An Algorithmic Approach for Quantitative Motion Artefact Grading in

HRpQCT Medical Imaging

Thomas A. Cox

1 a

, Sasan Mahmoodi

1 b

, Elizabeth M. Curtis

2 c

, Nicholas R. Fuggle

2 d

Rebecca J. Moon

2,4 e

, Kate A. Ward

2 f

, Leo D. Westbury

2 g

and Nicholas C. Harvey

2,3 h

Faculty of Engineering and Physical Sciences, Electronics and Computer Science, University of Southampton,

University Road, Southampton, U.K.

MRC Lifecourse Epidemiology Centre, University of Southampton, Southampton, U.K.

National Institute for Health Research (NIHR) Southampton Biomedical Research Centre, University of Southampton,

and University Hospital Southampton NHS Foundation Trust, U.K.

Paediatric Endocrinology, Southampton Children’s Hospital, University Hospital Southampton NHS Foundation Trust,

Southampton, U.K.

Keywords:

Medical Imaging, HRpQCT, High Resolution Peripheral Computed Tomography, Computed Tomography,

Motion Artefact, Artefact Detection.

Abstract:

High Resolution Peripheral Quantitative Computed Tomography (HRpQCT) is a modern form of medical

imaging that is used to extract detailed internal texture and structure information from non-invasive scans.

This greater resolution means HRpQCT images are more vulnerable to motion artefact than other existing

bone imaging methods. Current practice is for scan images to be manually reviewed and graded on a one to

ﬁve scale for movement artefact, where analysis of scans with the most severe grades of movement artefact

may not be possible. Various approaches to automatically detecting motion artefact in HRpQCT images

have been described, but these typically rely on classifying scans based on the qualitative manual gradings

instead of determining the amount of artefact. This paper describes research into quantitatively calculating

the degree of motion affecting an HRpQCT. This is approached by analysing the jumps and shifts present

in the raw projection data produced by the HRpQCT instrument scanner, rather than using the reconstructed

cross-sectional images. The motivation and methods of this approach are described, and results are provided,

along with comparisons to existing work.

1 BACKGROUND

There are various radiographic clinical methods

which allow clinicians to understand overall bone

health and detect and diagnose osteoporosis, a con-

dition characterized by low bone mineral density

and microarchitectural deterioration in bone struc-

ture. The current gold standard method is using Dual-

Energy X-ray Absorptiometry (DXA), which is used

to calculate areal bone mineral density (BMD). This

https://orcid.org/0000-0002-1343-8306

https://orcid.org/0000-0003-2507-659X

https://orcid.org/0000-0002-5147-0550

https://orcid.org/0000-0001-5463-2255

https://orcid.org/0000-0003-2334-2284

https://orcid.org/0000-0001-7034-6750

https://orcid.org/0009-0008-5853-8096

https://orcid.org/0000-0002-8194-2512

measurement is currently used in the deﬁnition of os-

teoporosis, but provides no detail on the microstruc-

ture of bone. High Resolution Peripheral Qualita-

tive Tomography (HRpQCT) is a modern computed

tomography (CT) technique for acquiring highly de-

tailed structural and texture information of bones. In

contrast to DXA, which generates only a two dimen-

sional image of the skeleton, HRpQCT provides a

three dimensional reconstruction of the internal struc-

ture of the subjects bones, which is used to derive

further quantitative parameters. A two dimensional

cross-sectional slice from a HRpQCT scan is shown

in ﬁgure 1. The unparalleled resolution and detail of

HRpQCT scans comes at the cost of high sensitiv-

ity to motion artefact. Motion artefact in CT scans

is quite different than what we traditionally think of

as motion artefact in regular images, such as those

taken with a standard camera. Whilst motion during

Cox, T., Mahmoodi, S., Curtis, E., Fuggle, N., Moon, R., Ward, K., Westbury, L. and Harvey, N.

An Algorithmic Approach for Quantitative Motion Artefact Grading in HRpQCT Medical Imaging.

DOI: 10.5220/0012434900003654

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 833-840

ISBN: 978-989-758-684-2; ISSN: 2184-4313

833

traditional photography may appear as a blurring or

smearing of the subject, motion artefact in CT scans

appears as tangential streaking and a distorted view of

the subject. Figure 2 shows an example of this.

Figure 1: An example of a single 2D slice from a Tibial

HRpQCT scan. The bright white lines around the edge of

the bones is the cortex, while inside of this the trabecular

structure can be seen.

