Automatic Valve Segmentation and

Measurement in Cardiac MRI

No Author Given

No Institute Given

Abstract. Cardiovascular Diseases CVDs are the leading cause of mor-

bidity and mortality. According to the World Health Organization, 17.9
million people died from CVDs. This statistic represents 32% of all the
global deaths, of these deaths 85% were due to health attacks and stroke.
Over the past decades, major advances have been made in cardiovascular
research and practice aimed at improving the diagnosis and treatment
of heart disease as well as reducing cardiovascular disease mortality. To
diagnose these pathologies, numerous cardiac imaging and modalities
are available, including Magnetic Resonance Imaging, computed tomog-
raphy, echocardiography. Cardiac MRI is considered the reference for
quantification of ventricular, volumes, mass, and measures, commonly
used to identify quantitative parameters of ventricle’s function. Cardiac
surgery is the heaviest medical procedure with high danger of mortality;
particularly cardiac Valve surgery and coronary artery surgery represent
the majority of indications in cardiac surgery. Undoubtedly, the coronary
life structures change tremendously from one individual to the next, as do
heart valves, and the best way to see the specific life systems is to exam-
ine them during a medical procedure. The current need in cardiovascular
medical surgery is to have an apparatus permitting both representations
of heart valves in three dimensions with the chance of making exact
estimations. This will permit the specialist: better preparation before
surgery, forecast of the instrumentation adapted to each patient, reduc-
tion in the duration of the operative act, and therefore of anesthesia,
with fewer complications. In addition, it allows the preparation of valve
medical surgery with specific or neurotic life systems. The aim objective
of this work is to propose rapid and precise identification and measure-
ment of the cardiac valve using computer vision and image processing
techniques. Finally, the experiments are conducted on Cardiac MRI im-
ages acquired from the Cheikh Zaid International University Hospital in
Rabat Morocco.

Keywords: MRI, Cardiac, Valve segmentation, Valve measurement

1 Introduction

Cardiovascular diseases (CVDs) are the leading cause of morbidity and mortal-
ity. According to the World Health Organization, 17.9 million people died from
CVDs in 2019, representing 32% of all global deaths. Of these deaths, 85% were
2 No Author Given

due to heart attacks and strokes. CVDs are a group of heart and vessel disorders,
including coronary heart disease, cerebrovascular disease, peripheral arterial dis-
ease, and rheumatic heart disease. The valvular disease mortality is close to 2%,
where one or more of the four cardiac valves, particularly the aortic and mitral
valves, are affected [1].
Over the past decades, major advances have been made in cardiovascular
research and practice aimed at improving the diagnosis and treatment of heart
disease and reducing cardiovascular disease mortality. Based on that, many sci-
entists started working on the automatic segmentation of the valves [7, 11, 8].
Segmentation is the process of extracting the desired objective (region of in-
terest), which can be done manually, semi-automatically, or automatically. In the
medical field, the most commonly used process is automatic or semi-automatic
detection within 2D or 3D images from different imaging modalities. The main
objective is to obtain the extracted organ to represent different anatomical struc-
tures [12, 13]. The challenging point with segmentation in the medical field is
the variability in medical images. Moreover, human anatomy itself shows prime
modes of variation.
In this work, we focused on valve heart segmentation, especially the aortic
valve. For that, we first segmented the region of interest that contains the heart
substructure, and then we segmented only the valve region in both states: open
and closed. The aim of our work is to propose a rapid and precise segmentation
of the aortic valve using computer vision and image processing techniques. In
the end, our approach will help clinicians minimize and reduce pre-surgery time
as well for better performance and diagnosis.
The rest of the paper is organized as follows. We introduce the heart anatomy
in Section 2. We describe related work in Section 3. Section 4 presents our
contribution and some additional details for each step of our approach. In Section
5, we present the experiments and results. We conclude this paper with several
interesting directions for future work in Section 6.

