittal slice from a FS SPGR.

Fig. 5 (a), (b), (c), (d), (e), (f), (g), (h) are a series of eight saggital slices arranged into pairs of comparative results, where: (a) is a pair of MR images,

(b) is a pair of Manual Segmentations,

(c) is a pair of NRR, and

(d) is a pair using the approach of the present invention. Fig. 6 is a surface rendering of the example of Fig. 5.

Fig. 7 is a graph of the DSC scores, for cases 1 to 20 excluding case 6, obtained for each individual cartilage from three sets: the patella, tibia and femur cartilage. Fig. 8(a) is a graph of mean patella volume; both manually and automatically derived. Thickness is calculated using a laplacian and exact Euclidean distance transform (EEDT).

Fig. 8(b) is a graph of mean patella thickness; both manually and automatically derived. Thickness is calculated before and after Shape based interpolation (SBI).

Fig. 9(a) is a graph of mean tibia volume; both manually and automatically derived. Thickness is calculated using a laplacian and exact Euclidean distance transform (EEDT).

Fig. 9(b) is a graph of mean tibia thickness; both manually and automatically derived. Thickness is calculated before and after Shape based interpolation (SBI).

Fig. 10(a) is a graph of mean femur volume; both manually and automatically derived. Thickness is calculated using a laplacian and exact Euclidean distance transform (EEDT). Fig. 10(b) is a graph of mean femur thickness; both manually and automatically derived. Thickness is calculated before and after Shape based interpolation (SBI). Fig 1 1 (a), (b) and (c) are sagittal slices with automatically obtained segmentation for case 9 with shading on the cartilage tissue reflecting the Laplacian Thickness value.

(a) is slice 17, (b) is slice 35, and

(c) is slice 47.

Fig 12 (a) and (b) are femoral cartilage volume renderings of the Laplacian thickness obtained for case 9. (a) is manually segmented, and

(b) is automatically segmented.

Fig 13 (a) and (b) are patella and tibia cartilage volume renderings of the Laplacian thickness obtained for case 9. (a) is manually segmented, and

(b) is automatically segmented.

Best Modes of the Invention Referring first to Figs. 1, 2 and 3 an example of the invention is presented for Tl weighted fat suppressed spoiled recall gradient recall MR image 1. An example of a sagittal slice 2 from such an image is shown in Fig. 4; in this case the image is a FS SPGR. In this image a femur 3, patella 4, and tibia 5 can be seen, as can regions of cartilage 6 around the bottom of the femur, top of the tibia and on the side of the patella facing the femur.

Initialisation Fig. 2(a) and Fig. 3(a) 10

Starting with the patient specific MR image data, the first step involves an affine registration with an a priori 3D model of the bone surfaces 12 to propagate the bone surfaces in the patient specific MR image data.

Bone Segmentation Fig. 2(a) and Fig. 3(b) 20

The result is from a 3D active shape model (ASM) 21. The model is matched to filtered image data, such as median filtering, which is used to smooth the image and generate image gradients.

Anisotropic Image Filtering Fig. 2(a) and Fig. 3(c) 30

The captured patient specific MR image data is also anisotropically filtered to smooth it and extracting gradients from it 32.

Initial Extraction of the Bone Cartilage Interface (BCD Fig. 2(a) and Fig. 3(c) 40 a priori knowledge about the probable location of the bone-cartilage interface, in the form of another model 42, is used together with the ASM to make an initial extraction of the bone-cartilage interface (BCI) [2] from the patient specific 3D model of the articular bone surfaces. The local image properties, namely that cartilage will be regions of hyper intense tissue in fat suppressed spoiled gradient recall images, may also be used, as may the anisotropically filtered MR image data 32 to improve extracted BCI.

Relaxation of the patient specific 3Dmodel Fig. 2(a) and Fig. 3(C) 50 The patient specific 3Dmodel of the bone surfaces is relaxed along a probable bone- cartilage interface (BCI) derived from a priori knowledge of the BCI, in order to improve the BCI and update the patient specific 3D model with the improved BCI. Again, the local image properties, namely that cartilage will be regions of hyper intense tissue in fat suppressed spoiled gradient recall images, may also be used, as may the anisotropically filtered MR image data 32 to improve extracted BCI.

