Automatic Spine Segmentation in CT Scans

Gabor Revy

, Daniel Hadhazi

and Gabor Hullam

Department of Measurement and Information Systems,

Muegyetem rkp. 3., H-1111 Budapest, Hungary

Keywords:

Spine Segmentation, Image Processing, CT, Dynamic Programming.

Abstract:

The segmentation of the spine can be an essential step in computer-aided diagnosis. Current methods aiming

to handle this problem generally employ an explicit model of some type. However, to create an adequately

robust model, a high amount of properly labeled diverse data is required. This is not always accessible. In this

research, we suggest an explicit model-free algorithm for spine segmentation. Our approach utilizes expert

algorithms that are built on medical expert knowledge to create a spine segmentation from thoracic CT scans.

Our system achieves an IoU (intersection over union) value of 0.7103±0.051 (mean±std) and a DSC (Dice

similarity coefﬁcient) of 0.8295±0.0343 on a subset of the CTSpine1K dataset.

1 INTRODUCTION

Algorithms aiming the higher level of automatization

of computer-aided detection (CADe) and diagnosis

(CADx) systems have been an important area of re-

search in recent decades. These systems can help

to make the detection of lesions and the planning of

medical interventions preventive, more accurate and

faster.

The aim of our research was to create an auto-

matic system to localize and segment the spine on CT

scans of full thoracic ﬁeld of view. This algorithm

is part of a larger system that segments bones, com-

plemented by the segmentation of the sternum and the

ribs. Based on the segmentation of the rib cage we can

detect several types of lesions or design medical inter-

ventions. Since the whole rib cage is to be segmented,

it is not critical if the segmentation of the spine runs

into the ribs. However, it is important to completely

segment the vertebral body and the spinous process.

There are other applications of spine segmenta-

tion. In the case of image-guided spinal surgery the

main focus is on the spine, therefore the 3D segmenta-

tion is also done differently. The CT scans are created

in a speciﬁc way: with a limited ﬁeld of view and with

the spine in the center. Then the segmentation should

not only consider the spine as a whole, but also dis-

tinguish the individual vertebrae. Therefore, such al-

https://orcid.org/0000-0002-4547-3923

https://orcid.org/0000-0002-6233-5530

https://orcid.org/0000-0002-4765-2351

gorithms must utilize an explicit model. Also, these

algorithms often require manual seed points in order

to function. In our work, however, we aimed to cre-

ate a simple but robust, explicit model-free algorithm

that does not involve manual intervention. Even in

such a simplistic case, there are still many challenges.

First, the varying quality of CT scans can cause difﬁ-

culties. For example, in some cases the vertebra falls

apart into many small segments of high density with

low-density material in between. Second, the con-

trast agent, which normally facilitates the segmenta-

tion process, complicates the applicability of classi-

cal image processing techniques in this case, as it can

have a high density, similar to bones.

This paper is structured as follows. Section 2 pro-

vides an overview on the relevant previous works.

Section 3 describes the process of the segmentation.

In subsection 3.1 the estimation of the center line of

the spine is described, subsection 3.2 details the cre-

ation of the main segmentation and in subsection 3.3

the reﬁnement of the segmentation mask is detailed.

In section 4 the proposed algorithm is evaluated on

a subset of the publicly available CTSpine1K (Deng

et al., 2021) dataset. The results of our work are sum-

marized in section 5.

The steps of the segmentation algorithm are vi-

sualized using the LIDC-IDRI (Armato et al., 2011)

dataset. The parameters of the proposed algorithms

were also tuned based on this database and expert

knowledge. This dataset contains 1308 thoracic CT

scans of diagnostic and lung cancer screening of 1010

patients.

Revy, G., Hadhazi, D. and Hullam, G.

Automatic Spine Segmentation in CT Scans.

DOI: 10.5220/0011660000003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 2: BIOIMAGING, pages 86-93

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

2 RELATED WORK

There are several approaches towards spine segmen-

tation in CT scans. The input can be a CT scan of

full thoracic ﬁeld of view or the ﬁeld of view may be

limited to the spine. For some medical procedures it

might be necessary to label the spine at vertebra-level

(multi-class labeling). In other cases, it is sufﬁcient

only to label the spine (binary labeling).