2 Heart anatomy

The heart is the center of the blood circulation and the organ that transfers the
oxygen blood to the body. It also transfers the carbon dioxide blood from the
organs to the lungs to exchange it.
The heart is divided into four chambers. The upper chambers are called the
left and right atrium, while the lower chambers are called the left and right ven-
tricle. As we can see in Figure 1, the left ventricular is the largest and strangest
chamber in the heart. A wall of tissue called the septum separates the left and
right atria and the left and right ventricle. Valves separate the atria from the
ventricles. The heart has two sides; the left and right sides of the heart work in
unison. The atria and ventricles contract and relax in turn, producing a rhythmic
heartbeat.
Valves are the keys to the heartbeat. As is shown in Figure 1, the heart has
four valves to ensure that the blood flow has only one direction.
 Valve Segmentation 3

(a) (b)

Fig. 1. (a) Heart anatomy, image source: [14] (b) Four valves, image source: [15]

– Aortic valve: This is between the left ventricle and the aorta.
– Mitral valve: This is between the left atrium and the left ventricle.
– Pulmonary valve: This is between the right ventricle and the pulmonary
artery.
– Tricuspid valve: This is between the right atrium and the right ventricle.
In this paper, we focused only on the aortic valve because it is the cardiac
centerpiece. Normally, its diameter is about 20 mm, and it contains three leaflets.

3 State-of-the-art
Aoyama et al. [2] proposed a fully segmentation approach based on the cascad-
ing of multiple deep neural networks by combining two neural networks, spatial
configuration-NET for anatomical landmarks, and U-NET for segmentation of
aortic valve components. They used ECG gated cardiac 3D–4D CT scans from
138 patients. The dataset was collected using Aquilion ONE (Canon Medical Sys-
tems Corporation, Otawara, Tochigi, Japan) or SOMATOM Forc (SIEMENS,
Munich, Germany). As a result, the proposed algorithm was able to automati-
cally segment different phases and types of aortic valves. Lai et al. [8] developed
a fully automatic method using deep learning on echocardiography images. The
dataset was collected on 58 patients from the Jantung Nagara institute. They
created their own database by selecting only the visible images. After that, they
applied black and white masks. They also cropped the images with the masks,
and after that, they selected 10 images randomly. For the rest, they applied im-
age augmentation (using rotation, translation, flip, Gausian blur, and median
filter). After that, they started training the model. In this step, they used two
4 No Author Given

architectures, the BCDU-NET, and the U-net. In the testing step, they suc-
cessfully proved the potential of using automatic segmentation by using deep
learning. Moreover, they argued that the BCDU-NET architecture is the best
one for image segmentation.
Devrim et al. [4] presented a novel automatic method for aortic segmenta-
tion based on the ECG-gated multislice CT data. Firstly, they applied contrast-
enhanced region detection through histogram analysis, then they applied circle
fitting and region growing based ascending aorta detection. Moreover, they used
supravalvular sinus detection with shape constraints. Finally, they performed a
3D-dimensional region growing over Hessian based aortic valve segmentation.
The proposed method was successively applied.
Pak et al. [11] proposed deep learning algorithms based on U-net with Spatial
Transformer Network (STN) in CT images to fix the need for accurate geometry
parameters of the aortic valve. STN performs an affine transformation to reduce
the input image size. They concentrate on leaflet segmentation in 3D CT im-
ages.The result of this method showed a mean of 0.717 in DSC (Dice Similarity
Coefficient) between the aortic valve root and its leaflets.
Bratt et al. [3] proposed a rapid neural network for aortic valve segmentation
in Phase Contrast (PC) - CMR images for aortic flow decisions. The method is
based on U-Net architecture. The fllow is calculated referring to the automated
segmentation in PC. The DICE output of the method between manual and
automated segmentation shows a mean DICE equal to 0.940 [CI 0.937–0.943].
Haque [7] developed a novel method using the Convolutional Neural Network.
He took the 3D CT images and considered them as a 2D slice to put them as
inputs into CNN for training as semantic segmentation. As a result, the semantic
technique is a tool that accomplishes the true virtuel of AI as a perspective for
human beings.
Liang et al. [9] developed a new 3D reconstruction and modeling of the aor-
tic valve based on machine learning. The image analysis methods are based on
clinical 3D CT images to perform Finite Element (FE) models using the recon-
structed geometries. They used as a database full-phase cardiac multi-slice CT
(MSCT) scans collected from patients at Hartford Hospital (Hartford, CT). This
method got good feedback from the expert. This technique offers great potential
to the computational modeling process, that helps diagnose the disease of valves
and also do pre-operative development.
Mirabella et al. [10] suggested semi-automatic methods to diagnose the cor-
relation of morphological and hemodynamics in BAV (Bicuspid Aortic Valve) for