Extracted BCI Fig. 2(a) and Fig. 3(c) 60

A final 3D bone surface model 52 is produced, and a final BCI 60 is extracted and marked in the MR image.

Distance Image Fig. 2(b) and Fig. 3(c) 70

Regions of the extracted BCI are masked and a distance image 70 is produced.

Estimation of Tissue Types Fig. 2(b) and Fig. 3(c) 80 The tissue types 82 are estimated in the patient specific 3D model. This involves the use of profiles of tissue properties that are compared with the image properties. Probability estimates of the patient specific tissue are then obtained, for instance using a three class expectation maximization using Gaussian mixture models. Parametizing Tissue Properties Fig. 2(b) and Fig. 3(c) 90

The estimates of tissue types 82 and the extracted gradients 32 are applied as process drivers to a cost function 90. The cost function operates in a similar fashion to parametize three properties of the cartilage, that is thickness, curvature and localised tissue appearance. Taking thickness as an example, in each iteration the cost function assigns the maximum cost function in the capture region to each thickness, irrespective of neighbouring assignments, to produce an estimated cartilage thickness map 92. The cost function itself is constrained by the values in a model of expected properties 94. A feedback loop passes the estimated cartilage thickness map back through the a priori model to constrain parametization by negative feedback.

Voxelisation Fig. 2(b) and Fig. 3(c) 100

The output is a final thickness map 92 having a coupled bone-cartilage segmentation having one common surface BCI in between them, which can then be analysed, voxelised and used in quantitative analysis. The resulting image is shown in Fig.5(d) where it can be compared to the results of other techniques, and Fig 6.

Although the invention has been described with reference to a particular example, it should be appreciated that it could be exemplified in many other forms and in combination with other features not mentioned above. For instance, the approach may be customised for use in other sequences, besides Tl , that have different appearances, which could require the use of slightly different priors and parameters. The invention may be applied to any MR sequence with appearance similar to FS SPGR, such as MEDIC or water excitation Dual Echo in the Steady State (weDESS)..

Any point or surface based registration or model based approach can be used to obtain the bone segmentations. As this allows the propagation of the corresponding point sets, hence the embedded prior knowledge.

Any model creation approach can be used to obtain the prior knowledge of the probable location of the bone-cartilage interface. In this example gradient information was obtained directly from a smoothed version of the image (anisotropic diffusion). There are many variations that could be used to obtain smoothed images and calculate gradients, including gradient vector flow and directional gradient vector flow. The cost function can be extended to explicitly incorporate other constraints and knowledge, including curvature, localised tissue appearance (from generative priors, ie homogeneity variation of cartilage tissue, MR artefacts) and likelihood weightings including regional analysis. There are many variations and schemes that can be used to obtain probability estimates for tissue, including texture and tissue classifiers.

Example implementation: The process was validated several different ways and compared to three other techniques: non-rigid registration, tissue classifier and a modified watershed algorithm.

The first validation approach was volume based measures compared to manual segmentations on a database of 20 FS SPGR images.

The volume was estimated directly for twenty cases (excluding case 6), and from the estimates we found that (respectively manual and automatic) segmentations had an average volume of (4245, 3912), (6026, 6056) and (14703, 14463) mm3 and median absolute volume difference error of 5.57%, 5.47% and 5.44% for the patellar, tibial and femoral cartilages respectively (excluding case 17). The thickness was calculated from the whole BCI using an approach based on [25], from which obtained an average thickness of (2.63, 2.44), (1.89, 1.56) and (1.84,1.80) mm and average absolute thickness difference of (0.19, 0.33, 0.10 mm) for the patellar, tibial and femoral cartilage respectively (excluding case 17). The surface area difference had a median DSC of (0.95, 0.88, 0.94); see Fig. 7. All the experiments presented below were performed using a leaveone-out approach. The cartilage segmentations automatically obtained were compared to the expert binary manual segmentations using the following volume-based measures:

• sensitivity = TP /(TP + FN)

• specificity = TN/(TN + FP )

• DSC = 2TP /(2TP + FP + FN) where TP is true positive, TN is true negative, FP is false positive and FN is false negative. The sensitivity is the 'true positive fraction' and specificity the 'true negative fraction', while DSC (Dice similarity coefficient) is a spatial overlap index. The value of all these measures ranges from 0 to 1. For DSC, a value of 0 indicates no spatial overlap and a value of 1 indicates a complete overlap between the two sets of binary segmentation.

