Author’s Accepted Manuscript

Detection and Prediction of Osteoarthritis in Knee

Gwidon W. Stachowiak, Marcin Wolski, Tomasz

PII: S2405-4518(16)30039-3
DOI: http://dx.doi.org/10.1016/j.bsbt.2016.11.004
Reference: BSBT44
To appear in: Biosurface and Biotribology
Received date: 7 October 2016
Revised date: 16 November 2016
Accepted date: 22 November 2016
Cite this article as: Gwidon W. Stachowiak, Marcin Wolski, Tomasz
Woloszynski and Pawel Podsiadlo, Detection and Prediction of Osteoarthritis in
Knee and Hand Joints Based on the X-ray Image Analysis, Biosurface and
Biotribology, http://dx.doi.org/10.1016/j.bsbt.2016.11.004
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
Detection and Prediction of Osteoarthritis in Knee and Hand Joints Based on the X-ray

Image Analysis

Gwidon W. Stachowiak • Marcin Wolski • Tomasz Woloszynski • Pawel Podsiadlo

Tribology Laboratory, School of Mechanical and Civil Engineering Curtin University, Bentley,

Western Australia 6102, Australia

Corresponding author:

Marcin Wolski

Department of Mechanical Engineering, Curtin University

GPO Box U1987 Perth, Western Australia 6845, Australia

Tel | +61-8-9266-1471, Fax | +61 8 9266 2681

Email | marcin.wolski@curtin.edu.au

1
Abstract

Current assessment of osteoarthritis (OA) is primary based on visual grading of joint space

narrowing and osteophytes present on radiographs. The approach is observer-dependent,

not sensitive enough for the detection of the early stages of OA and time consuming. A

promising solution is through fractal analysis of trabecular bone (TB) textures on

radiographs. The goal is to develop an automated decision support system for the

detection and prediction of OA based on TB texture regions selected on knee and hand

radiographs. In this review, we describe our progress towards this development which

was conducted in five stages, i.e., (i) development of automated methods for the selection

of TB texture regions on knee and hand radiographs (ii), development of fractal signature

methods for TB texture analysis, (iii) applications of the methods in the analysis of x-ray

images of knees and hands, (iv) development of TB texture classification system, and (v)

development of ReadMyXray website for knee x-ray analysis. The results achieved so far

are encouraging and it is hoped, that once the system is fully developed and evaluated, it

will be used to aid medical practitioners in the decision-making, i.e., in designing OA

preventative measures, treatments and monitoring the OA progression.

Keywords: osteoarthritis, trabecular bone, fractals, image analysis, classification

2
Introduction

Osteoarthritis (OA) is a debilitating and costly disease to all societies across the

globe [1]. As people live longer and the populations rapidly age, this cost is continuously

increasing and OA becomes an important issue that cannot be ignored. One way of

mitigating this cost is through early detection and, more importantly, prediction of OA. To

achieve this, new automated diagnosis and prognosis techniques designed to assist the

medical personnel in the decision-making must be developed. The techniques would

allow for identification of patients on a development pathway to OA. For these

individuals, OA management (e.g., exercise, dietary supplements, and lifestyle changes)

could be initiated in a timely manner, thus reducing impact of OA on quality of their lives

and slowing down its progression [2]. The patients would also benefit from inclusion in

clinical trials on disease-modifying osteoarthritis drugs (DMOADs), such as strontium

ranelate and bisphosphonates [3, 4]. The two DMOADs showed promising results in

animal models of OA. Strontium ranelate was found to reduce cartilage and bone

pathology, while bisphosphonates could prevent bone remodeling and preserve its

structural integrity.

Latest imaging techniques such as magnetic resonance imaging (MRI) and

computed tomography (CT) are routinely used in medical examinations. Classical x-ray

technique, even though more than 100 years old, is cheap, fast, easy to use and still

provides valuable information about changes occurring in human bones [5]. Current OA

3
assessment is based primarily on visual grading of joint space narrowing (JSN) and

osteophytes using atlases, e.g., the Osteoarthritis Research Society International (OARSI)

atlas [6] or Kellgren-Lawrence (K-L) grading scheme [7]. The individual grading is time

consuming, prone to high intra- and inter-rater variability, and is not sensitive enough to

detect changes occurring in joints at the early stages of OA [8-10]. For example, it was

shown that about 11% to 13% of cartilage volume can be lost before JSN of grade one is

observed on radiographs [8], and cartilage defect occur before radiographic OA or knee

joint pain are present [11].

A promising solution to this problem is through application of fractal analysis to

the trabecular bone (TB) texture radiographic images. The reasons are following: (i) TB

changes at the early stages of OA [12, 13], (ii) TB exhibits fractal properties, i.e., it is self-

similar over a range of scales [14, 15], (iii) TB texture images selected on 2D radiographs

are directly related to the underlying 3D bone structure [16], and (iv) OA changes have

been detected on TB texture images using fractal methods [17, 18]. In this work we focus

on OA in knee and hand joints, since these joints are most commonly affected by the

disease.

