A Report on Work: Cardiac MRI CBIR for Pathologies Detetion

Tomasz Michno

1 a

and Michal Jelonek

Kielce University of Technology, Kielce, Poland

Center for Cardiovascular Research and Development, American Heart of Poland, Katowice, Poland

Keywords:

CBIR, Medical Image Retrieval, Pathology Detection, MRI, Cardiac Magnetic Resonance Imaging.

Abstract:

The early detection of pathologies in the cardiovascular system is very important. One of the most accurate

imaging examinations of human tissues is magnetic resonance imaging (MRI), which is a very precise yet

non-invasive test. In order to process MRI images to detect pathologies, one of the most promising methods is

Content Based Image Retrieval (CBIR). This paper presents a report on the research on that topic as a result

of the Miniatura 5 Grant. The main contributions of the paper are: a review of the state-of-the-art methods,

a selection of the most promising image features that may be used to identify pathologies, a description of

the proposed system for preparing suggestions for doctors, which takes into consideration also methods for

presenting the results, which are most often omitted in other researches. The next step will be incorporating

full 3D MRI information into the pipeline.

1 INTRODUCTION

Magnetic Resonance Imaging (MRI) is one of the

most important tools in medical diagnostics. It

is a non-invasive, highly reproducible and accurate

method that incorporates the use of a very strong mag-

netic ﬁeld which is able to stimulate protons. The

physicians are able to change the imaging character-

istics by changing its parameters. Moreover, it is pos-

sible to obtain images of vessels and arteries using

MR Angiography (MRA) (Situ et al., 2019). Car-

diac Magnetic Resonance Imaging (CMR) is also able

to produce different types of images, including tis-

sue characterization, thrombus and scar capture which

provide key information for doctors. Patients are not

exposed to ionizing radiation (Peterzan et al., 2016).

There are many heart problems that can be solved

by CMR using different settings and imaging types

(Salerno and Kramer, 2009; Peterzan et al., 2016).

As an example, the presence of coronary artery dis-

ease (CAD) (Chang and Kim, 2016), an acute my-

ocardial infarction (AMI) or chronic myocardial in-

farction (CMI) can be identiﬁed with the CMR, which

may help in choosing a type of treatment (Peterzan

et al., 2016; Tahir et al., 2017). Moreover, CMR is

able to capture images for myocardial stress testing

(Peterzan et al., 2016) and is helpful in the evalua-

tion of Microvascular Obstruction (Perazzolo Marra

https://orcid.org/0000-0001-5437-8728

et al., 2010). Another example of CMR usage in

heart pathology detection may be the detection of Left

Ventricular Thrombus using different types of car-

diac magnetic resonance imaging (Chaosuwannakit

and Makarawate, 2021).

In order to process MRI images to detect patholo-

gies, one of the most promising methods is Content

Based Image Retrieval (CBIR), which is currently

used in many ﬁelds, such as photography, social net-

works, databases, but also in medicine. There are re-

searches focused on the diagnosis of various types of

pathologies, such as diabetic retinopathy (Sivakama-

sundari and Natarajan, 2015), breast cancer detection

(Carvalho et al., 2020) or some cardiovascular patho-

logical changes (Bergamasco et al., 2015b).

This paper presents a report on work on the

Miniatura 5 Grant funded by National Science Cen-

ter Poland, which is dedicated only to preparing ini-

tial research on the method for Cardiac MRI CBIR

for pathologies detection. The main contributions of

the paper are: a review of the state-of-the-art methods

related to the topic, a selection of the most promising

image features that may be used to identify patholo-

gies and a description of the proposed system for

preparing suggestions for doctors. The research also

takes into consideration methods for presenting the

results, which are most often omitted.

The next stage of the research will be incorporat-

ing full 3D MRI image information into the pipeline

Michno, T. and Jelonek, M.

A Report on Work: Cardiac MRI CBIR for Pathologies Detetion.

DOI: 10.5220/0012006800003467

In Proceedings of the 25th International Conference on Enterprise Information Systems (ICEIS 2023) - Volume 1, pages 667-674

ISBN: 978-989-758-648-4; ISSN: 2184-4992

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

667

in order to improve the results.