Currently motion artefact in HRpQCT scans is

manually graded using a scale from one to ﬁve. A

grade of one represents no motion artefact, and a

grade of ﬁve means the scan has an extreme amount

of motion artefact. According to the standard oper-

ating procedure provided by the hardware manufac-

turer (Scanco Medical), it is recommended that scans

with a motion grade of four or ﬁve are excluded from

analysis in research studies (Laib, 2023). Previous

research suggests that there is a a high level of agree-

ment between trained manual graders (Spearman cor-

relation of ρ

= 0.85), but it remains a subjective mea-

sure, and does not provide a continuous scale of arte-

fact severity (Pauchard et al., 2012).

The negative impact of motion artefact on

HRpQCT parameters has been investigated by both in

the scaphoid (Benedikt et al., 2023), and in the tibia

(Pauchard et al., 2011). Benedikt et al. compared

scaphoid bone parameters derived from HRpQCT

scans of the same patients with and without move-

ment, while Pauchard et al. compared images of ca-

daveric tibiae using a machine to precisely induce

movement. Both studies were able to demonstrate

signiﬁcant deviation in HRpQCT derived parameters,

highlighting the importance of considering motion

artefact while examining these scans.

There has been a wealth of previous research

into automatic motion artefact detection in HRpQCT

scans, and in the ﬁeld of motion correction for CT

scans in general. Walle et al. constructed and trained

Figure 2: An example of a single 2D slice from a tibial

HRpQCT with a high level motion artefact. This image

has been manually graded with a motion artefact score of

5. Streaking can clearly be seen across the image tangential

to the top and bottom of the bones. Additionally the cortex

of the tibia can be seen to be distorted as it does not connect

together at the top and bottom correctly. Compare to ﬁgure

1 for an image without motion artefact.

a deep convolutional neural network trained on 3312

2D slices taken from 414 manually graded HRpQCT

scans. This model was able to successfully classify

and differentiate scans graded less than three, from

those graded above three which would require exclu-

sion or a repeated scan. Walle et al. calculated the

f1 score, precision and recall of this model as 86.8%,

87.5% and 86.7% respectively. However, the authors

recognise that the performance of this model was

much lower when classifying the scans by grade with

an f1 score of 43.4% due to their model often mis-

classifying by one artefact grade (Walle et al., 2023).

The subjective nature of the manual grading makes

it impossible for a network trained using that data to

provide a completely objective motion artefact score.

Other researchers have been using convectional

neural networks to classify artefact in other types of

CT images. In a study identifying motion in head

CT images Liu et al. developed a graph based ap-

proach whereby two dimensional slices of cranial CT

scans were converted into a complex graph and then

used to train a convolution neural network to clas-

sify scans into those affected by motion artefact or

unaffected. From the construction of these graphs,

they demonstrated the signiﬁcant differences in graph

node clustering and degree between the two groups

of scans, and when trained they showed the model

outperformed traditional pixel based learning meth-

ods (Liu et al., 2022).

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

834

Throughout this paper, we will describe our novel

approach for detecting HRpQCT motion artefact.

First we provide our motivation for a sinogram based

detection method, then we describe our algorithm

and implementation using a U-Net architecture. Fi-

nally we discuss our preliminary results, and share our

plans for future work.

2 ANALYSING HRPQCT

SINOGRAMS

2.1 The Radon Transform

Before a CT image can be evaluated, the raw projec-

tion data needs to be processed and transformed to

produce a cross sectional slice of the region of in-

terest. CT instruments are able to do this by emit-

ting x-rays through the region of interest at many dif-

ferent angles measuring the attenuation of the x-ray

(the loss of energy) caused by the internal parts of

the object at each angle. At each of these different

angles the x-rays will pass through different planes

through the subject, and result in different attenua-

tions or projections of these planes. By combining

these projections by using the Radon transform (also

known as back projection) the scanner can produce a

reconstructed cross-sectional image of the region of

interest such that the internal structure of their bones

can be analysed. The raw projection data from a CT

scan is known as a sinogram, as off-centre objects in

the subject will appear to oscillate as the scanner ro-

tates around them. An example sinogram from a sin-

gle 2D slice is shown in Figure 3. The Xtreme CT II

scanner (developed by Scanco Medical in Brutteslein,

Switzerland) takes all 110 2D slices over an area of

the subject simultaneously and gradually rotates 180

◦

around the subject, constructing a sinogram for each

slice row by row. Because of this, each row of the

sinogram represents a projection taken of the subject

from a given angle. In Figure 3 there are three distinct

objects: the soft tissue of the leg appearing in light

grey, and inside of this there are two darker cross-

ing sections representing the bones. The wider sec-

tion is the tibia and the narrower section is the ﬁbula.