In this paper, the progress towards the development of the detection and

prediction of OA based on TB texture images is summarized. The work conducted has

been divided into five stages, i.e.: (i) development of automated methods for the selection

of TB texture regions on knee and hand radiographs (ii), development of fractal signature

methods for the bone texture analysis in knees and hands, (iii) applications of the methods

developed to the analysis of x-ray images of knees and hands, (iv) development of bone
4
texture classification system, and (v) development of dedicated website for the automated

analysis of x-ray images of knee joints.

Automated bone texture region selection on knee and hand radiographs

In the first stage, an automated methods for the selection of TB texture regions of

interest (ROI) on knee [19] and hand [20] x-ray images were developed. Currently, the

bone regions are usually selected manually by a trained observer which is time-consuming

and observer-dependent.

Knee radiographs

The method developed selects two square ROIs on an x-ray image on the TB

immediately under the medial and lateral cortical plates of the tibia without the need for a

human operator (Fig. 1(a)). The method has three major components: image

preprocessing, delineation of cortical bone plates and location of the ROIs. In the first

component, image is resized to 512 × 512 pixels, its contrast is adjusted, edges are

enhanced, and knee borders together with the joint space line are found. In the second

component, active shape models (ASMs) are used to delineate superior borders of the

medial and lateral cortical tibia plates. The final component consists of placing the ROIs

under the cortical borders and performing horizontal and vertical adjustments of their

positions so that they do not overlap periarticular osteopenia, subchondral bone sclerosis

and fibula head (only lateral ROI). The accuracy of the method in the selection of the

5
regions was compared against gold standard ROIs selected manually by an experienced

radiologist on 132 knee x-ray images. The agreement with the gold standard was

evaluated by means of similarity index (SI). SI is an overlap measure between the region

selected by the automated method and the corresponding gold standard region. It takes

values between 0 (no overlap) and 1 (entire overlap). In general, values of SI ≥ 0.70 indicate

excellent agreement [21]. The results obtained showed that the method developed is

accurate and reliable (SI ≥ 0.81). However, errors in ROI selection can occur. These errors

could be due to low contrast between knee and background and the deformation of knee

bones. For such cases, a special user interface was developed to enable manual adjustment

of ROIs.

Hand radiographs

For hand radiographs, automated selection of bone regions is far more difficult

than for knee joints. Compared to tibia, finger bones and subsequently ROIs are much

smaller, e.g., 20 × 20 pixels as opposed to 256 × 256 pixels for knee ROIs. Also, there are

more bones in hand, i.e. regions to be selected (16 in hands as opposed to 2 in knees). In

addition finger bones have varying orientations with respect to the image horizontal axis.

There is no such problem with knees. The method developed selects square ROIs above

and below cortical plates of the 2nd-5th distal interphalangeal (DIP) and proximal

interphalangeal (PIP) joints as these joints are often affected by OA. For one hand, all

together 16 ROIs (2 regions per joint) are selected (Fig. 1(b)). The ROIs selection is executed

in four stages: (i) x-ray image segmentation into hand and background, (ii) identification
6
of the 2nd-5th finger regions, (iii) localization of centre points of distal phalanges heads

and DIP, PIP and metacarpophalangeal (MCP) joints, (iv) location of the ROIs under and

above the DIP and PIP joints. The accuracy of the method developed was evaluated using

gold standard ROIs selected manually by a trained observer on 40 hand x-ray images.

Effects of hand translation, rotation, finger separation and flexion were tested. It was

found that the selection of the ROIs on hand radiographs using the new method is

accurate and reliable (SI ≥ 0.7). Nevertheless, in some isolated cases the selection of ROIs

can be inaccurate, especially for hand radiographs with severe OA or non-anatomical

structures present (e.g., rings). Thus, a mouse pointer tool was developed to adjust the

regions’ position and orientation if necessary.

Development of directional fractal signature methods for bone texture analysis in knees

and hands

In the second stage, methods for the analysis of the ROIs selected on knee and

hand radiographs were developed. The methods would have to be not only sensitive

enough to differentiate between healthy and osteoarthritic joints, but also be able to detect

minute changes occurring in TB due to OA, i.e., to detect an early development of OA.

As bone texture images exhibit anisotropic and multiscale nature (i.e., their

characteristics change with direction and scale) we decided to use fractal signature (FS)

based techniques, principles of which were originally developed by Peleg [22]. FS is a set

of fractal dimensions (FDs) calculated at individual scales (i.e., trabecular image sizes)

7
while FD is the most popular fractal measure of texture roughness. High (low) FD values

indicate rougher (smoother) textures.