Previously, the segmentation of the spine has

mainly been performed by ﬁtting a shape prior and

deforming it to the actual spine. Athertya et al. used

a set of feature markers from the CT scan to cre-

ate an initial contour for an active contour model

(ACM) (Athertya and Kumar, 2016). This is further

reﬁned by utilizing a fuzzy corner metric, which is

based on the image intensity. Castro-Mateos et al.

utilized a type of statistical shape models, statisti-

cal interspace models (SIMs) to signiﬁcantly reduce

the overlap between the different vertebrae (Castro-

Mateos et al., 2015). Ibragimov et al. proposed a

multi-energy segmentation framework, which com-

bines landmark detection and shape-based segmenta-

tion (Ibragimov et al., 2017). In their work, landmarks

are used for the non-rigid deformation of their model.

They utilized Laplacian coordinates to ﬁnd the opti-

mal deformation.

With the emergence of machine learning in image

processing, a signiﬁcant amount of data-driven learn-

ing algorithms have been proposed. Sekuboyina et al.

utilized two neural networks: one for localisation and

another for segmentation (Sekuboyina et al., 2017).

A 2D attention network provided a low-resolution lo-

calisation of the spine. Then, a 2D-3D U-shaped

network generated high-resolution binary segmenta-

tions. Lessmann et al. proposed an iterative vertebra

segmentation approach utilizing a fully convolutional

network to segment and label vertebrae one after the

other (Lessmann et al., 2019). The network is com-

bined with a memory component in order to retain

information about the vertebrae already segmented.

As it can be seen, most of the algorithms require a

lot of labeled samples to create a model. This is usu-

ally a bottleneck in realization of the methods. Fur-

thermore, the number of manually labeled CT scans

with full thoracic view is limited. Our method is ex-

plicit model-free thus it does not require voxel-level

segmented samples.

3 METHOD

The segmentation pipeline works as follows. First,

(1) a center line of the spine is estimated based on the

Hounsﬁeld unit values using a convolution-based al-

gorithm. From this centerline, the (2) spine is roughly

segmented by utilizing morphological reconstruction.

The (3) boundaries of the resulting segmentation are

truncated based on anatomical prior knowledge. Fi-

nally, the (4) contour of the vertebral body is reﬁned

by a dynamic programming-based parabola ﬁtting al-

gorithm.

3.1 Spine Center Line Estimation

In the ﬁrst step, a rough bone mask (I

thresh

) is cre-

ated. This is obtained by thresholding the radioden-

sity (Hounsﬁeld unit) values. The effects of different

thresholds were analyzed in a series of experiments,

using CT scans with and without contrast. Based on

these results a threshold value of 150 was selected in

order to ensure, that the bones are extracted even if

the density values are degraded by high amount of

noise. However, this interval overlaps with the density

value of the blood, containing contrast agent. In such

a case, the descending aorta and the heart may also be

segmented and might become merged with the bone

mask. Therefore, in order to avoid such an inclusion,

the segmentation of the descending aorta is subtracted

from the resulting mask. Furthermore, on the axial

slices, the region that is anterior relative to the de-

scending aorta is removed from the mask. In the next

step, the center of the spine is estimated. Among the

sagittal slices the one with the most segmented vox-

els is selected. The index of this slice estimates the

column where the spine is located on the axial slices.

Furthermore, a spine region localization map (M) is

created (see Figure 1), searching for disk-like shapes.

This map is created by convolving the segmentation

(Figure 1b) with the upper (d

) and the lower (d

) half

of a disk separately and taking the element-wise min-

imum of the results:

(x, y) = (x

+ y

≤ r

) ∧ (y > 0) (1)

(x, y) = (x

+ y

≤ r

) ∧ (y < 0) (2)

M = minimum(I

thresh

∗ d

, I

thresh

∗ d

) (3)