The paper is organized as follows: in the next sec-

tion there are described methods in medical CBIR.

The section 2 contains a presentation of the proposed

method. Next, the results of initial experiments are

shown. The last section contains a summary and de-

scription of future work directions.

2 LITERATURE OVERVIEW

One of the biggest challenges in medical imaging is

detecting pathologies with high certainty. This could

be very beneﬁcial for choosing appropriate treatment.

Due to the recent advances in computer vision and

image processing, there are more and more possible

usages of computer aided CMR imaging. There is a

great area where a lot of tasks are similar to multime-

dia database retrieval, which in many cases may be

formulated as the problem of Content Based Image

Retrieval (CBIR) (M

unzer et al., 2016).

One of the main challenges in CBIR is proper im-

age segmentation. This is very crucial, especially

in medical image processing, because proper sepa-

ration of, e.g. desired tissues, may highly increase

the precision of a later diagnosis suggestion. For

segmentation different algorithms are used, e.g. for

2D images a fuzzy connectedness image segmenta-

tion with geometric moments (FCISGMs) may be ap-

plied (Atique and Bhagat, 2016). Another idea is to

use a region growing process with OTSU threshold-

ing (Ramos et al., 2014). There are also methods us-

ing U-Net and CNN networks, e.g. (Guo et al., 2020).

Another idea is to ﬁrst localize and then make seg-

mentation of a 3D Cardiac MRI image using a modi-

ﬁed version of the 3D U-Net network named 3D DR-

UNet (Vesal et al., 2020). As another example of

medical image segmentation, a CXR chest image seg-

mentation using a so called CardioNet may be given

(Jafar et al., 2022). Sometimes, for CMR image seg-

mentation, there is also used Fully Convolutional Net-

work (FCN) for semantic segmentation (Bai et al.,

2018). There are also methods for CRM image acqui-

sition plane recognition using CNN networks (Mar-

geta et al., 2017) or some novel image analysis tech-

niques, like Radiomics, which use information about

shape and tissues characteristics as numerical values

(Raisi et al., 2020). Additionally, autoencoder ap-

proaches for segmentation are also researched (Wi-

bowo et al., 2022), as well as transfer learning usage

(Ankenbrand et al., 2021).

Another problem for medical image CBIR is how

to retrieve results after segmentation. There are CBIR

systems that are universally designed for different

types of medical images, like CT, MRI, PET or X-

Ray. One of the problems here is the computation

time, which can be reduced by dividing an image into

a set of blocks (Atique and Bhagat, 2016). There is

also an afﬁne transformation used for image align-

ment along with different metrics (Ayyachamy and

Manivannan, 2013). As different features mean, stan-

dard deviation, entropy, skewness and energy may be

used (Ayyachamy and Manivannan, 2013). For the

detection of Interstitial Lung Diseases in CT images,

not only the information from images may be used but

also from radiology reports with extracting textual in-

formation (Ramos et al., 2014). Another method may

incorporate information from four types of image pro-

jections, like Perfusion and Tagged MRI, LT-SENC,

HT-SENC and Cine Images. As feature vectors, infor-

mation about the dynamic range of the image, Image

Spectrum Histograms and Image Gradient Projectors

may be used (Wael and Fahmy, 2012).

A huge problem in CMR CBIR is using all 3D

depth information stored in the volumetric image. In

order to overcome this problem, spectral clustering

and as a feature vector the 3D Hough Transform De-

scriptor (3DHTD) have been proposed (Bergamasco

et al., 2015b). Another method for querying for 3D

cardiac models is based on the Local Shape Distance

Descriptor (LSDD) and a bipartite graphs (Bergam-

asco et al., 2015a). To summarize, the most often used

descriptors for cardiac MRI are shape, texture, motion

and clinical-based (Delmondes and Nunes, 2022).