Knee radiographs

For the analysis of x-ray images of knee joints, we developed the variance

orientation transform (VOT) method [23]. The method calculates FS in all possible

directions (i.e., directional FS (DFS)) allowing for detailed analysis of multiscale bone

texture roughness. Briefly, in the VOT method, absolute differences of grayscale level

values are calculated for pairs of pixels within a circular search region of a fixed size,

which moves across the bone texture image (Fig. 2(a)). The differences between the pixels’

grey-scale level values along with the corresponding directions and distances between the

pixels are stored. The direction is defined as an angle between a line running through the

pair of pixels and the x-ray horizontal axis. For each direction, the variances of the

differences stored are plotted in log–log coordinates against distances (Fig. 2(b)). The log–

log data points are then divided into overlapping sub-sets of points, a line is fitted to each

sub-set, and its slope is used to calculate the Hurst coefficient (H) at an individual scale. H

relates to FD as FD = 3 – H. The individual scale is represented by the distance associated

with the central log–log data point in the corresponding sub-set. At each scale, the Hurst

coefficients obtained are plotted in polar coordinates against directions and an ellipse is

fitted to the resulting plot (Fig. 2(c)). Using the ellipses fitted, the following bone texture

parameters are calculated (Fig. 2(d)):

8
 • Roughest part FS (FSSta). FSSta quantifies the texture roughness in the roughest

direction of TB at individual scales. FSSta is calculated as 3 - Sta, where Sta is half

the minor axis length of the ellipse fitted at a given scale.

• Texture aspect ratio signature (StrS). The parameter measures a degree of bone

texture anisotropy at different scales. It takes values between 0 and 1, with lower

values representing higher anisotropy. It is calculated as the ratio of the minor to the

major axes of the ellipses fitted at different scales.

• Texture direction signature (StdS). It describes dominating texture direction at

individual scales. It is defined as an angle between the major axis of the ellipse fitted

and the horizontal image axis at each scale.

As an example, the VOT method was applied to two bone texture regions selected in

the medial compartment of tibia from subjects with and without knee OA (Figs. 3(a) and

3(b)). Corresponding rose plots obtained for these two regions are shown in Figs. 3(e) and

3(f), respectively. From the Fig. 3, it can be seen that the plot corresponding to the healthy

knee approximates an ellipse, while the plot corresponding to the OA knee exhibits

roughly circular shape. This indicates that bone texture roughness in the healthy joint, as

compared to the diseased bone, varies considerably with direction, i.e., OA texture is less

anisotropic than in the healthy knee [24]. The mean StrS values obtained were 0.72 (OA)

and 0.52 (non-OA). The plots also show that Hurst coefficients calculated for healthy bone

are, in general, lower (i.e., higher FDs) than those for OA bone. This means that the

healthy bone texture is generally rougher than the OA bone [24]. This agrees with the
9
mean FDSta values obtained for the non-OA and OA subjects, i.e., 2.87 and 2.83,

respectively. The means of StdS obtained were 105° (non-OA) and 60° (OA), showing the

dominating texture directions, i.e., directions in which bone texture is the smoothest.

The effects of radiographic conditions on the VOT method were evaluated, i.e.,

radiographic noise, blur, exposure, magnification, projection angle, translation of the ROIs,

and image resolution/size [25, 26]. The results obtained showed that the method is

considerably affected by image noise (greater than 5%), blur, magnification, projection

angle (more than 10°), and resolution (less than 48 × 48 pixels).

Hand radiographs

ROIs selected on hand radiographs are considerably smaller than those on knee

radiographs, e.g., 20 × 20 (hand) vs 256 × 256 (knee) pixels. We found that the VOT at

image sizes less than 48 × 48 pixels does not provide reliable results as there is not enough

data for the analysis [27]. To address this problem, we developed an augmented VOT

method (AVOT) [27].

The AVOT method’s operating principles are similar to VOT, except that the size

of the search region is automatically adjusted to the size of a TB texture ROI, and FSs can

be calculated with respect to any reference direction. This was achieved through the

application of a recursive multi-directional pixel selection (RMPS) algorithm and a

reduction of the marginal subsets of points in the log–log plot from five to three. Details of

the algorithm and the reduction procedure are given in [27]. As a result, the AVOT

method can reliably calculate FSs of small bone texture regions selected on arbitrary
10
oriented finger bones.

Applications of the methods developed to the analysis of x-ray images of knees and

hands

In the third stage, the potential of the VOT and AVOT methods for the detection

and prediction of OA was evaluated in a number of following studies [27-31], i.e.:

 Detection of differences in TB texture between subjects with and without knee OA,

 Detection of differences in TB texture between subjects without radiographic knee

OA (K-L grade < 2), but with and without cartilage defects,

 Association of baseline bone texture with knee joint replacement (KJR) at 6-year

follow-up,

 Association of baseline bone texture with incident of knee OA at 30-, 60- and 84-

month follow-ups, and

 Detection of differences in TB texture between subjects with and without hand OA.

Detection of differences in TB texture between subjects with and without OA

Initially a pilot study was conducted to evaluate the VOT method abilities to

differentiate between OA and non-OA textures in knees [28].