The radius of the disk is set to an average-sized verte-

bra. The motivation of utilizing this type of ﬁltering

is that convolving with a disk shape (i.e. without the

separation of the kernel) resulted in many false posi-

tions due to the contrast agent. This can occur in those

CT image slices, where the mask of the descending

aorta is not present thus the anterior part of the ax-

ial mask cannot be removed. The robustness, in this

case, was further improved by applying the convolu-

tion only to the posterior part of the body (i.e. only to

the lower half of the axial slices). To get an estimated

center line of the vertebrae, in each axial slice of the

Automatic Spine Segmentation in CT Scans

ﬁltered image M the point with the highest value is se-

lected. The maximum search is performed in a region

that is deﬁned by the previously estimated column of

the spine center. This search results in a single point

on each slice. Quadratic polynomials are ﬁtted to the

coordinates of the points to ﬁlter out outliers. To en-

sure that the ﬁt is robust with respect to outliers, the

RANSAC (Fischler and Bolles, 1981) method is uti-

lized with L1 loss. The projection of the smoothed

center line is shown in Figure 2.

(a) Input CT slice (b) Bone mask created by

thresholding the input CT

map

(d) Spine region localization

map with overlay

Figure 1: The main steps of the generation of the spine re-

gion localization map for determining a preliminary center

line of the spine. LIDC 0014:50.

3.2 Boundary of the Vertebra

For clarity, the results of the different steps of the al-

gorithm are shown in different colors in Figure 3, and

are referred to in italics in the text. To obtain the

boundary of the vertebra, we start from the bone mask

(red mask) and the center line (green dot C in the mid-

dle) created in the previous step. Several lines are de-

ﬁned, which will determine the bounding region of

the vertebral mask. These lines are deﬁned based on

anatomical prior knowledge approved by medics. To

obtain the end of the spinous process (blue line H3),

maximum intensity projection is performed perpen-

dicular to the sagittal slices followed by thresholding.

This can be computed directly from the bone mask.

Figure 2: Projection of the center line onto a sagittal slice

after polynomial smoothing. Original center points are

marked in blue, whereas the ﬁtted polynomial is the green

line. LIDC 0014.

Since the end of the spinous process is located further

back than the ribs, the bottommost point of the result-

ing segmentation is considered to be the edge of the

spinous process.

In the next step, a basic segmentation mask is

created using morphological reconstruction by dila-

tion (Gonzalez, 2009) with a disk, based on the bone

mask. The seed of the algorithm is created from the

center line, obtained in the previous step. The cen-

ter line only provides a seed region. As long as it

is located within the spine, the morphological recon-

struction applied here can highlight the vertebrae. To

create the seed of the reconstruction the center line

is dilated to ensure that it covers the vertebra. The

algorithm is performed in 2D for each slice. The re-

sulting segmentation mask may overlap with the ribs.

Furthermore, the segmentation may not contain the

spinous process, since it might be detached from the

vertebral body in some slices of the bone mask (as

it can be also seen in Figure 3). This latter prob-

lem was solved by ﬁrst dilating the resulting mask

across the axial slices (M

), intersecting it with the

bone mask (M

), and performing a morphological re-

construction with a limited number of iterations using

the result (M

∩ M

) as the base mask of the recon-

struction. The number of iterations is limited so that

the segmentation does not overlap with nearby struc-

BIOIMAGING 2023 - 10th International Conference on Bioimaging

V1 V2

V5 V6

Figure 3: Segmentation of the vertebra (LIDC 0014:31).

The bone mask (red+yellowish brown+blue) is created by

thresholding the CT slices. Based on this mask a baseline

spine mask is created (yellowish brown+blue). The colored

lines constrain (yellowish brown → blue mask) the bound-

ing region of the vertebrae by utilizing a priori knowledge.

The magenta and the blue lines denote the top of the ver-

tebra and the end of the spinous process, respectively. The

drab line is drawn to the upper third of the part deﬁned by

these two lines. The width of the baseline spine mask is

marked with the two brown lines. These lines are shifted

both directions with the width resulting in the green lines.

The mint green and purple lines denote the intermediate and

ﬁnal center line of the spine. The algorithm is detailed in

subsection 3.2.

tures e.g. smaller vessels near the vertebral body. The

obtained segmentation mask is shown in saffron and

blue in Figure 3.

This reconstructed mask is utilized to obtain the

top of the vertebral body (magenta line H1) in the ax-

ial slices. Around the center line, several columns of

the reconstructed mask are selected in a band. In this

band, the upper part of the segmentation is considered

to be the most anterior part of the vertebral body (see

Figure 3).