M EAN (STA NDA R D DEVIATION) ( M ED[ AN ) OF VALI DATION VI EAS URES .

Af fine Sensitivity Specificity DSC

- (Patella) 0.450 (0.163) (0.502) 0.998 (0.001 ) (0.998) 0.422 (0.164) (0.49 h

- (Tibia) 0 460 (0.170) (0.491 ) 0.998 (0.001) (0.998) 0 473 (0.166) (0.519)

- (Femur) 0.418 (0.143) (0.474) 0.994 (0.002) (0.994) 0.427 (0.138) (0.478)

Non-Rigid Sensitivity Specificity DSC

After 10 mm

- (Patella) 0.506 (0.178) (0.524) 0.998 (0.001) (0.998) 0.479 (0.186) (0.51 1 )

- (Tibia ) 0 652 (0.153) (0.704) 0.999 (0.001 ) (0.999) 0.671 (0.139) (0.735)

- (Femur) 0.664 (0.154) (0.71 1) 0.997 (0.002) (0.997) 0.682 (0.144) (0.734) After 2.5 mm

- (Patella) 0.787 (0.120) (0.833) 0.999 (0.001) (0.999) 0.736 (0.148) (0.786)

- (Tibia ) 0.751 (0.1 10) (0.778) 0.999 (0.001 ) (0.999) 0.769 (0.098) (0.814)

- (Femur) 0.777 (0.159) (0.823) 0.997 (0.002) (0.997) 0.754 (0.147) (0.796) After I mm

- (Patella1) 0.803 (0.1 19) (0 848) 0.999 (0.001) (0.999) 0 732 (0.156) (0.787)

- (Tibia ) 0.781 (0.156) (0 804) 0.999 (0.001) (0.999) 0.785 (0.095 ) (0.829)

- (Femur) 0.795 (0.162) (0.836) 0.997 (0.002) (0.997) 0.758 (0.148) i0.795)

Tissue Classifier Sensitivity Specificity DSC

- (Patella) 0.750 (0.124) (0.781 ) NA 0.81 1 (0.1 1 ) (0.853)

- (Tibia) 0.707 (0.065) (0.716) NA 0.795 (0.07) (0.820)

- (Femur) 0.838 (0.045) (0.829) NA 0.849 (0.07) (0.886)

Our Approach Sensitivity Specificity DSC

- (Patella) 0.821 (0.135) (0.849) 1.000 (0.000) ( 1.000) O.S33 (0.135) (0.870)

- (Tibia) 0.829 (0.207) (0.860) 0.999 (0.000) (0.999) 0.826 (0.083) (0.855)

- (Femur) 0.837 (0.162) (0.865) 0.999 (0.000) (0.999) 0.848 (0.076) (0.870) Table 2. Average (from 5 segmentations) of the total cartilage (patellar , tibial and fe mural) results obtained using our algorithm compared to the improved watershed approach of Grau (S] . Only the total cartilage was compared as Gran's approach cannot obtain the individual cartilage segmentations that are necessary to perform statistically significant quantitat ive analysis

I nip rove d Waters hed Our approach

Scan Sens, Spec. DSC Sens. Spec, DSC

(1) 0.8965 0.9987 0.8988 0 8410 0.9993 0.8897 (2) 0.8649 0,9990 0.8907 0 8402 CL 9994 0.8898

(3) 0.8763 0.9990 0.8984 0 8490 0.9992 0.8902

(4) 0.8905 0,9988 0.8978 0 8591 0.9992 0.8959

QUANTITATIVE ANALYS IS PERFORM ED ON THE CARTILAGES OF

EACH SCAN. MEAN (STD) (MAXI MU M ) THICKNESS CALCU LATED

ON THE BCL FOR EACH SCAN .