For this study 26 cases and 26 controls were used. Controls were the subjects with

(K-L grade ≥ 2) and without (K-L grade < 2) knee OA, respectively. The subjects were

11
individually matched by sex, age and body mass index (BMI). Tibial bone texture ROIs

(two per knee joint) were selected on cases and controls knee x-rays, and analysed by the

VOT method. Examples of the ROIs selected on a knee x-ray are shown in Fig. 1(a). Bone

texture parameters obtained for cases and controls were compared using paired t-test. The

results showed that OA textures were smoother and less anisotropic than textures in

healthy joints, i.e., that trabeculae in OA joins are thicker and less structured than in non-

OA joints. For this study, a database from Lund University (Sweden) was used [32].

Detection of differences in TB texture between subjects with and without cartilage defects

In this case-control study, the VOT method’s sensitivity to pre-radiographic OA

bone changes in joints with and without cartilage defects was investigated [29]. Cases (n =

28) were subjects with cartilage defects (grade ≥ 2), whereas controls (n = 28) exhibited

normal cartilage (grade < 2). The subjects were individually matched by sex, age, BMI,

whether arthroscopic partial meniscectomy (APM) was performed or not, and APM’s

location (i.e., medial or lateral compartment). All subjects were without radiographic knee

OA (K-L grade < 2). Cartilage defects were identified and graded based on MRI scans [29].

An example of a MRI scan of a knee joint with cartilage defects is shown Fig. 4 [33].

Databases from the University of Western Australia and Monash University (Australia)

were used [34]. On each x-ray image, bone regions in the medial and lateral compartment

were selected and analysed using the VOT method. Paired t-tests were used to evaluate

differences in fractal parameters between cases and controls. The results showed that TB

texture was rougher for cases than for controls, suggesting that thinning and fenestration
12
of trabeculae occur in knee joints with early OA (i.e., with cartilage defects). This finding

seems to contradict our previous results [28] which showed that OA knees exhibit lower

roughness than healthy joints. However, a possible explanation is that TB changes occur in

two stages due to OA. In the early OA stage, trabecular structures undergo thinning and

fenestration, whereas in the late OA, TB thickens. This explanation is consistent with other

studies in which it was shown that early knee OA exhibits bone resorption, while later

stages of OA are characterized by bone formation, suggesting non-linear relationship

between TB changes and OA severity [35, 36].

The results from this study show that the VOT method is sensitive enough to

detect pre-radiographic OA bone changes.

Association of baseline bone texture with KJR at 6-year follow up

In another study, a potential of the VOT method in predicting KJR in patients with

OA was assessed [37]. Medial and lateral bone ROIs were selected on knee radiographs of

114 subjects with symptomatic knee OA. A database from Monash University (Australia)

was used [30]. 28 subjects (25%) underwent KJR over 6-years. The associations between

tertiles of bone texture parameters obtained at baseline and the incidence of KJR were

studied using logistic regression. The statistical analyses were performed with adjustment

for baseline age, sex, BMI, Western Ontario and McMaster Universities Osteoarthritis

Index (WOMAC) score, and osteophyte and JSN grades. It was found that in knees

exhibiting higher baseline bone texture roughness, the risk of KJR is reduced. This could

be the result of the decrease in size of trabeculae and increase in interconnectivity of TB

[38].

These results showed that the VOT method could identify subjects with knee OA

who are likely to undergo KJR.

Association of baseline bone texture with incident of knee OA at 30-, 60- and 84-month follow-ups

Recently, a much wider study, involving 3026 subjects (6052 knees) from the

Multicentre Osteoarthritis Study (MOST) [University of California, San Francisco], was

conducted [31]. In the study, 894 subjects (626 knees) recruited by University of Alabama

(UAB), and 807 subjects (1158 knees) from University of Iowa (UIowa) were used.

Associations between baseline TB texture and radiographic OA incidence at three follow-

up periods (30, 60 and 84 months) were investigated. At the baseline, subjects were

without radiographic knee OA (i.e., K-L grade 0 or 1). There were 195 (22%, UAB) and 303

(26%, UIowa) OA incident knees cumulative to 84 months. OA incidence was defined as

K-L grade 2 or more at follow-up.

The associations of quartiles of baseline fractal parameters calculated using the

VOT method with incident OA were evaluated using logistic regression adjusted for age,

sex, race, BMI, leg alignment and baseline K-L grade (0 or 1). The analyses were stratified

by two radiographic modalities used, i.e., digitized film in the UAB and computer

radiography in the UIowa. Results obtained showed that baseline bone texture parameters

are associated with the OA incident at follow-up. Specifically, for subjects from UAB, it
14
was found that the baseline higher medial bone roughness was associated with increased

odds of OA incidence at a follow-up. For subjects from UIowa, lower anisotropy of bone

textures at baseline was associated with increased odds of OA. The higher roughness of

bone texture at baseline in subjects who had incident of OA, as compared to those who did

not, could be a result of fenestration and breakage of TB structures due to early OA. The

results agree with our previous findings [29] that in knee joints with pre-radiographic OA,

bone texture is rougher than in healthy joints.

These results are promising, showing that the VOT method can be used in the

development of a system for knee OA prediction.