A more accurate center column (mint green line

V4) of the vertebral body can be determined as fol-

lows. First, the cumulative sum of the number of seg-

mentation pixels is calculated across the columns in

each axial slice. Then, the median of the cumulative

sum determines the center column of the vertebrae.

Using this center column (mint green line V4), the

maximal width (brown lines V2 and V5) of the ver-

tebral body can be determined. This width is used

later to determine the width of the whole vertebra. To

measure the width of the vertebral body in each slice,

the upper third (drab line H2) of the vertebral body is

utilized. This can be derived from the top of the ver-

tebral body and from the end of the spinous process

(blue line H3) calculated earlier.

Separating the ribs from the spine is a difﬁcult

task: there is often no visible sharp line of separa-

tion. To ensure that the mask contains the transverse

process, rather than the ribs, a simple heuristic is uti-

lized: in both directions from the edge of the vertebral

body, the edges are shifted by the width of the verte-

bral body. This approach deﬁnes two vertical lines

(green lines V1 and V6), from which the segmenta-

tion is discarded outwards.

After this step, a new, more accurate vertebra cen-

ter column (light purple line V3) can be selected by

utilizing the same algorithm as before.

3.3 Parabola Fitting

The presence of the contrast agent in the blood can

cause non-bony parts to be included in the segmenta-

tion. This means, that small vessels and vessel parts

close to the spine may merge with the spine in the seg-

mentation mask. The anterior part of the vertebrae is

semicircular, which can be segmented more robustly

by curve ﬁtting. Thus we designed a parabola ﬁt-

based correction of the contour of the vertebral body

that utilizes dynamic programming.

First, a family of parabolic curve segments is gen-

erated with the same height and different focal lengths

(see Figure 4). Furthermore, a directional gradient

Figure 4: Family of parabolas that are tested to best ﬁt the

contour of the vertebra.

image is produced from the segmentation mask gen-

erated in the previous step. The gradient image is cal-

culated for each axial slice. Its center is deﬁned by

the ﬁnal horizontal center and the bottom line of the

upper third of the vertebral body. This gradient im-

age takes a large absolute value with a negative sign

at the upper parabolic border of the vertebral body.

Using exhaustive search, all the parabolas are tested

with different shifts along the y axis. We select the

parabola along which the sum of the gradient values

normalized by the length of the parabola is minimal.

In the next step, the best-ﬁtting parabolas are ﬁt to the

corresponding axial CT slice using dynamic program-

ming (DP). The algorithm is performed separately in

each slice and has two main inputs. The ﬁrst is the

best-ﬁtting parabola retrieved from the previous step

Automatic Spine Segmentation in CT Scans

and sampled at equal distances along the curve. The

second is a directional gradient image produced in the

same way as in the previous step, but this time, it is

based on the input CT slices:

dir grad

(x) = I

grad

(x)

x − c

||(x − c)||

, (4)

where I

dir grad

(x) is the gradient image of the CT slice

and c is the centerpoint.

The algorithm can be described by the following

recursion:

T [r, q] =

max

′

{T [r

′

, q − 1]−

σ ·



mult

[r] · ||p[q − 1] − c|| − µ

mult

′

] · ||p[q] − c||



}

+ (−1) · I

dir grad

(µ

mult

′

] · (p[q] − c) + c), (5)

where p[q] is the q

point of the ﬁtted parabola, c is

the center of the parabola, σ is a multiplicative penalty

factor, and µ

mult

[k] is the k

multiplication factor that

controls the length of the vector that points from the

center c to the direction of the current point p[q] of the

parabola. Here, T indicates the quality of the multi-

plicative offset r along point q and it is initialized with

zeros: T [row, −1] = 0. As it follows from the recur-

sion, this DP approach tries to balance between align-

ing the contour of the mask to points in the gradient

image that are negative and have a high absolute value

while still maintaining the parabolic shape. In the last

step, the best offset o is selected:

o = argmax

′

(T [r

′

, end]), (6)

where ”end” indexes the last point. The ideal offset

for each point of the parabola can then be traced back

from this offset. Based on the points thus ﬁtted, the

segmentation mask can be reﬁned by morphological

reconstruction. Here, the seed mask is the mask under

the ﬁtted points. To create the base mask for the re-

construction, morphological opening is performed on

the segmentation mask above the ﬁtted points. This

ensures, that components linked by only a few pix-

els are detached so that they can be eliminated from

the ﬁnal segmentation. Figure 7, Figure 8 and Fig-

ure 6 show examples of the results of the algorithm

and Figure 5 shows a full spine segmentation mask in

3D.