Towards the top of the scan, the tibia and ﬁbula ap-

pear to overlap as those at the angles those rows were

captured at the x-rays passed through both bones. In

other words, the scan began with a side on view of the

leg but about half way down the scan, the tibia and

ﬁbula stop overlapping, and in each row there are two

distinct darker patches where there is bone. This is

because after the scanner has rotated halfway through

the scan, x-rays either pass though one bone or the

other, effectively meaning that once the scanner has

rotated halfway through the scan (about 90

◦

) it is pro-

jecting a side on view of the leg. Once the scanner has

ﬁnished rotating around the subject and has all 110

full sinograms, it applies the Radon transform to each

one resulting in the cross sectional HRpQCT image

shown in Figure 2. The Radon transform effectively

maps and combines each row of the sinogram onto a

different diameter of a circle, reconstructing the im-

age into a cross section. It is able to extract depth in-

formation by the amplitudes of the vertical sine waves

in the sinogram. If there was a circular object directly

at the centre of the scan, it would remain in the same

horizontal position in each row of the sinogram. The

further away from the centre a circular object is, the

greater its position will vary in different rows of a

sinogram.

Additional research that deals with detecting mo-

tion artefact in HRpQCT scans focuses on properties

of these sinograms, as in these images, motion arte-

fact is much clearer. As each angle of the sinogram is

taken sequentially, the vertical axis can be treated as

time. Therefore, if a subject moves half way through a

scan, the sinogram should show motion artefact in the

rows in the middle of the sinogram. Figure 3 shows

an example of a sinogram with motion artefact. In

this sinogram the subject has moved their leg near the

beginning of the scan, so when the scanner moves on

to subsequent angles, the positions of their soft tissue

and bones seem shifted to the side. The single jump

shown in Figure 3 is what causes the artefact shown in

Figure 2, as the rows in the radon transform no longer

agree on where objects are, causing streaking and dis-

connects in the edges of the bones.

2.2 Artefact Detection

By inspecting sinograms, motion artefact is much eas-

ier to understand. Artefact is seen simply as the

amount that the sinogram jumps or shifts, which be-

gins to provide a way of quantitatively measuring the

amount of motion in a HRpQCT image. Pauchard et

al. have proposed a method to automatically quantify

motion artefact based on the ﬁrst and second moments

of the sinogram (Pauchard et al., 2011). The authors

show that these moments which they use to calculate

the in-plane translation during the scan are accurate

when compared to experiments where a scan is taken

of a sample with controlled motion.

Sode et al. also use HRpQCT sinograms to im-

plement a quantitative way of measuring the motion

artefact (Sode et al., 2011). The authors exploited

the nature of the sinogram in order to compare the

ﬁrst and last lines and determine the net difference.

An Algorithmic Approach for Quantitative Motion Artefact Grading in HRpQCT Medical Imaging

835

Figure 3: An example sinogram from a slice of a tibial HRpQCT scan with motion artefact. Each row of the scan represents

an angle where X-rays were passed through the subject. The pixels in each row show the projection created by a plane of

X-rays passing though the subjects leg at a given angle. Here a clear jump can be seen roughly a quarter of the way down the

scan, this represents where the subject has moved and their leg is now in a different location. This jump is the cause of the

motion artefact seen in Figure 2.

This is effective as the HRpQCT scanner emits x-rays

through a 200

◦

range. As such, the rows at the top of

the sinogram, from 0

◦

to 20

◦

, should mirror those at

the bottom, from 180

◦

to 200

◦

, as they represent pro-

jections of the subject from the opposite side. Given

this, assuming the subject has not moved during the

duration of the scan, the difference between the top

and bottom rows of the sinogram should be minimal.