Detection of differences in TB texture between subjects with and without hand OA

Hand x-ray images from 40 subjects from the Osteoarthritis Initiative (OAI)

database [University of California, San Francisco] were analyzed using AVOT method.

The subjects were individually matched by sex, age, BMI and race [27]. Half of the subjects

(i.e., cases) had radiographic OA (approximate K-L grade ≥ 2) in the 5th DIP joint. Controls

did not have OA (approximate K-L grade < 2) in the joint. On each hand radiograph, two

ROIs were selected, i.e., the first in the base of the 5th distal phalanx and the second in the

head of the 5th middle phalanx (Fig. 5). The regions were analyzed with the AVOT

method and the differences in the parameters obtained between cases and controls were

evaluated using paired t-tests. We found that TB texture was smoother for cases than for

controls in the horizontal directions, rougher in the vertical direction, and less anisotropic.

These differences might indicate thickening and realignment of trabeculae in the

The results obtained were encouraging, showing that the AVOT method could be

used for OA detection in hands. A study with larger number of subjects will be conducted

in the future.

Development of bone texture classification system

In the fourth stage, we focused on the development of an automated system for

the classification of knee x-ray images based on bone textures. To simplify the

development, we initially decided to use a signature dissimilarity measure (SDM) method

which calculates only three parameters (i.e., overall roughness, anisotropy and direction)

[39]. The parameters quantify texture features sensitive to OA changes in trabecular bone.

The VOT method calculates a relatively large number of parameters (i.e., 27) which

considerably hinders image classification due to the “curse of dimensionality”. Full

description of the SDM method is given in [39]. Briefly, the method is executed in several

steps. First, a scale-space representation of a TB texture image is constructed by

convoluting the image data with Gaussian kernels at different scales. Second, the gradient

and Laplacian operators are applied over all scales and their extremum values are found.

Next, histograms of the differences of the extremum values in different directions are

constructed. Using the histograms, the three, mentioned above, parameters are then

obtained.

We assessed the performance of the SDM method in prediction of radiographic

16
OA progression and achieved high prediction scores in the sample of patients tested [40].

In the study, we analyzed the radiographs from 105 subjects (90 men and 15 women with

a mean age of 54 years) [40]. Altogether, radiographs from 203 knees, 68 with and 135

without radiographic tibiofemoral OA (i.e., K-L grade < 2), were analyzed. Two sets of

knee radiographs (baseline and follow-up) were obtained 4 years apart. ROIs of medial

compartment tibial TB texture were selected using an automated region selection method.

TB texture was then analyzed using the SDM method and three parameters were

produced for each ROI: roughness, degree of anisotropy and direction of anisotropy. A

logistic regression model based on the parameters was evaluated in the prediction of

medial tibiofemoral OA progression defined as an increase in medial JSN grade between

baseline and follow-up. The model was adjusted for age, sex, and BMI. For knees with and

without pre-existing radiographic OA at baseline, the regression model based on the

parameters calculated on medial TB texture had the prediction accuracy of 0.77 and 0.75

area under the curve (AUC), respectively.

We then used the SDM method to develop a dissimilarity-based multiple classifier

(DMC) system for the classification of knee x-ray images [41]. The schematic illustration of

the DMC is shown in Fig. 6. The DMC consisted of a heterogeneous ensemble of classifiers

(linear, quadratic, tree, k nearest neighbors, Parzen and support vector machine). To

produce a class for TB texture (e.g., healthy or OA), classifiers in the ensemble were

combined using a specially developed probabilistic model [42]. The model selectively

combined only those classifiers which performed well on similar TB texture images during

system training. We assessed the DMC in the detection of radiographic OA and prediction
17
of radiographic OA progression, and the results were encouraging.

For the detection, in a pilot study [41], knee radiographs from 51 subjects were

initially analyzed. Healthy and OA classes were defined as K-L grade 0 in both

tibiofemoral compartments in both knees and K-L grades 2 or 3 in at least one tibiofemoral

compartment in at least one knee. Prediction of radiographic OA progression was assessed

on two sets of knee radiographs taken from 50 subjects 4 years apart. The definition of

non-progressive and progressive classes was based on the increase in the sum of

tibiofemoral JSN and osteophytes grades between baseline and follow-up. The DMC

achieved classification accuracies of 90.51% and 80% (averages over 5 repetitions of two-

fold cross validation) for the detection of radiographic OA and prediction of radiographic

OA progression, respectively.

The results obtained indicate that detection and prediction of radiographic OA

using the classification system developed is possible. Thus, our future research will

involve the modifications of the classification system developed to include parameters

calculated by the VOT method on knee radiographs. Also, the system will be extended to

hand radiographs using parameters calculated by the AVOT method.

Development of ReadMyXray website

For the purpose of assisting the medical personnel in knee OA diagnosis we

developed ReadMyXray website (http://readmyxray.curtin.edu.au/). The website

provides, free of charge, fractal-based assessment of bone texture of knee x-rays. Medical

18
practitioners, researchers and members of a general public can upload knee radiographs

for analysis, which are then analysed using the VOT method. A bone roughness report is

prepared and forwarded to the uploader. An example report is shown in Fig. 7.