4 EVALUATION

Although several spine segmentation datasets were

investigated we chose a dataset that is most relevant

Figure 5: 3D reconstruction of the resulting segmentation.

LIDC 0014.

Figure 6: Coronal and sagittal view of the resulting segmen-

tation on the LIDC 0014 CT scan.

to our target area of application. This means, that the

input CT should be a thoracic CT scan with full tho-

racic ﬁeld of view. Thus, our proposed spine segmen-

tation method was evaluated on the COVID-19 (An

et al., 2020) (Harmon et al., 2020) subdataset of the

CTSpine1K (Deng et al., 2021) dataset retrieved from

TCIA (Clark et al., 2013a). The COVID-19 dataset

consists of unenhanced chest CT scans from 632 pa-

tients with COVID-19 infections at initial point of

care. 20 of these were selected and manually anno-

tated in the CTSpine1K dataset.

Two evaluation metrics were used for the evalua-

tion of the accuracy of the proposed algorithms. In-

tersection over union (IoU) is the ratio of the intersec-

tion and union of the segmentation mask deﬁned by

the algorithm (A) and the manual labeling (B):

IoU(A, B) =

|A ∩ B|

|A ∪ B|

|A ∩ B|

|A| + |B| − |A ∩ B|

(7)

The Dice similarity coefﬁcient (DSC) equals

twice the intersection of the segmentation mask vol-

umes divided by the sum of the volumes:

DSC(A, B) =

2|A ∩ B|

|A| + |B|

2|A ∩ B|

|A ∪ B| + |A ∩ B|

(8)

The results of the evaluation are shown in Table 1.

Furthermore, Figure 9 shows the distribution of the

evaluation results.

BIOIMAGING 2023 - 10th International Conference on Bioimaging

(a) LIDC 0005:81 (b) LIDC 0014:11 (c) LIDC 0014:30 (d) LIDC 0014:32

(e) LIDC 0021:78 (f) LIDC 0022:13 (g) LIDC 0022:65 (h) LIDC 0024:54

(i) LIDC 0024:58 (j) LIDC 0030:106 (k) LIDC 0030:110

Figure 7: Examples, where the dynamic programming-based parabola ﬁtting algorithm improved the spine segmentation. The

best-ﬁtting parabola is colored red and the DP-ﬁtted points are marked in green. The resulting spine segmentation is marked

in purple, while the excluded part is marked in blue.

(a) LIDC 0005:31 (b) LIDC 0018:45 (c) LIDC 0018:92 (d) LIDC 0024:55

Figure 8: Examples, where the dynamic programming-based parabola ﬁtting algorithm did not (fully) improve or worsened

the spine segmentation. The best-ﬁtting parabola is colored red and the DP-ﬁtted points are marked in green. The resulting

spine segmentation is marked in purple, while the excluded part is marked in blue.

The inaccuracy is mainly due to the fact that the

ribs are not completely removed from the segmenta-

tion in cases where they appear to be attached to the

vertebral body in the scan.

Another factor that reduces the measured accuracy

is caused by the inaccuracy of the labeling. Figure 10

Automatic Spine Segmentation in CT Scans

shows an example of this phenomenon. As shown

in the ﬁgure, based on the Hounsﬁeld unit values the

ground truth labeling is not complete.

Table 1: Spine segmentation accuracy results on the

COVID-19 (An et al., 2020) subdataset from the CT-

Spine1K (Deng et al., 2021) dataset.