Our approach

Scan Volume (nιmΛ ) La c>laci an BCI (m m) πCDT BCl {mm )

Pntella ( 1 ) 2717 2.32 (0.94) <4 J4) 2.16 (0.97) (4 22)

(2) 2684 2.34 (096) (4.49) 2.17 (0.99) (4 25)

(3) 3093 2.47 (0.88) (4.32) 2.13 (0.W) (4 .09)

(4) 2808 23 (0.95) (4.42) 2.12 (0.9*? ) (4.1 2)

TiMa ( 1) 4471 .68 (0-83X4.18) 1 .58 (0.80) (3.96)

(2) 4395 .72 (OSO) (4.05) 1 .62 (0.78) (3.98)

U) 4225 .77 (0 -82) (4.1 1) 1 .67 (0.79) (3.99)

<4) 4323 .74 (0 ,86X4.41 ) 1 .63 (0.83) (4.15)

I'emu r ( 1 ) 10277 .«7 (0.69) (4.01 ) 1 . «0 (0.70) (3.96)

(2) 10246 .86 (0.69) (4.08) 1 ,78 (0.70) (4.05)

(3) 10644 .90 (0.69) (4.05) 1 .82 (0.71 ) (399)

(4) 10634 .87 (0.69X4.12) 1 .79 (0.70) (4 .1 5)

Quantitative analysis performed using the automatic segmentations compared to the manual segmentations are presented in the table above and in Figs. 8, 9 and 10.

Figs. 1 1 are sagittal slices with automatically obtained segmentation for case 9 with colourmap on the cartilage tissue reflecting the Laplacian Thickness value. Figs. 12 are femoral cartilage volume renderings of the Laplacian thickness obtained for case 9. And Figs. 13 (are patella and tibia cartilage volume renderings of the Laplacian thickness obtained for case 9. References:

[1] S. Ourselin, A. Roche, G. Subsol, X. Pennec, and N. Ayache, "Reconstructing a 3D Structure from Serial Histological Sections." Image and Vision Computing, 19(l-2):25- -31, January 2001.

[2] Jurgen Fripp. Stuart Crozier, Simon Warfield, and Sebastien Ourselin "Automatic segmentation of the bone and extraction of the bone-cartilage interface from magnetic resonance images of the knee." Phys Med Biol. 2007 Mar 21 ;52(6): 1617-31. The entire contents of reference [2] are incorporated herein by reference.

[3] Jurgen Fripp, Pierrick Bourgeat, Andrea Mewes, Simon Warfield, Stuart Crozier, and Sebastien Ourselin. "3D Statistical Shape Models to Embed Spatial Relationship Information". In Computer Vision for Biomedical Image Applications: Current techniques and future trends, an ICCV workshop, October 2005. The entire contents of reference [3] are incorporated herein by reference.

[4] Kapur T, Beardsley P, Gibson S, Grimson W, and Wells W. M, "Model-based segmentation of clinical knee MRI," in Proceedings IEEE International Workshop on Model-Based 3D Image Analysis (in conjuction with ICCV), Bombay, India, Jan. 1998, pp. 97-106.

[5] Folkesson, J., Dam, E. B., Olsen, O. F., Pettersen, P. C, Christiansen, C: Segmenting articular cartilage automatically using a voxel classification approach. IEEE Trans. Medical Imaging 26(1) (2007) 106-1 15

[6] Grau, V., Mewes, A., Alcaniz, M., Kikinis, R., Warfield, S.: Improved watershed transform for medical image segmentation using prior information. IEEE Trans. Medical Imaging 23(4) (2004) 447-458

[7] J. G. Tamez-Pena, M. Barbu-Mclnnis, and S. Totterman, "Knee cartilage extraction and bone-cartilage interface analysis from 3D MRI data sets," in SPIE: Medical Imaging 2004, vol. 5370, San Diego, CA, USA, May 2004, pp. 1774-1784. [8] Li, K.,Millington, S. ,Wu, X., Chen, D.Z., Sonka,M.: Simultaneous segmentation of multiple closed surfaces using optimal graph searching. In: Information Processing in Medical Imaging. Volume 3565 of LNCS., Glenwood Springs, CO, USA, Springer Verlag (2005) 406^17 [9] Cicuttini, F., Hankin, J., Jones, G., Wluka, A.: Comparison of conventional standing knee radiographs and magnetic resonance imaging in assessing progression of tibiofemoral joint osteoarthritis. Osteoarthritis Cartilage 13(8) (2005) 722-727