Conversely, if the subject has shifted, these will be

noticeably different. This process is, however, com-

plicated by the nature in which the Xtreme CT II scan-

ner emits x-rays; the cone beam conﬁguration causes

slight magniﬁcation of the subject to different extents

at 0

◦

and 180

◦

ranges. Sode et al. were able to correct

this by transforming the sinogram from the cone beam

domain (where x-rays are emitted in a cone from a

single point) to the parallel beam domain, as if the x-

rays were emitted equally spaced and in parallel. Af-

ter performing this transformation, they showed that

their Quantitative Motion Estimates (QMEs) derived

from the sum squared of the difference between the

top and bottom rows of the sinogram correlate with

the manual grades given to HRpQCT images. For ex-

ample, they show that HRpQCT images with a grade

of ﬁve have signiﬁcantly increased QMEs that those

with lower artefact grades.

This method is effective at detecting overall mo-

tion during the scan, however it has two major draw

backs. First, it only determines the net motion be-

tween the start and end of the scan. Thus if a sub-

ject was to twitch, brieﬂy moving and then moving

back, it is possible that this method would result in

a misleading low to zero QME, as their method only

considers the net difference between the ﬁrst and last

lines of the sinogram. Second, this method provides

no means of detecting where or when the jumps oc-

curred during the scan, and therefore does not assist

in correcting the artefact. Sode et al. address the fact

that while there is a difference in QME derived from

scans with a manual quality grade of ﬁve, the QMEs

overlap for scans graded between one and four. The

authors attribute this in part to the subjective nature of

manual grades, but we hypothesis that it is also caused

by the global nature of the way their QMEs are calcu-

lated. In Section 3 we discuss our proposed strategy

for detecting all motion that occurs throughout a sino-

gram.

3 IMPLEMENTATION

We used HRpQCT scans taken as part of the Mater-

nal Vitamin D Osteoporosis Study (MAVIDOS) (Har-

vey et al., 2012) and Southampton Women’s Survey

(SWS) (Inskip et al., 2006) studies in order to investi-

gate methods of detecting motion artefact from sino-

gram data. We initially tried to detect motion arte-

fact entirely algorithmically in the sinogram by us-

ing a combination of edge detection and the Hough

transform. However we found that the edges were

simply too faint in comparison to the CT noise to ac-

curately detect this. This method relied on detecting

horizontal lines across the image where ”jumps” had

occurred, as seen near the top of Figure 3. Unfor-

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

836

tunately, where jumps occurred, due to the nature of

trabecular bone, there were little to no pixel to pixel

differences between lines around the jump, resulting

in failure in both edge and ridge detection failed to

detect where the jumps occurred. Nonetheless, using

this method in conjunction with the Hough transform

with tuned parameters would often detect small seg-

ments of lines where jumps occurred along the edges

of the bone. Still, we found it difﬁcult to ensure that

these jumps were accurately detected in the presence

of noise without over-smoothing the image and losing

edge information.

Realizing the main indicators of a jump occurring

were most evident along the edge of the bones, we

began working on detecting and tracing the six ma-

jor vertical edges of the image, corresponding to the

edges of the soft tissue, tibia and ﬁbula. Our mo-

tivation was that if we could extract traces of these

sinusoidal edges, we would be able to detect where

anomalies or jumps occurred in each edge. Then hav-

ing detected these anomalies in the same place on a

sufﬁcient number of edges, we would be able to con-

ﬁdently say that a jump had occurred at that location.

With this method, severity and direction of the shift

would also be detected by analysing how much the

edge deviated from its sinusoidal pattern. This ap-

proach has motivated the remainder of our research

into motion artefact detection and correction, as it

should be able to accurately detect where in a sino-

gram jumps have occurred, as well as measure the

severity of each jump independently. This approach

will be beneﬁcial both for automatically calculating

a quantitative measure of motion artefact in the scan,

and to subsequently inform how individual shifts can

be corrected or reduced. Initially we used a simple

canny edge detection algorithm to detect and trace

these six major edges; however, too much noise was

included in the traces to allow for accurate detection

of where anomalies occurred. To solve these prob-

lems, we segmented the sinograms using machine

learning to quickly and accurately ﬁnd the traces of

the edges needed.

3.1 U-Net Segmentation

In order to accurately segment the sinogram images,

we used the state-of-the-art U-Net architecture. U-

Net is a fully convolutional neural network that is

widely used for CT and other biomedical imaging

tasks due to its fast training speed and high accu-

racy even with comparatively low amounts of data

(Yin et al., 2022). Before we could begin training

a U-Net model to segment our HRpQCT sinograms,

we split the combined scans from the MAVIDOS and

SWS cohorts equally into a training and testing group.