Patient id IoU DSC

A-0003 0.7218 0.8384

A-0011 0.7637 0.8660

A-0013 0.8226 0.9027

A-0014 0.7632 0.8657

A-0016 0.8100 0.8951

A-0025 0.6938 0.8192

A-0046 0.7090 0.8297

A-0070 0.6575 0.7934

A-0073 0.6583 0.7939

A-0090 0.6971 0.8215

A-0096 0.7403 0.8507

A-0106 0.6499 0.7878

A-0120 0.7462 0.8546

A-0147 0.7269 0.8418

A-0154 0.7193 0.8368

A-0173 0.6665 0.7999

A-0187 0.6863 0.8140

A-0202 0 0.6683 0.8012

A-0215 0.6333 0.7755

A-0237 0.6722 0.8040

Figure 9: Spine segmentation accuracy statistics of the eval-

uation: intersection over union (iou) and Dice similarity co-

efﬁcient (dsc).

We also investigated the effect of the parabola

ﬁtting-based reﬁning step. The comparison of the re-

sults is shown in Table 2. This reﬁnement yielded

only a slight improvement: from 0.7086 to 0.7103

and from 0.8284 to 0.8295 in terms of the average IoU

and DSC, respectively. This shows, that in a general

case, it can only slightly improve the segmentation

by removing the segmentation of high-density tissues

close to the vertebral body. This step, however, can

be crucial if a bigger component is still present in the

segmentation due to the morphological reconstruction

step.

For comparison, Deng et al. (Deng et al., 2021)

provide a U-Net (Ronneberger et al., 2015) based

benchmark solution for the CTSpine1K dataset. Their

approach reached a DSC value of 0.985 on the CT-

Figure 10: An example from the evaluation dataset. Blue

denotes the ground truth labeling, whereas the mask created

by the algorithm is marked in green. Based on simply the

Hounsﬁeld unit values, the labeling is not complete.

Table 2: Mean (avg), standard deviation (std) and minimum

(min) of the intersection over union (IoU) and Dice simi-

larity coefﬁcient values from the evaluation results of the

segmentation masks before (baseline) and after (parabola)

applying the parabola ﬁtting-based reﬁnement.

avg std min

IoU

baseline 0.7086 0.051 0.6344

parabola 0.7103 0.051 0.6332

DSC

baseline 0.8284 0.0344 0.7763

parabola 0.8295 0.0343 0.7754

Spine1K and 0.929 on the VerSe dataset (Sekuboy-

ina et al., 2020). However, it is worth noting that the

VerSe database and the majority of the CTSpine1K

database consists of CT scans with a limited ﬁeld of

view, focusing on the spine. Furthermore, the neural

network based solution provided by Altini et al. (Al-

tini et al., 2021) reached a DSC value of 0.8917 ±

0.0363 on the VerSe dataset.

5 CONCLUSIONS

In this paper we have presented an explicit model-

free segmentation technique for spine segmentation.

The input for the segmentation system is a CT scan

of full thoracic ﬁeld of view. We have used classical

image processing algorithms and dynamic program-

ming that leverage medical expertise. The advantage

of this approach is that it does not require an explicit

model and is fully automatic. This can be a viable

choice if a fast and simple segmentation is desired

and if it is not critical that the segmentation of the

ribs is not perfectly separated from the spine. How-

ever, in cases where individual segmentation of each

vertebra is required, it is necessary to choose a model-

based solution. A possible limitation of this study

is that the presented method was ﬁne-tuned on the

LIDC-IDRI dataset. Further evaluation on additional

BIOIMAGING 2023 - 10th International Conference on Bioimaging

datasets would be preferable, however the availability

of such datasets that contain labeled CT scans of full

thoracic ﬁeld of view is limited.

ACKNOWLEDGEMENTS

The authors acknowledge the National Cancer Insti-

tute and the Foundation for the National Institutes of

Health, and their critical role in the creation of the free

publicly available (Clark et al., 2013b) LIDC/IDRI

Database (Armato III, Samuel G. et al., 2015) used

in this study; and also the Multi-national NIH Con-

sortium for CT AI in COVID-19. This research was

funded by the National Research, Development, and

Innovation Fund of Hungary under Grant TKP2021-

EGA-02.

REFERENCES

Altini, N., De Giosa, G., Fragasso, N., Coscia, C., Sibi-

lano, E., Prencipe, B., Hussain, S. M., Brunetti, A.,

Buongiorno, D., Guerriero, A., et al. (2021). Segmen-

tation and identiﬁcation of vertebrae in ct scans using

cnn, k-means clustering and k-nn. In Informatics, vol-

ume 8, page 40. Multidisciplinary Digital Publishing

Institute.