There are also cardiac CBIR systems dedicated to

3D echo images (Doppler images) using dimensions

of cardiac ventricles and texture properties, sucha as

kurtosis, skewness, edge gradient, color histogram as

features (Nandagopalan et al., 2012). There are also

methods that decompose medical image features into

Discrete Latent Codes using GANs (Kobayashi et al.,

2021). For 3D Brain MRI images there are also meth-

ods that incorporate dimensionality reduction using

so called Loc-VAE (Nishimaki et al., 2022). An-

other approach for dimensionality reduction may be

the 3D convolutional autoencoders (3D-CAE) (Arai

et al., 2018). Moreover, there are also SURF, PHOG

and VGG16 network based descriptors used to cre-

ate a one-dimensional vector (Rinaldi and Russo,

2020). For brain images, there are also different fea-

tures used, like pixel, local and global features that

are incorporated together to make one feature vec-

tor (Rizvi, 2020). Another example of a deep learn-

ing based method may be the usage of Convolutional

Siamese Neural Networks for distinguishing between

lung cancer and tuberculosis (Zhang et al., 2022). The

summary of all the aforementioned methods is shown

in Tab. 1.

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

668

Table 1: Comparison of chosen CBIR methods.

Research Supported images Algorithms/Methods Metrics

(Atique and Bhagat, 2016)

2D: general

CT, MRI, PET, X-Ray

segmentation: CISGMs,

features/image retrieval: LEBIR,

additional: dividing an image into blocks

no strict information

(Nandagopalan et al., 2012)

2D: cardiac

Doppler image,

(echocardiography)

segmentation: FKM-SQL algorithm

features/image retrieval: dimensions of cardiac ventricles,

texture properties,

kurtosis, skewness, edge gradient, color histogram, etc.

Euclidean distance

(Ayyachamy and Manivannan, 2013)

2D: brain, chest, liver,

limbs

CT, MRI, PET, X-Ray, US

segmentation: no information

registration: aligning images using afﬁne

transformation

features/image retrieval:

sum of squared distance, mutual information

7 metrics tested

Euclidean,

Manhattan,

Mahalanobis,

Canberra,

Bray-curtis,

Squared chord,

chi-squared distances

(Wael and Fahmy, 2012)

2D: cardiac

MRI

segmentation: no information

features/image retrieval:

Dynamic Range, Image Spectrum Histogram,

Image Gradient Projectors

additional: detecting the image type

(Cine images, LT-SENC, HT-SENC, Tagged MRI

and Perfusion MRI)

Euclidean distance

(Ramos et al., 2014)

3D: interstitial lung diseases

segmentation: region growing process + OTSU,

extracting VOIs

features/image retrieval: text distance calculation,

local mean, standard deviation,

skew, kurtosis

additional: supporting CBIR with information

from text reports

Euclidean distance

(Bergamasco et al., 2015b)

3D: cardiac

MRI

segmentation: Seg3D + ImageVis3D software

features/image retrieval: 3DHTD used to extract features

additional: method compares 3D models,

Spectral clustering is used

Euclidean distance

(Bergamasco et al., 2015a)

3D: cardiac

MRI

segmentation: Seg3D + ImageVis3D software

features/image retrieval: LSDD

additional: a bipartite graph is constructed

form two 3D models

Euclidean

and Manhattan

distances

(Kobayashi et al., 2021)

2D: glioma/brain

MRI

segmentation: segmentation decoder

features/image retrieval: autoencoder, decomposing into normal

and abnormal anatomy codes

additional: semantic components decomposition

Euclidean distance

(Nishimaki et al., 2022)

3D: brain

MRI

segmentation and features/image retrieval:

localized variational autoencoder (Loc-VAE)

additional:

dimensionality reduction method

not applicable

(Rinaldi and Russo, 2020)

2D: general

MRI, X-Ray, TAC

segmentation: no information

features/image retrieval: descriptors: SURF, PHOG,

VGG16-based

cosine distance

(Rizvi, 2020)

2D: brain

MRI

segmentation: no strict information

features/image retrieval: pixel, local and global features

no information

(Arai et al., 2018)

3D: brain

MRI

segmentation and features/image retrieval:

3D convolutional autoencoder

additional: dimensionality reduction method

not applicable

(Zhang et al., 2022)

2D: lungs

segmentation and features/image retrieval: CSNN

L2 norm distance

Most of the research on Medical CBIR does not

consider the issue of preparing efﬁcient visualizations

for doctors. However, this is a very crucial part of

the system because it helps in a faster and more pre-

cise understanding of the obtained results. One of the

methods that can be used for a medical diagnosis re-

port is a Venn-style diagram (Huang et al., 2020). Es-

pecially in cardiac reports, there are used Bull’s Eye

A Report on Work: Cardiac MRI CBIR for Pathologies Detetion

669

Plots as well as 2D and 3D maps (Kreiser et al., 2018).