Currently, a work is undertaken to extend the website usability for analysis of hand

radiographs using the AVOT method and also to enable x-ray image classification by the

SDM method for both knee and hand x-rays.

Summary

The current research efforts directed towards the development of methods for

diagnostic and prognostic purposes of OA were described. The methods include

automated selection of TB texture regions on knee and hand radiographs, followed by a

numerical analysis of the regions selected and their classification.

Radiographs of knee joints were analyzed using the VOT method. A number of

studies were conducted. We found that the method could not only differentiate between

OA and healthy knee joints and identify pre-radiographic OA bone changes, but it also

could be used in prediction of OA and KJR.

For the analysis of very small images, like those found in hand radiographs, the

VOT method was augmented (i.e., AVOT), and successfully applied to hand radiographs.

We showed that the method is sensitive enough to show the differences between healthy

and OA hand joints.

Classification system of knee bone textures was developed. Initially, for the

19
classification, the bone textures selected were analyzed by means the SDM method. The

SDM calculates much lower number of bone texture parameters than the VOT method,

thus reducing the complexity of the classification system developed. As the system was

shown to have good predictive capability of radiographic OA progression, we are

planning to extend it by including support for the VOT method, and also to extend it to

hand radiographs analyzed by the AVOT method.

The results obtained show that automated OA detection and prediction based on x-

ray images of knee and hand joints is possible. To assist the medical practitioners in knee

OA diagnosis we developed ReadMyXray website.

It should be noted that as the imaging technologies advance rapidly, the techniques

developed can be adapted to be used in the analysis and classification of MRI images, CT

scans, ultrasound images, and others.

Our future work will include development of regression models for prediction of

OA in knee and hand joints and evaluation of the models developed in longitudinal and

prospective studies. It is also planned to modify and test the methods developed for

analysis of other joint disorders than OA, such as rheumatoid arthritis and osteoporosis.

ReadMyXray website will be further developed to provide fractal-based assessment and

classification of TB textures of not only knees, but also hands and other joints.

Acknowledgements

The authors wish to thank the Curtin University, Department of Mechanical Engineering

20
and the School of Civil and Mechanical Engineering for their support during preparation

of the manuscript.

References

[1] R. Bitton, The economic burden of osteoarthritis, Am J Manag Care. 15 (2009) S230-5.
[2] E. Kon, G. Filardo, M. Drobnic, H. Madry, M. Jelic, N. van Dijk, et al., Non-surgical
management of early knee osteoarthritis, Knee. Surg. Sports Traumatol. Arthrosc. 20
(2012) 436-449.
[3] D.G. Yu, H.F. Ding, Y.Q. Mao, M. Liu, B. Yu, X. Zhao, et al., Strontium ranelate
reduces cartilage degeneration and subchondral bone remodeling in rat osteoarthritis
model, Acta Pharmacol. Sin. 34 (2013) 393-402.
[4] G. Mohan, E. Perilli, I.H. Parkinson, J.M. Humphries, N.L. Fazzalari, J.S. Kuliwaba,
Pre-emptive, early, and delayed alendronate treatment in a rat model of knee
osteoarthritis: effect on subchondral trabecular bone microarchitecture and cartilage
degradation of the tibia, bone/cartilage turnover, and joint discomfort, Osteoarthritis
Cartilage. 21 (2013) 1595-1604.
[5] C. Buckland-Wright, Which radiographic techniques should we use for research and
clinical practice?, Best Pract. Res. Clin. Rheumatol. 20 (2006) 39-55.
[6] R.D. Altman, G.E. Gold, Atlas of individual radiographic features in osteoarthritis,
revised, Osteoarthritis Cartilage. 15 (2007) A1-A56.
[7] J.H. Kellgren, J.S. Lawrence, Radiological assessment of osteo-arthrosis, Ann. Rheum.
Dis. 16 (1957) 494-502.
[8] G. Jones, C. Ding, F. Scott, M. Glisson, F. Cicuttini, Early radiographic osteoarthritis
is associated with substantial changes in cartilage volume and tibial bone surface
area in both males and females, Osteoarthritis Cartilage. 12 (2004) 169-174.
[9] K.D. Brandt, R.S. Fife, E.M. Braunstein, B. Katz, Radiographic grading of the severity
of knee osteoarthritis: relation of the Kellgren and Lawrence grade to a grade based
on joint space narrowing, and correlation with arthroscopic evidence of articular
cartilage degeneration, Arthritis Rheum. 34 (1991) 1381-1386.
[10] C. Ding, F. Cicuttini, J. Jones, How important is MRI for detecting early
osteoarthritis?, Nat. Clin. Pract. Rheumatol. 4 (2008) 4-5.
[11] M.F. Sowers, C. Hayes, D. Jamadar, D. Capul, L. Lachance, M. Jannausch, et al.,
Magnetic resonance-detected subchondral bone marrow and cartilage defect
characteristics associated with pain and X-ray-defined knee osteoarthritis,
Osteoarthritis Cartilage. 11 (2003) 387-393.
[12] M. Ding, A. Odgaard, I. Hvid, Changes in the three-dimensional microstructure of
human tibial cancellous bone in early osteoarthritis, J. Bone Joint Surg. Br. 85 (2003)
906-912.