An, P., Xu, S., Harmon, S. A., Turkbey, E. B., Sanford,

T. H., Amalou, A., Kassin, M., Varble, N., Blain,

M., Anderson, V., Patella, F., Carraﬁello, G., Turkbey,

B. T., and Wood, B. J. (2020). Ct images in covid-19.

Armato, S., MacMahon, H., Engelmann, R., Roberts, R.,

Starkey, A., Caligiuri, P., McLennan, G., Bidaut,

L., Qing, D., McNitt-Gray, M., et al. (2011). The

lung image database consortium (lidc) and image

database resource initiative (idri): A completed ref-

erence database of lung nodules on ct scans. Medical

Physics, 38(2):915–931.

Armato III, Samuel G., McLennan, G., Bidaut, L., et al.

(2015). Data from lidc-idri.

Athertya, J. S. and Kumar, G. S. (2016). Automatic seg-

mentation of vertebral contours from ct images using

fuzzy corners. Computers in biology and medicine,

72:75–89.

Castro-Mateos, I., Pozo, J. M., Perea

nez, M., Lekadir, K.,

Lazary, A., and Frangi, A. F. (2015). Statistical in-

terspace models (sims): application to robust 3d spine

segmentation. IEEE transactions on medical imaging,

34(8):1663–1675.

Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J.,

Koppel, P., Moore, S., Phillips, S., Mafﬁtt, D., Pringle,

M., et al. (2013a). The cancer imaging archive (tcia):

maintaining and operating a public information repos-

itory. Journal of digital imaging, 26(6):1045–1057.

Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J.,

Koppel, P., Moore, S., Phillips, S., Mafﬁtt, D., Pringle,

M., Tarbox, L., and Prior, F. (2013b). The cancer

imaging archive (TCIA): Maintaining and operating

a public information repository. Journal of Digital

Imaging, 26(6):1045–1057.

Deng, Y., Wang, C., Hui, Y., Li, Q., Li, J., Luo, S., Sun, M.,

Quan, Q., Yang, S., Hao, Y., et al. (2021). Ctspine1k:

A large-scale dataset for spinal vertebrae segmenta-

tion in computed tomography. arXiv e-prints, pages

arXiv–2105.

Fischler, M. A. and Bolles, R. C. (1981). Random sample

consensus: a paradigm for model ﬁtting with appli-

cations to image analysis and automated cartography.

Communications of the ACM, 24(6):381–395.

Gonzalez, R. (2009). Digital Image Processing. Pearson

Education.

Harmon, S. A., Sanford, T. H., Xu, S., Turkbey, E. B., Roth,

H., Xu, Z., Yang, D., Myronenko, A., Anderson, V.,

Amalou, A., et al. (2020). Artiﬁcial intelligence for

the detection of covid-19 pneumonia on chest ct us-

ing multinational datasets. Nature communications,

11(1):1–7.

Ibragimov, B., Korez, R., Likar, B., Pernu

s, F., Xing, L.,

and Vrtovec, T. (2017). Segmentation of pathological

structures by landmark-assisted deformable models.

IEEE transactions on medical imaging, 36(7):1457–

1469.

Lessmann, N., Van Ginneken, B., De Jong, P. A., and

sgum, I. (2019). Iterative fully convolutional neu-

ral networks for automatic vertebra segmentation and

identiﬁcation. Medical image analysis, 53:142–155.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Sekuboyina, A., Bayat, A., Husseini, M. E., L

ofﬂer, M.,

Rempﬂer, M., Kuka

cka, J., Tetteh, G., Valentinitsch,

A., Payer, C., Urschler, M., et al. (2020). Verse: a ver-

tebrae labelling and segmentation benchmark. arXiv.

org e-Print archive.

Sekuboyina, A., Kuka

cka, J., Kirschke, J. S., Menze,

B. H., and Valentinitsch, A. (2017). Attention-driven

deep learning for pathological spine segmentation. In

International Workshop on Computational Methods

and Clinical Applications in Musculoskeletal Imag-

ing, pages 108–119. Springer.

Automatic Spine Segmentation in CT Scans