There are also some reports where there is a need to

annotate uncertainties on images, e.g. using different

colors or grayscale areas (Gillmann et al., 2021).

3 PROPOSED CBIR METHOD

FOR DIAGNOSTICS

The main goal of the research is to prepare an ini-

tial method for efﬁcient and precise providing sugges-

tions for doctors with graphical, easy to understand

results presentation. Due to the fact that this task is

very complex, as the ﬁrst stage making a review of

the state-of-the-art methods and preparing an initial

2D method was chosen, which has been presented in

this paper. The next step will be preparing research

on the 3D method.

The proposed system consists of the following

logical modules: preprocessing and segmentation,

feature extraction, the CBIR suggestion module and

visualization module. All of the modules are de-

scribed in more detail in the following subsections.

3.1 Preprocessing and Segmentation

In Computer Vision and Image Processing one of the

most important steps is the proper preparation of the

image for further processing and analysis. One of

the types of segmentation that has been considered

is neural network based. For such an approach, the

U-Net architecture has been chosen, which is one

of the most efﬁcient image segmentation methods in

medical imaging (Siddique et al., 2021),(Tong et al.,

2018). Additionally for higher performance, it can be

supported by other network architectures like R-CNN

(Xu et al., 2018).

One of the problems that may occur during U-Net

usage is the complexity of the input data. In order to

reduce it but also to leave different pixel values in tis-

sues, the Mean shift segmentation has been proposed

as an initial image preprocessing step before using it

as an input for the U-Net. The idea of heart extraction

and example result is shown in Fig. 1.

During the research, we also noticed the problem

of image alignment and registration. Images captured

during acquisition may have different orientations and

scalings. For more precise work with different algo-

rithms from the pipeline, there may be at least some

rotation needed. There are many researches dedicated

to image registration using more classical approaches

(Hill et al., 2001) or Deep Learning (Fu et al., 2020).

During this initial research, that topic was not taken

into consideration and was left for further develop-

ment - all images were rotated manually by a human.

Another idea that has been considered during the

research is the MRI image type detection. It has

been proved that such an additional step can improve

the overall precision of obtained results (Wael and

Fahmy, 2012), thus it was added to the pipeline. A

pre-trained CNN neural network is used in order to

detect image type, which may be e.g. a cine-SSFP or

TSE image. This step is important because it helps to

separate the pipeline for more precise training models

for heart extraction from an MRI image.

3.2 Features Extraction

Extracting features from the image is one of the most

crucial parts of Content Based Image Retrieval sys-

tems. Choosing proper features may greatly improve

the precision of returned results. After analysis of the

aforementioned state-of-the-art methods, described in

Section 2, the most promising ones were chosen.

Statistical and global features are able to represent

information about some features that describes the

image globally like the mean pixel value or the char-

acteristic of the image histogram. Additionally, with

e.g. Gabor ﬁlters, they are able to compare different

textures present in images (Barbu, 2009). Global fea-

tures are commonly and successfully used in CBIR

methods (Varish and Pal, 2015).

Another types of features which are very promis-

ing are heart-based. Some diseases are characterized

by e.g. overgrowth of one of the cardiac ventricles

and such information may be crucial for precise diag-

nosis. Due to that fact, we propose to use two types

of features: proportions of heart width and height and

proportions of heart ventricles. There are some al-

gorithms for automatic heart ventricles segmentation

(Peng et al., 2016) and measurement (Wang et al.,

2019). Extracting these features automatically from

images will be researched in the future and at this

stage has been done manually.