®
24 study subjects. The aortic valve segmentation was made manually by selecting
control points and interpolating them through OsiriX processing. The focus
on segmentation in the valve plane results in geometry variables that quantify
the difference between BAV and TAV (Tricuspid Aortic Valve) in value terms.
The limitations of the method present time-consuming 2D and 3D segmentation
frames that are manual.
Elattar et al. [5] presented a fully automated ascending aorta segmentation
based on selected geometrical features of the connected components. They also
 Valve Segmentation 5

used normalized cut segmentation to detect the contour for a 20-subject database
who made CTA and extracted aortic root from volumes before TAVI (Tran-

±
scatheter Aortic Valve Implantation) surgery. The result of this method showed
a mean dice coefficient of 0.95 0.03.

4 Automatic Valve Segmentation

Segmentation of the heart valve is one of the crucial task in medical imaging. It
helps in different cardiac measurement and treatments, e.g., aortic valve repair
and replacement. In our approach, we use computer vision as well as image
processing techniques to segment at the first stage the different anatomical heart
substructures. In the second stage, we isolate only the aortic valve from the
segmented substructure for both cases, valve open and closed.

4.1 Heart Segmentation

Heart segmentation aims to extract substructures such as the left ventricular

myocardium, the four-chamber blood cavities, and the great vessels. Our pro-
posed approach is fully automatic without expert supervision and fast in terms
of cardiac substructures segmentation.
As shown in Figure 2, the segmentation process is carried out using basic
image processing techniques.
MRI input: Cardiac MRI slices are used as the input of our approach which
contains all the information available to analyze the patient’s condition.
Image rescaling: This step scales the MRI input slice to the interval [0,1]
in order to normalize the input.
Image binarization: After that, we used the thresholding method that
represents the simplest way to binarize the rescaled images. The threshold was
determined automatically based on the content of MRI slices using the Otsu
method. The proposed approach used the same threshold value for all slices to
separate cardiac structures from the rest of the image.
As shown in Figure 3, the resulting image of binarization is deficient since
some parts of the bone, pulmonary veins, and parts of the rib cage are still
present after the thresholding. After this step, the resulting binary image needs
more processing to distinguish the heart organ from the background and other
structures. However, the use of some morphological techniques in image process-
ing may eliminate undesirable characteristics and also the aortic valve.
Image filtering: To handle the previous problem, we perform median fil-
tering of the rescaled image. According to the median filter, the center pixel
of a n × m neighborhood is replaced by the median value of the corresponding
window. Then, we extract again the binary image of the filtered image.
AND operation: In this step, logical AND operation is performed between
two binary images : 1) Binary rescaled image, and 2) binary filetered image.
The aim of this operation is to keep only pixels in common. By this way, we
6 No Author Given

Fig. 2. Flowchart of the heart substructure segmentation.

guarantee the segmentation of the heart substructure with the presence of the
aortic valve.
Heart substructure segmentation: At this stage, our approach is able to
segment the cardiac substructure in open and closed valve cases. This output
serves as the input of the second stage of our approach to isolate only the aortic
valve.

4.2 Aortic segmentation

Aortic valve opens during systole, the blood is ejected from left ventricle into
aorta. Then, it closes during diastole and prevents reflux from aorta into left
ventricle. In fact, the nature of its functionning makes the valve segmentation
very difficult (Figure 4).
In the second stage, we perform the aortic valve segmentation (Figure 5). For
this purpose, we extract the binary mask of heart substructure (i.e., the resulting
 Valve Segmentation 7

Rescaled image Binarized image Filtered image

Logical AND Substructure mask Substructure segmentation

Fig. 3. Heart substructure segmentation.