[10] Waterton, J., Solloway, S., Foster, J., Keen, M., Gandy, S., Middleton, B., Maciewicz, R., Watt, I., Dieppe, P., Taylor, C: Diurnal variation in the femoral articular cartilage of the knee in young adult humans. Magnetic Resonance in Medicine 43 (2000) 126-132 [1 1] Stammberger, T., Eckstein, F., Englmeier, K., Reiser, M.: Determination of 3D cartilage thickness data from MR imaging: computational method and reproducibility in the living. Magnetic Resonance in Medicine 41(3) (1999) 529-536

[12] Gougoutas, A., Wheaton, A., Borthakur, A., Shapiro, E., Kneeland, J., Udupa, J., Reddy, R.: Cartilage volume quantification via live wire segmentation. Academic Radiology 1 1(12) (2004) 1389-1395. [13] F. Eckstein, F. Cicuttini, J. Raynauld, J. Waterton, and C. Peterfly, "Magnetic resonance imaging (MRI) of cartilage in knee osteoarthritis (OA): morphological assessment," Osteoarthritis and Cartilage, vol. 14, pp. 46-75, May 2006.

[14] G. E. Gold, D. Burstein, B. Dardzinski, P. Lang, F. Boada, T. Mosher, "MRI of articular cartilage in OA: novel pulse sequences and compositional/functional markers", Osteoarthritis and Cartilage, vol 14, Supplement 1, pp 76-86, 2006.

[15]. Piplani M. A, Disler D.G, McCauley T.R, Holmes TJ, Cousins J.P. Articular cartilage volume in the knee: semiautomated determination from three-dimensional reformations of MR images. Radiology 1996; 198:855-9.

[16] Lynch J. A, Zaim S, Zhao J, Stork A, Peterfy CG, Genant HK. Cartilage segmentation of 3D MRI scans of the osteoarthritic knee combining user knowledge and active contours. Proceedings of SPIE. International

Society for Optical Engineering 2000; Volume 3979:925-935.

[17] Solloway S, Hutchinson CE, Waterton JC, Taylor CJ. The use of active shape models for making thickness measurements of articular cartilage from MR images. Magn Reson Med 1997;37:943e52.

[18] Ghosh S, Beuf O, Newitt D.C, Ries M, Lane N, and Majumdar S, "Watershed segmentation of high resolution articular cartilage images for assessment of osteoarthritis," ISMRM, 2000.

[19] Kshirsagar A. A, Watson PJ, Tyler J.A, Hall L.D. Measurement of localized cartilage volume and thickness of human knee joints by computer analysis of threedimensional magnetic resonance images. Invest Radiol 1998;33:289-99.

[20] Jaremko, J, Cheng R, Lambert R, Habib A, and Ronsky J, "Reliability of an efficient MRI-based method for estimation of knee cartilage volume using surface registration," Osteoarthritis and Cartilage, vol. 14, no. 9, pp. 914-922, Sept. 2006

[21] Kauffmann C, Gravel P, Godbout B, Gravel A, Beaudoin G, Raynauld JP, et al. Computer-aided method for quantification of cartilage thickness and volume changes using MRI: validation study using a synthetic model. IEEE Trans Biomed Eng 2003;50: 978-88. [22] Warfield S. K, Kaus M, Jolesz F. A, and Kikinis R, "Adaptive, template moderated, spatially varying statistical classification," Medical Image Analysis, vol. 4, no. 1, pp. 43-55, 2000.

[23] Pakin S. K, Tamez-Pena J. G, Totterman S, and Parker K. J, "Segmentation, surface extraction, and thickness computation of articular cartilage," in SPIE: Medical Imaging, Image Processing, vol. 4684, May 2002, pp. 155-166.