It has been proven that Scale-Invariant Feature

Transform (SIFT), Orientated FAST and Robust

BRIEF (ORB) or Speeded Up Robust Feature (SURF)

descriptors are able to provide precise results for

CBIR (Chhabra et al., 2020). Due to that fact, they are

also promising for their usage in the medical CBIR.

The last type of considered feature descriptors

are neural network-based descriptors. Recently Deep

Learning methods are becoming more and more pop-

ular in many areas, including Content Based Image

Retrieval systems (Staszewski et al., 2021; Rinaldi

and Russo, 2020). For extracting deep features, most

often VGG-16 and VGG-19 are used.

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

670

Figure 1: Heart mask generation and extraction from the MRI image. Images represent actual trained U-Net results. Original

MRI image source: (Campello et al., 2021).

Choosing proper descriptors is not the only prob-

lem in feature extraction. Another crucial thing is the

selection of the area of the image from which these

features should be extracted. During the initial re-

search, it was discovered that higher precision could

be obtained using an extracted heart image. More-

over, because the neighboring area to the extracted

heart may contain some important data for image pro-

cessing and classiﬁcation, we propose to use them

together with the heart features, creating a ”ring”

around the extracted heart.

3.3 Making Recommendations for

Doctors

Preparing recommendations for doctors is a very

complex task. When CBIR is used, one of the

main problems is: storing features together with im-

ages and querying the database in an efﬁcient way

(Deniziak and Michno, 2019). In this research, we

focused only on the querying part, with a simple

database structure based on a vector. A more thor-

ough investigation on how to efﬁciently store infor-

mation about images will be done in future research.

The CBIR suggestions module works as follows:

ﬁrstly, as an input retrieve MRI type and feature vec-

tor. Next, all images of the same MRI type are se-

lected. After that, for each image and feature set

s ∈ (heart, ring):

a) ﬁnd the L2 norm distance between input fea-

ture vector |a| and feature vector of the image from

database |b| using following equation:

dist

(|a|, |b|) =

∑

k=0

− b

, (1)

b) compute the total distance:

val

= 0.75 · dist

heart

(|a|, |b|) (2)

val

= 0.25 · dist

ring

(|a|, |b|) (3)

dist(|a|, |b|) = val

+ val

, (4)

c) store id of the image from database together with

total distance.

Next all stored distances are sorted and top 5 of them

are selected and returned as a main disease sugges-

tion.

The metadata connected with the images contains

the main cardiac disease and a list of other problems

that a patient may have, like thick adipose tissue or a

past myocardial infarction.

What was not covered by most of the researches is

data visualization for doctors. This is done by the last

system’s module, Visualization Module, which as an

input receives data about the ﬁve most similar images

together with their metadata.

The module generates charts that contains recom-

mendation for doctors. Because some diagrams may

show information more clearly, three types of charts

are generated: pie chart, words cloud, bar chart.

The idea of showing more than one diagram is to

choose the most efﬁcient one after consultations with

different doctors, which should be covered in future

research. During the research it occurred that it may

be helpful to not only show suggestions about the

main disease that is present on the MRI image, but

also some additional information that may be con-

nected to the patient. Thus, there are generated two

sets of charts: one containing the main disease and

the other showing all stored additional information.

4 INITIAL EXPERIMENTS

Due to the main idea of the Miniatura 5 grant, which

is gathering knowledge and performing some initial

research that is designated to be highly extended by

future research, more theoretical considerations have

been made and only initial experiments have been per-

formed using the dataset (Campello et al., 2021).

One set of tests was dedicated to checking which

type of segmentation would be most suitable for

initially preprocessing the image in order to make

the heart more signiﬁcant than the background and

other tissues. For initial preprocessing, the follow-

ing segmentation algorithms have been considered:

K-Means, DBSCAN, OTSU, Mean shift, Watershed

(with OTSU thresholding applied before to the im-

age), U-Net. The example segmentation results for

the unsupervised methods are shown in Fig. 2. As

can be seen, each algorithm, except DBSCAN, was

able to segment the heart from other tissues. Another

type of test was made in order to check the U-Net

A Report on Work: Cardiac MRI CBIR for Pathologies Detetion

671

based segmentation, which proved the high efﬁciency

of that method.