Then to increase the amount of data we had available,

we decided to sample three random slices from each

HRpQCT scan in the training set and treat them as

separate images. While different slices of the same

scan will be similar, we hypothesised that the differ-

ences caused by viewing a different section of bone

increased size of our dataset without the risk of over-

ﬁtting. This resulted in a training sample of up to

1,110 different sinogram slices. We manually masked

600 of these different sinograms in order to gather

some preliminary results and determine whether U-

Net segmentation would be effective.

Figure 4: Here an overlay of masks predicted by our U-

Net model for the distinct segmented areas in a HRpQCT

sinogram is shown. The areas segmented in this image,

shown in different colours accurately trace out the appro-

priate areas even where motion artefact has occurred during

the scan. The scan shown here was manually assigned a

motion artefact grade of ﬁve.

With this training sample, we constructed a U-

Net model in python with Keras using the segmen-

tation methods library. As a proof of concept we used

the default U-Net construction of 4 encoding and 4

decoding layers. We deﬁned the loss function us-

ing a combination of categorical focal loss and dice

loss, used a softmax activation function to assign each

pixel to its appropriate mask/class. Once the net-

work was constructed, we trained it on the labeled

HRpQCT sinograms for 80 epochs with a learning

rate of 0.00001 and a batch size of 10 images. Af-

ter 80 epochs of training the network reached a loss

of 0.0512, an intersection over union score of 0.9154

and an f1-score of 0.9492. Once we were satisﬁed

that it was accurate, we began predicting masks for

the sinograms in the test set and were able to conﬁrm

that the network was automatically masking images

correctly. An example of the output masks that the

model predicted for an image with high motion arte-

fact is shown below in Figure 4. As seen in the ﬁgure

An Algorithmic Approach for Quantitative Motion Artefact Grading in HRpQCT Medical Imaging

837

we segmented the sinogram into ﬁve main areas: left

and right soft tissue, central soft tissue between the

tibia and ﬁbula, and where the tibia and ﬁbula overlap

at the top and bottom of the sinogram. By segmenting

these areas, we ensured that every edge was captured

by at least one of these segments, ensuring that we

can capture any distortion caused by motion artefacts

in these masks.

From visual inspection of the results produced by

our U-Net model for the testing set, we determined

that jumps and shifts in the sinogram were being cap-

tured by the masks. Our next challenge was devel-

oping an algorithm to piece together traces of the six

major edges in the image from these masks. We be-

gan by using a simple Canny ﬁlter to extract the edges

from each of the masks. Then after combining the

edges into one image, such as the image shown in Fig-

ure 5, our algorithm detects the points where each of

the masks get closest to each other to ﬁnd the points

where the edges of the tibia and ﬁbula would intersect

in the image. Then it splices together the appropriate

parts of each of the edges around these points in or-

der to construct the six major edge traces which are

displayed in Figure 6. This ﬁgure in fact shows the

distance between each of the traces and the vertical

centre line of the sinogram.

Figure 5: This ﬁgure shows a combined image of all of the

edges detected from a masks produced by U-Net. In order

to convert these edges into the actual edges of the subjects

leg, tibia and ﬁbula, our algorithm detects where the edges

would cross and splices together edges at these points to

construct the edge signals shown in 6.

One disadvantage of the method used to segment

the sinograms is that our algorithm will fail to capture

any motion artefact in the unlikely case that it occurs

in a sinogram exactly where the edges of the tibia and

ﬁbula intersect. In the future it may be possible to

retrain the U-Net with images where the soft tissue,

tibia and ﬁbula are segmented more directly, but it

would require manually tracing some edges multiple

times plus more segmentation regions where the tibia

and ﬁbula overlap. For the time being we have ac-

cepted this as a shortcoming of our current approach,

but any net shift in these regions should still be de-

tected.

Figure 6: This ﬁgure shows the edge traces captured from

the masks shown in Figure 4. Each different coloured

curve in the image plots the distance between each of the

traces and the vertical center line of the sinogram, where

the horizontal-axis shows the row of the sinogram, and the

vertical axis shows the distance . The lines are constant

at the beginning and end of the plot where the traces were

cropped in order to make the edge detection more robust.