(a) (b)

Fig. 4. Valve states. (a) Closed. (b) Open.

output from the first stage) to compute the substructure area. This information
is very helpful to detect the valve state, and hence we select the appropriate
treatment to segment the valve. After many experiments and case studies, we
conclude that the mask area is greater than the Area threshold = 3, 000 when
the valve is open, and smaller than this value otherwise. Based on this study of
valve states relative to the substructure area, we can select the best treatment
for the valve segmentation.
Valve open: Initially, we used mathematical morphology to extract only use-
ful information. First, we perform a morphological erosion with a disc structuring
element which effectively reduces the area of some unwanted regions such as right
and left atrium, and Right Ventricular Outflow Tract Tachycardia (RVOT).In
this step, the valve is disconnected from the other regions. Then, we apply a mor-
phological dilation with a disc structuring element to enlarge the boundaries of
the aortic valve part to be easly identified. Moreover, we extract circularity and
eccentricity parameters of all the small regions in the image. As shown in Figure
8 No Author Given

Fig. 5. Flowchart of the valve segmentation.

6, the valve region represents the highest circularity and the lowest eccentricity
values compared to other regions.
Valve closed: In the other case, when the valve is closed, we first apply the
morphological opening (i.e., the dilation of the erosion) with a disc structuring
element to eliminate the valve region. Then, we apply XOR operation on both
the resulting mask of morphological opening and the one generated from the
first stage of the approach. A logical XOR operation returns 1 if and only if one
of the two pixels is equal to one. By this way, XOR operation permits to isolate
the valve region which need only few processing to be robustly segmented. As
depicted in Figure 7, to segment the valve region, we perform a surface filetering
to keep only the largest region. After that, we perform a morphological erosion
with a disc structuring element to reduce the valve region. Finally, we segment
the valve region using Chan-Vese [6] active contour segmentation. This method
segments objects without clearly defined edges.

5 Experimental Evaluation

In this section, we will outline the used dataset as well as the qualitative and
quantitative results of our segmentation approach.
 Valve Segmentation 9

Substructure Mask Erosion

Dilation Valve segmentation

Fig. 6. Valve open segmentation based on circularity and eccentricity parameters.

Substructure Mask Morphological opening Logical XOR

Surface filtering Erosion Active contour

Fig. 7. Valve closed segmentation.

10 No Author Given

5.1 Clinical case

The proposed approach is also applied to proprietary data of Cheikh Zaid Inter-
national University Hospital in Rabat Morocco, using Siemens Amira equipment.
A concerned patient is a 24-year-old man infected with COVID-19. He weighs
80kg and stands 1.80m tall. The patient was positioned toward the front of the
imaging equipment where the face is in an upward direction (i.e., Head First-
Supine). The patient’s dataset is saved in DICOM 3.0 format. A volume of CINE
VALVE FLASH series has a specific dimension (i.e., weight and height=208 ×
168) and contains 100 slices in grayscale. Whereas, the slice thickness (i.e., the
distance between two adjacent slices) is equal to 8mm.

5.2 Qualitative and quantitative results

Figures 8 and 9 show the segmentation results of our proposed approach. We
can clearly note that both the heart substructure and the aortic valve are well
segmented regardless of whether the valve is open or closed.

Fig. 8. Heart substructure segmentation results. First line: valve open. Second line:
valve closed.

Moreover, we provide statistical analysis to evaluate the performance of the

automatic valve segmentation on our dataset.
Jaccard is defined as the intersection between the ground truth (G) and the
segmented surface (S), divided by their union.

|G∩S |
Jaccard = (1)
|G∪S |
Dice also called the overlap index, is defined as the overlap ratio between
ground truth (G) and the segmented surface (S).
2|G∩S |
Dice = (2)
|G|+|S|
 Valve Segmentation 11

Heart substructure Valve segmentation Ground truth

Fig. 9. Aortic valve segmentation results.

12 No Author Given

Hausdorff distance (HD) is defined as the maximum distance between the

ground truth (G) and the segmented surface (S).

HD = max(D(G, S), D(S, G)) (3)

Table 1 reports the statistical results of our automatic whole segmentation.
Our method achieved a Dice coefficient of 90.09% and 81.86%, a Jaccard index of
0.8197 and 0.7029, as well as 76.01mm and 74.89mm as the Hausdorff distance for
valve open and closed, respectively. These results demonstrate the effectiveness
of our automatic segmentation.

Table 1. Valve segmentation results.