Figure 2: The results of different segmentation algorithms.

Original MRI image source: (Campello et al., 2021).

In order to test the idea of MRI image type detec-

tion, a CNN neural network has been implemented

using Python and Keras library. Cine and LV im-

ages from (Campello et al., 2021) has been chosen

for training and testing the network. The average pre-

cision for LV images was 0.87 and for Cine images

was 0.83. This results should be improved by tuning

the parameters and architecture of the network.

Another experiments were made in order to ini-

tially test module for generating suggestions for doc-

tors. As features mean, variance, median, kurtosis,

skewness has been chosen. During the tests, it has

been seen that the variance values are not able to be

used for differentiating classes of images on its own.

The results for kurtosis were much better, but for Con-

genital Arrhythmogenesis class the differences were

still high. This should be investigated more during

future research together with using Gabor ﬁlters.

The next tests were performed to create sugges-

tions using CBIR. Four types of queries have been

made in order to check if the system proposes the cor-

rect disease. For the Hypertrophic Cardiomyopathy,

Tetralogy of Fallot and Dilated Left Ventricle the sys-

tem proposed correct suggestions. The problematic

one was Calcaneal insufﬁciency avulsion, where the

system was not certain about the disease and proposed

all classes with the same probability. This may be the

result of an insufﬁcient number of features - as a fur-

ther research direction it should be investigated which

ones should be the most suitable.

The last set of experiments has been made to test

the correction of chart generation. For that purpose,

the Matplotlib was used to generate bar and pie charts

and WordCloud for generating word clouds. The

module was tested to see if it generates correct charts

for speciﬁed data - for each query it was successful.

5 CONCLUSIONS

In this paper, a report on work on Cardiac MRI CBIR

Retrieval for pathology detection has been presented.

During the research, a review of different methods

that are applied to medical CBIR has been made.

Next, all of them were analyzed for used segmenta-

tion, features and metrics for comparisons. Addition-

ally, an initial method for CBIR has been proposed

that includes: a selection of the most promising image

features that can be extracted from MRI images with

the proposition of using the closest heart’s neighbor-

ing pixels, a description of the comparison and sug-

gestion method together with results visualization and

an overview of the architecture of the proposed sys-

tem.

Moreover, some initial experiments were per-

formed in order to check the main assumptions of the

proposed method. The results showed that for image

segmentation and heart extraction, the U-Net based

method should be sufﬁcient. However, the feature set

should be investigated more thoughtfully in order to

ﬁnd the features that have the best performance when

comparing images connected to different diseases.

When analyzing existing methods and working on

the research, there appeared many further directions

and ideas. Not only improving the performance of

the feature set should be made, but also some efﬁcient

metrics for comparing them during CBIR comparison

phase. More work should also be done in order to

check if an additional feature set for the heart’s neigh-

boring area may improve the precision of suggestions

or not. Another way of improving the results may be

by adding information extracted from e.g. ECG anal-

ysis or patient information, like e.g. age. Moreover,

because at this stage only 2D images were used, an-

other future research direction will be using whole 3D

MRI image data together with the time series.

ACKNOWLEDGEMENTS

For T. Michno this research was funded by ”Miniatura

5” grant from the Polish National Science Centre (ID

533304; No. 2021/05/X/ST6/01794).

REFERENCES

Ankenbrand, M. J., Lohr, D., Schl

otelburg, W., Reiter, T.,

Wech, T., and Schreiber, L. M. (2021). Deep learning-

based cardiac cine segmentation: Transfer learning

application to 7T ultrahigh-ﬁeld MRI. Magnetic Res-

onance in Medicine, 86:2179–2191.

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

672

Arai, H., Chayama, Y., Iyatomi, H., and Oishi, K. (2018).

Signiﬁcant Dimension Reduction of 3D Brain MRI

using 3D Convolutional Autoencoders. volume 2018,

pages 5162–5165.

Atique, M. and Bhagat, A. P. (2016). A Novel Localized

Entropy-based Medical Image Retrieval. IETE Jour-

nal of Research, 62(5):721–732.