[24]. Aldasoro, C. R and Bhalerao A, "Volumetric texture segmentation by discriminant feature selection and multiresolution classification". IEEE Trans. Medical Imaging 26(1) (2007) 1-15

[25] Jones, G. Ding, C. Scott, F. Glisson, M. and Cicuttini, "Early radiographic osteoarthritis is associated with substantial changes in cartilage volume and tibial bone surface area in both males and females.". OAC 12(2) (2004) 169-174

[26] Duryea, J. Neumann, G. Brem, M.H.Koh, W. Noorbakhsh, F. Jackson, R.D. Yu, Y. Eaton, CB, and Lang P, "Novel fast semi-automated software to segment cartilage for knee MR acquisitions", OAC 15(5) (2007) 487-492.

Claims

Claims
1. An automatic process for the segmentation of articular cartilage in magnetic resonance (MR) images, comprising the steps of: filtering captured patient specific MR image data to smooth it, and extracting gradients from it; generating a patient specific 3D model of the articular bone surfaces, by fitting an a priori 3D model of the bone surfaces to the captured patient specific MR image data; relaxing the patient specific 3Dmodel of the bone surfaces along a probable bone-cartilage interface (BCI) derived from a priori knowledge of the BCI, and updating the patient specific 3D model to identify the BCI; estimating the properties of the tissue types in the updated patient specific 3D model; and, applying the estimates of tissue types and the extracted gradients as process drivers to a cost function that iteratively assigns values that define the cartilage in the updated patient specific 3D model using an a priori model of expected cartilage properties, to produce a bone-cartilage segmentation with one common BCI.
2. A segmentation process according to claim 1, wherein the step of generating a patient specific 3D model of the articular bone surfaces involves an affine registration of the a priori 3D model of the bone surfaces to the acquired patient specific MR image data.
3. A segmentation process according to claim 2, wherein filtering is applied to the MR image data to smooth the image data.
4. A segmentation process according to claim 1 , wherein relaxation of the patient specific 3Dmodel of the bone surfaces utilizes the results of the filtering applied directly to the captured patient specific MR image data.
5. A segmentation process according to claim 1 or 4, wherein the relaxation process takes place iteratively.
6. A segmentation process according to claim 1 , 4 or 5, wherein the estimation of the tissue types in the updated patient specific 3D model is represented as probabilities.
7. A segmentation process according to claim 1, wherein the properties in the a priori model used by the cost function may include cartilage thickness, curvature and appearance.
8. A segmentation process according to claim 7, wherein the cartilage appearance is represented by probabilities.
9. A segmentation process according to claim 1, wherein the cost function operates, in each iteration to assign values to be the place of maximum cost found in the capture region, irrespective of neighbouring assignments.
10. A segmentation process according to claim 1 or 9, wherein, the cost function is constrained by the values in the model of expected properties.
1 1. A segmentation process according to claim 1 , 9 o r 10, wherein a feedback loop is used to constrain parametization of the values by negative feedback.
12. A software program comprising machine readable code on a machine readable medium to perform the method according to any preceding claim and allow submission, processing and return of segmentation results and quantitative analysis.
PCT/AU2008/001559 2007-10-23 2008-10-22 Automatic segmentation of articular cartilage in mr images WO2009052562A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
AU2007905805 2007-10-23
AU2007905805A AU2007905805A0 (en) 2007-10-23 Automatic segmentation of articular cartilage in MR images

Publications (1)

Publication Number Publication Date
WO2009052562A1 true WO2009052562A1 (en) 2009-04-30

Family

ID=40578962

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/AU2008/001559 WO2009052562A1 (en) 2007-10-23 2008-10-22 Automatic segmentation of articular cartilage in mr images

Country Status (1)

Country Link
WO (1) WO2009052562A1 (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
USD702349S1 (en) 2013-05-14 2014-04-08 Laboratories Bodycad Inc. Tibial prosthesis
USD752222S1 (en) 2013-05-14 2016-03-22 Laboratoires Bodycad Inc. Femoral prosthesis
CN106202738A (en) * 2016-07-14 2016-12-07 哈尔滨理工大学 Method for establishing joint cartilage two-phase model on the basis of hyperelastic solid phase characteristics
WO2017081373A1 (en) 2015-11-13 2017-05-18 University Of Oulu An assessment system and method for determining at least one of macro-topology, milli-topology, micro-topology and nano-topology
USD808524S1 (en) 2016-11-29 2018-01-23 Laboratoires Bodycad Inc. Femoral implant