Approach Dice (%) Jaccard HD (mm)

Open 90.09% 0.8197 76.01
Closed 81.86% 0.7029 74.89

6 Conclusion
In this paper, we focused on valve heart segmentation, especially the aortic valve.
For that, we first segmented the region of interest that contains the heart sub-
structure, and then we segmented only the valve region in both states: open
and closed. The approach was conducted on MRI images using computer vision
and image processing techniques that have guaranteed promising results. We
have also surveyed recent work that has been done on the aortic valve segmen-
tation. Our perspective in future work is to stimulate this method and test it
for various aortic valve subjects and extend the application to different anatom-
ical valves of the heart. Finally, our approach will help the clinicians to perform
fully automatic segmentation of the valves for better diagnosis and also make
the pre-surgery process easy and more precise with time-saving.

References
1. Cardiovascular diseases (CVDs). http://www.who.int
2. Aoyama, G., Zhao, L., Zhao, S., Xue, X., Zhong, Y., Yamauchi, H., Tsukihara, H.,
Maeda, E., Ino, K., Tomii, N., et al.: Automatic aortic valve cusps segmentation
from ct images based on the cascading multiple deep neural networks. Journal of
Imaging 8(1), 11 (2022)
3. Bratt, A., Kim, J., Pollie, M., Beecy, A.N., Tehrani, N.H., Codella, N., Perez-
Johnston, R., Palumbo, M.C., Alakbarli, J., Colizza, W., et al.: Machine learning
derived segmentation of phase velocity encoded cardiovascular magnetic resonance
for fully automated aortic flow quantification. Journal of Cardiovascular Magnetic
Resonance 21(1), 1–11 (2019)
 Valve Segmentation 13

4. Devrim, Ü., Harmankaya, İ., Öksüz, İ., Çubuk, R., Çelik, L., Kadipaşaoğlu, K.:
Model-free automatic segmentation of the aortic valve in multislice computed
tomography images. Pamukkale Üniversitesi Mühendislik Bilimleri Dergisi 27(2),
122–128 (2021)
5. Elattar, M.A., Wiegerinck, E., Planken, R., van Assen, H., Baan, J., Marquering,
H., et al.: Automatic segmentation of the aortic root in ct angiography of can-
didate patients for transcatheter aortic valve implantation. Medical & biological
engineering & computing 52(7), 611–618 (2014)
6. Getreuer, P.: Chan-vese segmentation. Image Processing On Line 2, 214–224 (2012)
7. Haque, H.Z.: Aortic valve segmentation using convolutional neural network with
skip mechanism. Communications 7, 1–5
8. Lai, K.W., Shoaib, M.A., Chuah, J.H., Nizar, M.H.A., Anis, S., Ching, S.L.W.:
Aortic valve segmentation using deep learning. In: 2020 IEEE-EMBS Conference
on Biomedical Engineering and Sciences (IECBES). pp. 528–532. IEEE (2021)
9. Liang, L., Kong, F., Martin, C., Pham, T., Wang, Q., Duncan, J., Sun, W.: Machine
learning–based 3-d geometry reconstruction and modeling of aortic valve deforma-
tion using 3-d computed tomography images. International journal for numerical
methods in biomedical engineering 33(5), e2827 (2017)
10. Mirabella, L., Barker, A.J., Saikrishnan, N., Coco, E.R., Mangiameli, D.J., Markl,
M., Yoganathan, A.P.: Mri-based protocol to characterize the relationship between
bicuspid aortic valve morphology and hemodynamics. Annals of biomedical engi-
neering 43(8), 1815–1827 (2015)
11. Pak, D.H., Caballero, A., Sun, W., Duncan, J.S.: Efficient aortic valve multilabel
segmentation using a spatial transformer network. In: 2020 IEEE 17th Interna-
tional Symposium on Biomedical Imaging (ISBI). pp. 1738–1742. IEEE (2020)
12. Shariaty, F., Hosseinlou, S., Rud, V.Y.: Automatic lung segmentation method in
computed tomography scans. In: Journal of Physics: Conference Series. vol. 1236,
p. 012028. IOP Publishing (2019)
13. Zaimi, I., Zrira, N., Benmiloud, I., Marzak, I., Megdiche, K., Ngote, N.: Towards
an improved 3d reconstruction by the use of automatic bone segmentation from ct
scan images. In: 2021 Fifth International Conference On Intelligent Computing in
Data Sciences (ICDS). pp. 1–5. IEEE (2021)
14. The heart poster. https://www.behance.net
15. Heart valves. http://www.commons.wikimedia.org