4 PRELIMINARY RESULTS

As shown in Figure 6, shifts and jumps in the the

sinogram slice of an HRpQCT scan can be captured

via examining edge traces, even when to the edge

is very small. Although the original HRpQCT scan

shown in Figure 4 has been assigned a motion artefact

grade of ﬁve and has pronounced discontinuities, the

peaks and troughs detected from the edge traces are

quite smooth. Nonetheless we implemented a sim-

ple anomaly detection algorithm, where we detected

where local peaks and troughs occurred in each sig-

nal and measured their prominences. Then to create

a simple quantitative measure of motion artefact that

occurred in the image, we summed the prominences

at each location across all six of the edges, and set

a threshold to determine that if more than one edge

agreed, there was a shift or jump at a particular loca-

tion. This gives a list of locations and prominences

for where motion artefact had occurred in the sino-

gram, which we summed to derive an estimate of how

much motion had occurred in a sinogram slice. Sub-

sequently for 600 images in the test set, the U-Net

model was used to predict masks. These were then

used to capture edge traces and calculate a quanti-

tative motion artefact estimate for them as described

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

838

above. Figure 7 shows the range of grades this algo-

rithm assigned to the test sinograms organized by the

preassigned manual motion artefact grading. While

this ﬁgure does show that our rudimentary algorithm

gives a larger number of high scores to sinograms

from images with a higher manual motion artefact

grade, it also suggests that some of the images with

high manual grade motion artefact are given very low

motion artefact estimates. We had hoped to produce a

similar diagram to that by Sode et al. showing a cor-

relation between our motion estimates and the man-

ual grades, however, it appears as though many in-

stances of motion artefact are not being appropriately

detected. We are hopeful that through reﬁnement of

our algorithm, we will be able to use the edge traces

we extracted to better detect motion artefact.

Figure 7: This plot shows the spread and means of motion

artefact estimates our algorithm assigned to sinograms from

the testing set, organized by their manually assigned quality

scores. This shows an association between our calculated

quantitative measure of artefact in scans with the manually

assigned grade. Scans with a high motion artefact grade (5)

receive higher estimated scores.

5 DISCUSSION

While our current scoring algorithm fails to fully cor-

relate with the manual motion artefact grades, it does

still somewhat capture the levels of motion artefact,

generally giving scans with motion artefact grade ﬁve

higher scores. The main drawback of of current al-

gorithm is assigning low scores to scans that contain

high levels of motion artefact. From inspection of

the algorithm’s results, it is clear that the main cause

of this miss-classiﬁcation are cases where the current

anomaly detection method fails to detect jumps in the

sinogram. While we hope to improve by retraining

our U-Net model and improving our edge detection,

even our current algorithm shows promising results in

cases where motion artefact is scored appropriately.

Visual inspection of the graded sinograms conﬁrms

that where our algorithm assigns a high motion arte-

fact score, artefact is detected where jumps and shifts

occur in the sinograms. This feature of our algorithm

forms the basis for subsequent approaches to correct

motion artefact in HRpQCT scans.

Figure 8: The correlation between QMEs and manual grad-

ing of scans our testing dataset are shown in this plot. For

each of the ﬁve categories of manually assigned scores, the

spread and standard deviation of QMEs in each of those

categories are shown. QMEs shown in this plot where cal-

culated using the methods and algorithm described by Sode

et al. (Sode et al., 2011).

In order to compare our results, we re-

implemented the QME algorithm described by Sode

et Al. (Sode et al., 2011) to classify the motion arte-

fact in our dataset. However, their methods lower per-

formance when applied to our dataset, displaying to a

weaker association between high quantitative scores

of scans with a high manual grade than in our algo-

rithmic results. Our numerical results clearly demon-

strate a superior performance for our algorithm in

motion artefact detection in comparison with our re-

implementation of the previous work done by Sode et

al. especially in scans with motion artefact of grade

ﬁve, our results show a signiﬁcant separation between

motion artifacts of grade ﬁve with the other motion

artefact levels. This could be attributed to two fac-

tors. First, as part of their approach the HRpQCT

sinogram data is converted into a parallel beam format

using proprietary code provided by Scanco Medical.

This ensures that differences between the projections

from opposite ends of the sinogram are minimized

and the QMEs can be calculated correctly. When re-

implementing this we were unable to gain access to

this proprietary code. Instead we coded our own con-

version to the parallel beam format manually, which

may have introduced some differences in the results.