21
[13] S.K. Boyd, R. Müller, J.R. Matyas, G.R. Wohl, R.F. Zernicke, Early morphometric and
anisotropic change in periarticular cancellous bone in a model of experimental knee
osteoarthritis quantified using microcomputed tomography, Clin. Biomech. 15 (2000)
624-631.
[14] N.L. Fazzalari, I.H. Parkinson, Fractal dimension and architecture of trabecular bone,
J. Pathology. 178 (1996) 100-105.
[15] I.H. Parkinson, N.L. Fazzalari, Methodological principles for fractal analysis of
trabecular bone, J. Microsc. 198 (2000) 134-142.
[16] R. Jennane, R. Harba, G. Lemineur, S. Bretteil, A. Estrade, C.L. Benhamou, Estimation
of the 3D self-similarity parameter of trabecular bone from its 2D projection, Med.
Image Anal. 11 (2007) 91-98.
[17] J.A. Lynch, D.J. Hawkes, J.C. Buckland-Wright, Analysis of texture in
macroradiographs of osteoarthritic knees using the fractal signature, Phys. Med. Bio.
36 (1991) 709-722.
[18] V.B. Kraus, S. Feng, S. Wang, S. White, M. Ainslie, A. Brett, et al., Trabecular
morphometry by fractal signature analysis is a novel marker of osteoarthritis
progression, Arthritis Rheum. 60 (2009) 3711-3722.
[19] P. Podsiadlo, M. Wolski, G.W. Stachowiak, Automated selection of trabecular bone
regions in knee radiographs, Med. Phys. 35 (2008) 1870-1883.
[20] M. Wolski, M. Englund, G.W. Stachowiak, P. Podsiadlo. Automated selection of
trabecular bone texture regions on hand radiographs. in ISMRM Workshop on Imaging
Based Measures of Osteoarthritis 2015. Pacific Grove, CA, USA.
[21] A.P. Zijdenbos, B.M. Dawant, R.A. Margolin, A.C. Palmer, Morphometric analysis of
white-matter lesions in mr-images - method and validation, IEEE Trans. Med.
Imaging. 13 (1994) 716-724.
[22] S. Peleg, J. Naor, R. Hartley, D. Avnir, Multiple resolution texture analysis and
classification, IEEE Trans. Pattern Anal. Mach. Intel. 6 (1984) 518-523.
[23] M. Wolski, P. Podsiadlo, G.W. Stachowiak, Directional fractal signature analysis of
trabecular bone: evaluation of different methods to detect early osteoarthritis in knee
radiographs, P. I. Mech. Eng. H. 223 (2009) 211-236.
[24] P. Podsiadlo, L. Dahl, M. Englund, L.S. Lohmander, G.W. Stachowiak, Differences in
trabecular bone texture between knees with and without radiographic osteoarthritis
detected by fractal methods, Osteoarthritis Cartilage. 16 (2008) 323-329.
[25] M. Wolski, P. Podsiadlo, G.W. Stachowiak, Directional fractal signature analysis of
trabecular bone: evaluation of different methods to detect early osteoarthritis in knee
radiographs, P. I. Mech. Eng. H. 223 (2009) 211-236.
[26] M. Wolski, P. Podsiadlo, G.W. Stachowiak, A new method to measure trabecular
bone texture on hand radiographs: data from the Osteoarthritis Initiative,
Osteoarthritis Cartilage. 22 (2014) S257-S258.
[27] M. Wolski, P. Podsiadlo, G.W. Stachowiak, Directional fractal signature methods for
trabecular bone texture in hand radiographs: Data from the Osteoarthritis Initiative,
Med. Phys. 41 (2014)