Ayyachamy, S. and Manivannan, V. S. (2013). Medical im-

age registration-based retrieval using distance metrics.

International Journal of Imaging Systems and Tech-

nology, 23(4):360–371.

Bai, W., Sinclair, M., Tarroni, G., Oktay, O., Rajchl, M.,

Vaillant, G., Lee, A. M., Aung, N., Lukaschuk, E.,

Sanghvi, M. M., Zemrak, F., Fung, K., Paiva, J. M.,

Carapella, V., Kim, Y. J., Suzuki, H., Kainz, B.,

Matthews, P. M., Petersen, S. E., Piechnik, S. K.,

Neubauer, S., Glocker, B., and Rueckert, D. (2018).

Automated cardiovascular magnetic resonance image

analysis with fully convolutional networks. Journal of

Cardiovascular Magnetic Resonance, 20(1):65.

Barbu, T. (2009). Content-based image retrieval using Ga-

bor ﬁltering. Proceedings - International Workshop

on Database and Expert Systems Applications, DEXA,

pages 236–240.

Bergamasco, L. C., Oliveira, H., B

ıscaro, H., Wechsler, H.,

and Nunes, F. L. (2015a). Using Bipartite Graphs for

3D Cardiac Model Retrieval. In 2015 IEEE 28th In-

ternational Symposium on Computer-Based Medical

Systems, pages 232–237.

Bergamasco, L. C., Oliveira, R. A., Wechsler, H., Da-

juda, C., Delamaro, M., and Nunes, F. L. (2015b).

Content-Based Image Retrieval of 3D Cardiac Mod-

els to Aid the Diagnosis of Congestive Heart Failure

by Using Spectral Clustering. In 2015 IEEE 28th In-

ternational Symposium on Computer-Based Medical

Systems, pages 183–186.

Campello, V. M., Gkontra, P., Izquierdo, C., Martin-Isla,

C., Sojoudi, A., Full, P. M., Maier-Hein, K., Zhang,

Y., He, Z., Ma, J., Parreno, M., Albiol, A., Kong, F.,

Shadden, S. C., Acero, J. C., Sundaresan, V., Saber,

M., Elattar, M., Li, H., Menze, B., Khader, F., Haar-

burger, C., Scannell, C. M., Veta, M., Carscadden, A.,

Punithakumar, K., Liu, X., Tsaftaris, S. A., Huang, X.,

Yang, X., Li, L., Zhuang, X., Vilades, D., Descalzo,

M. L., Guala, A., Mura, L. L., Friedrich, M. G., Garg,

R., Lebel, J., Henriques, F., Karakas, M., Cavus, E.,

Petersen, S. E., Escalera, S., Segui, S., Rodriguez-

Palomares, J. F., and Lekadir, K. (2021). Multi-

Centre, Multi-Vendor and Multi-Disease Cardiac Seg-

mentation: The M&Ms Challenge. IEEE Transactions

on Medical Imaging, 40:3543–3554.

Carvalho, E. D., Filho, A. O., Silva, R. R., Ara

ujo, F. H., Di-

niz, J. O., Silva, A. C., Paiva, A. C., and Gattass, M.

(2020). Breast cancer diagnosis from histopathologi-

cal images using textural features and CBIR. Artiﬁcial

Intelligence in Medicine, 105:101845.

Chang, S.-A. and Kim, R. (2016). The Use of Cardiac Mag-

netic Resonance in Patients with Suspected Coronary

Artery Disease: A Clinical Practice Perspective. Jour-

nal of Cardiovascular Ultrasound, 24:96.

Chaosuwannakit, N. and Makarawate, P. (2021). Left Ven-

tricular Thrombi: Insights from Cardiac Magnetic

Resonance Imaging. Tomography, 7(2):180–188.

Chhabra, P., Garg, N. K., and Kumar, M. (2020).

Content-based image retrieval system using ORB and

SIFT features. Neural Computing and Applications,

32:2725–2733.

Delmondes, P. H. M. and Nunes, F. L. S. (2022). A system-

atic review of multi-slice and multi-frame descriptors

in cardiac MRI exams. Computer Methods and Pro-

grams in Biomedicine, 221:106889.