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001032079A2 (en) * 1999-11-01 2001-05-10 Arthrovision, Inc. Evaluating disease progression using magnetic resonance imaging
WO2002023483A2 (en) * 2000-09-14 2002-03-21 Leland Stanford Junior University Technique for manipulating medical images

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001032079A2 (en) * 1999-11-01 2001-05-10 Arthrovision, Inc. Evaluating disease progression using magnetic resonance imaging
WO2002023483A2 (en) * 2000-09-14 2002-03-21 Leland Stanford Junior University Technique for manipulating medical images

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
FRIPP ET AL.: "Automatic Segmentation of Articular Cartilage in Magnetic Resonance Images of the Knee", MEDICAL IMAGE COMPUTING AND COMPUTER-ASSISTED INTERVENTION, 12 October 2007 (2007-10-12), ISBN: 978-3-540-757, Retrieved from the Internet <URL:http://www.springerlink.com/content/u317x7876635011k> *
KAUFFMANN ET AL.: "Computer-Aided method for Quantification of Cartilage Thickness and Volume Changes Using MRI: Validation Study Using a Synthetic Model", IEEE TRANSACTION ON BIOMEDICAL ENGINEERING, vol. 50, 2003, pages 978 - 988 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
USD702349S1 (en) 2013-05-14 2014-04-08 Laboratories Bodycad Inc. Tibial prosthesis
USD752222S1 (en) 2013-05-14 2016-03-22 Laboratoires Bodycad Inc. Femoral prosthesis
WO2017081373A1 (en) 2015-11-13 2017-05-18 University Of Oulu An assessment system and method for determining at least one of macro-topology, milli-topology, micro-topology and nano-topology
CN106202738A (en) * 2016-07-14 2016-12-07 哈尔滨理工大学 Method for establishing joint cartilage two-phase model on the basis of hyperelastic solid phase characteristics
USD808524S1 (en) 2016-11-29 2018-01-23 Laboratoires Bodycad Inc. Femoral implant

Similar Documents

Publication Publication Date Title
Lynch et al. Automatic segmentation of the left ventricle cavity and myocardium in MRI data
Li et al. Group-sparse representation with dictionary learning for medical image denoising and fusion
Cocosco et al. A fully automatic and robust brain MRI tissue classification method
Ruiz-Alzola et al. Nonrigid registration of 3D tensor medical data
Iglesias et al. Multi-atlas segmentation of biomedical images: a survey
Collins et al. Model-based segmentation of individual brain structures from MRI data
Vovk et al. A review of methods for correction of intensity inhomogeneity in MRI
Shen et al. Measuring size and shape of the hippocampus in MR images using a deformable shape model
Wells et al. Adaptive segmentation of MRI data
Withey et al. Medical image segmentation: Methods and software
Iglesias et al. Robust brain extraction across datasets and comparison with publicly available methods
US7876938B2 (en) System and method for whole body landmark detection, segmentation and change quantification in digital images
US7024027B1 (en) Method and apparatus for three-dimensional filtering of angiographic volume data
Kamber et al. Model-based 3-D segmentation of multiple sclerosis lesions in magnetic resonance brain images
Despotović et al. MRI segmentation of the human brain: challenges, methods, and applications
US6842638B1 (en) Angiography method and apparatus
US20080292194A1 (en) Method and System for Automatic Detection and Segmentation of Tumors and Associated Edema (Swelling) in Magnetic Resonance (Mri) Images
US7787927B2 (en) System and method for adaptive medical image registration
Zhu et al. Multicontext fuzzy clustering for separation of brain tissues in magnetic resonance images
Kybic et al. Fast parametric elastic image registration
Suri et al. Computer vision and pattern recognition techniques for 2-D and 3-D MR cerebral cortical segmentation (Part I): a state-of-the-art review
Hou A review on MR image intensity inhomogeneity correction
Studholme et al. Accurate alignment of functional EPI data to anatomical MRI using a physics-based distortion model
Lundervold et al. Segmentation of brain parenchyma and cerebrospinal fluid in multispectral magnetic resonance images
Belaroussi et al. Intensity non-uniformity correction in MRI: existing methods and their validation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 08843200

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase in:

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 08843200

Country of ref document: EP

Kind code of ref document: A1