22
[28] M. Wolski, P. Podsiadlo, G.W. Stachowiak, L.S. Lohmander, M. Englund, Differences
in trabecular bone texture between knees with and without radiographic
osteoarthritis detected by directional fractal signature method, Osteoarthritis
Cartilage. 18 (2010) 684-690.
[29] M. Wolski, G.W. Stachowiak, A.R. Dempsey, P.M. Mills, F.M. Cicuttini, Y. Wang, et
al., Trabecular bone texture detected by plain radiography and variance orientation
transform method is different between knees with and without cartilage defects, J.
Orthop. Res. 29 (2011) 1161-7.
[30] F.M. Cicuttini, G. Jones, A. Forbes, A.E. Wluka, Rate of cartilage loss at two years
predicts subsequent total knee arthroplasty: a prospective study, Ann. Rheum. Dis.
63 (2004) 1124-1127.
[31] P. Podsiadlo, M. Nevitt, M. Wolski, G. Stachowiak, J. Lynch, I. Tolstykh, et al.,
Baseline trabecular bone and its relation to incident radiographic knee osteoarthritis
and increase in joint space narrowing score: directional fractal signature analysis in
the MOST study, Osteoarthritis Cartilage. 24 (2016) 1736-1744.
[32] H. Roos, M. Laurén, T. Adalberth, E.M. Roos, K. Jonsson, L.S. Lohmander, Knee
osteoarthritis after meniscectomy: prevalence of radiographic changes after twenty-
one years, compared with matched controls, Arthritis Rheum. 41 (1998) 687-693.
[33] A.E. Wluka, Y. Wang, M. Davies-Tuck, D.R. English, G.G. Giles, F.M. Cicuttini, Bone
marrow lesions predict progression of cartilage defects and loss of cartilage volume
in healthy middle-aged adults without knee pain over 2 yrs, Rheumatology. 47 (2008)
1392-1396.
[34] P.M. Mills, Y. Wang, F.M. Cicuttini, K. Stoffel, G.W. Stachowiak, P. Podsiadlo, et al.,
Tibio-femoral cartilage defects 3–5 years following arthroscopic partial medial
meniscectomy, Osteoarthritis Cartilage. 16 (2008) 1526-31.
[35] T. Hayami, M. Pickarski, Y. Zhuo, G.A. Wesolowski, G.A. Rodan, L.T. Duong,
Characterization of articular cartilage and subchondral bone changes in the rat
anterior cruciate ligament transection and meniscectomized models of osteoarthritis,
Bone. 38 (2006) 234-243.
[36] D.L. Batiste, A. Kirkley, S. Laverty, L.M.F. Thain, A.R. Spouge, D.W. Holdsworth, Ex
vivo characterization of articular cartilage and bone lesions in a rabbit ACL
transection model of osteoarthritis using MRI and micro-CT, Osteoarthritis Cartilage.
12 (2004) 986-996.
[37] P. Podsiadlo, F.M. Cicuttini, M. Wolski, G.W. Stachowiak, A.E. Wluka, Trabecular
bone texture detected by plain radiography is associated with an increased risk of
knee replacement in patients with osteoarthritis: a 6 year prospective follow up
study, Osteoarthritis Cartilage. 22 (2014) 71-75.
[38] S.K. Tanamas, A.E. Wluka, J.P. Pelletier, J.M. Pelletier, F. Abram, P.A. Berry, et al.,
Bone marrow lesions in people with knee osteoarthritis predict progression of
disease and joint replacement: a longitudinal study, Rheumatology. 49 (2010) 2413-
2419.

23
[39] T. Woloszynski, P. Podsiadlo, G.W. Stachowiak, M. Kurzynski, A signature
dissimilarity measure for trabecular bone texture in knee radiographs, Med. Phys. 37
(2010) 2030-2042.
[40] T. Woloszynski, P. Podsiadlo, G.W. Stachowiak, M. Kurzynski, L.S. Lohmander, M.
Englund, Prediction of progression of radiographic knee osteoarthritis using tibial
trabecular bone texture, Arthritis Rheum. 64 (2012) 688-695.
[41] T. Woloszynski, P. Podsiadlo, G. Stachowiak, M. Kurzynski, A dissimilarity-based
multiple classifier system for trabecular bone texture in detection and prediction of
progression of knee osteoarthritis, P. I. Mech. Eng. H. 226 (2012) 887-894.
[42] T. Woloszynski, M. Kurzynski, P. Podsiadlo, G.W. Stachowiak, A measure of
competence based on random classification for dynamic ensemble selection, Inf.
Fusion. 13 (2012) 207-213.

24
 (a)

(b)
Fig. 1. X-ray images of (a) knee and (b) hand

with selected TB texture ROIs (white squares).

Hand image adapted from OAI database.

25
Fig. 2. A schematic illustration of the VOT method: (a) a search region that moves across the
image, (b) a log–log plot, (c) a rose plot of Hurst coefficients and (d) texture parameters
calculated from the plot.

26
 (a) (b)

(c) (d)

(e) (f)
Fig. 3. X-ray images of (a) normal and (b) osteoarthritic knee joints

with marked ROIs (white squares), corresponding (c, d) bone texture

images, and (e, f) rose plots of Hurst coefficients, along with ellipses

fitted, at scale of 0.55 mm obtained for the regions.

27
Fig. 4. Example of grade 2 and grade 3 cartilage defects
visible on a MRI scan. Image adapted from [33].

28
 (a) (b)
Fig. 5. X-ray images of (a) normal and (b) osteoarthritic DIP

joints with marked bone texture regions (20 × 20 pixels).

Images adapted from [27].

29
 Selected TB bone texture ROIs

Measurement of distances between TB texture images

Generation of a classifier ensemble

Selection of accurate classifiers

Classification of the ROIs

Classified ROIs

Fig. 6. A schematic illustration of the DMC system for classification of TB bone textures.

30
Fig. 7. Example of a bone roughness report from ReadMyXray website.

31