Second, the dataset used by Sode et al. is composed

An Algorithmic Approach for Quantitative Motion Artefact Grading in HRpQCT Medical Imaging

839

of middle-aged adults (Sode et al., 2011), whereas

our dataset included only children under ten years

old. Because of this, our dataset contains many ex-

amples of extreme motion artefact, which the QMEs

calculated by Sode et al. in their paper may not be

able to accurately depict. Nonetheless we would still

like to highlight the potential of clinical applications

of our algorithm. Not only would clinicians be able

to immediately repeat a scan in which motion arte-

fact has been automatically detected, but additionally,

scans were artefact may be unavoidable, such as those

with young children, may be corrected and included

in study.

6 FUTURE WORK

Our preliminary results have shown that our al-

gorithmic approach to detecting motion artefact in

HRpQCT scans can distinguish between scans with

high and low levels of motion artefact. However,

as described above, there are some limitations, as

our current model cannot accurately distinguish the

amounts of motion contained in scans graded one to

four. In our further research we plan to increase our

training samples and reﬁne our U-Net model to im-

prove our sinogram segmentation and make it more

robust to noise. We predict that this will improve the

accuracy of our detection of jumps in the sinogram

data, and therefore increase the accuracy of our nu-

merical motion artefact measures. We plan to con-

tinue to iterate on this approach to ensure that our

quantitative results accurately reﬂect the amount of

distortion caused by motion in the image, and can be

used to inform the accuracy of HRpQCT parameters

derived from artefact affected scans.

ACKNOWLEDGEMENTS

TC is supported through a doctoral studentship at

the University of Southampton funded jointly by

the MRC Lifecourse Epidemiology Centre and the

Institute for Life Sciences. This work was sup-

ported by MRC [MC PC 21003; MC PC 21001],

and National Institute for Health Research (NIHR)

Southampton Biomedical Research Centre, Uni-

versity of Southampton, and University Hospital

Southampton NHS Foundation Trust, UK.

REFERENCES

Benedikt, S., Horling, L., Stock, K., Degenhart, G., Pal-

lua, J., Schmidle, G., and Arora, R. (2023). The im-

pact of motion induced artifacts in the evaluation of

HR-pQCT scans of the scaphoid bone: an assessment

of inter- and intraobserver variability and quantitative

parameters. Quant. Imaging Med. Surg., 13(3):1336–

1349.

Harvey, N. C., Javaid, K., Bishop, N., Kennedy, S., Papa-

georghiou, A. T., Fraser, R., Gandhi, S. V., Schoen-

makers, I., Prentice, A., and Cooper, C. (2012).

MAVIDOS maternal vitamin D osteoporosis study:

study protocol for a randomized controlled trial. the

MAVIDOS study group. Trials, 13:13.

Inskip, H. M., Godfrey, K. M., Robinson, S. M., Law, C. M.,

Barker, D. J. P., Cooper, C., and SWS Study Group

(2006). Cohort proﬁle: The southampton women’s

survey. Int. J. Epidemiol., 35(1):42–48.

Laib, A. (2023). Sop quality grading.

Liu, Y., Wen, T., Sun, W., Liu, Z., Song, X., He, X., Zhang,

S., and Wu, Z. (2022). Graph-based motion artifacts

detection method from head computed tomography

images. Sensors, 22(15):5666.

Pauchard, Y., Ayres, F. J., and Boyd, S. K. (2011). Auto-

mated quantiﬁcation of three-dimensional subject mo-

tion to monitor image quality in high-resolution pe-

ripheral quantitative computed tomography. Physics

in Medicine & Biology, 56(20):6523.

Pauchard, Y., Liphardt, A.-M., Macdonald, H. M., Hanley,

D. A., and Boyd, S. K. (2012). Quality control for

bone quality parameters affected by subject motion in

high-resolution peripheral quantitative computed to-

mography. Bone, 50(6):1304–1310.

Sode, M., Burghardt, A. J., Pialat, J.-B., Link, T. M., and

Majumdar, S. (2011). Quantitative characterization of

subject motion in hr-pqct images of the distal radius

and tibia. Bone, 48(6):1291–1297.

Walle, M., Eggemann, D., Atkins, P. R., Kendall, J. J.,

Stock, K., M

uller, R., and Collins, C. J. (2023). Mo-

tion grading of high-resolution quantitative computed

tomography supported by deep convolutional neural

networks. Bone, 166:116607.

Yin, X.-X., Sun, L., Fu, Y., Lu, R., and Zhang, Y. (2022).

U-Net-based medical image segmentation. J. Healthc.

Eng., 2022:4189781.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

840