Deniziak, R. S. and Michno, T. (2019). World wide web

cbir searching using query by approximate shapes. In

Rodr

ıguez, S., Prieto, J., Faria, P., Kłos, S., Fern

andez,

A., Mazuelas, S., Jim

enez-L

opez, M. D., Moreno,

M. N., and Navarro, E. M., editors, Distributed Com-

puting and Artiﬁcial Intelligence, Special Sessions,

15th International Conference, pages 87–95, Cham.

Springer International Publishing.

Fu, Y., Lei, Y., Wang, T., Curran, W. J., Liu, T., and Yang,

X. (2020). Deep learning in medical image regis-

tration: a review. Physics in Medicine & Biology,

65(20):20TR01.

Gillmann, C., Saur, D., Wischgoll, T., and Scheuermann,

G. (2021). Uncertainty-aware Visualization in Medi-

cal Imaging - A Survey. Computer Graphics Forum,

40:665–689.

Guo, F., Ng, M., Goubran, M., Petersen, S. E., Piechnik,

S. K., Neubauer, S., and Wright, G. (2020). Improving

cardiac MRI convolutional neural network segmenta-

tion on small training datasets and dataset shift: A

continuous kernel cut approach. Medical Image Anal-

ysis, 61:101636.

Hill, D. L. G., Batchelor, P. G., Holden, M., and Hawkes,

D. J. (2001). Medical image registration. Physics in

Medicine & Biology, 46(3):R1.

Huang, J.-H., Yang, C.-H. H., Liu, F., Tian, M., Liu,

Y.-C., Wu, T.-W., Lin, I.-H., Wang, K., Morikawa,

H., Chang, H., Tegner, J., and Worring, M. (2020).

DeepOpht: Medical Report Generation for Retinal

Images via Deep Models and Visual Explanation.

Jafar, A., Hameed, M. T., Akram, N., Waqas, U., Kim,

H. S., and Naqvi, R. A. (2022). CardioNet: Auto-

matic Semantic Segmentation to Calculate the Car-

diothoracic Ratio for Cardiomegaly and Other Chest

Diseases. Journal of Personalized Medicine, 12(6).

Kobayashi, K., Hataya, R., Kurose, Y., Miyake, M., Taka-

hashi, M., Nakagawa, A., Harada, T., and Hamamoto,

R. (2021). Decomposing Normal and Abnormal

Features of Medical Images into Discrete Latent

Codes for Content-Based Image Retrieval. CoRR,

abs/2103.12328.

Kreiser, J., Meuschke, M., Mistelbauer, G., Preim, B., and

Ropinski, T. (2018). A Survey of Flattening-Based

Medical Visualization Techniques. Computer Graph-

ics Forum, 37:597–624.

Margeta, J., Criminisi, A., Lozoya, R. C., Lee, D., and Ay-

ache, N. (2017). Fine-tuned convolutional neural nets

for cardiac MRI acquisition plane recognition. Com-

puter Methods in Biomechanics and Biomedical Engi-

neering: Imaging & Visualization, 5(5):339–349.

A Report on Work: Cardiac MRI CBIR for Pathologies Detetion

673

unzer, B., Sch

offmann, K., and B

osz

orm

enyi, L. (2016).

Content-based processing and analysis of endoscopic

images and videos: A survey. Multimedia Tools and

Applications, 77:1323–1362.

Nandagopalan, S., Adiga, B. S., Sudarshan, T. S. B.,

Dhanalakshmi, C., and Manjunath, C. N. (2012). A

content-based image retrieval system for echo images

using SQL-based clustering approach. The Imaging

Science Journal, 60(5):256–271.

Nishimaki, K., Ikuta, K., Onga, Y., Iyatomi, H., and Oishi,

K. (2022). Loc-VAE: Learning Structurally Local-

ized Representation from 3D Brain MR Images for

Content-Based Image Retrieval.

Peng, P., Lekadir, K., Gooya, A., Shao, L., Petersen, S., and

Frangi, A. (2016). A review of heart chamber seg-

mentation for structural and functional analysis using

cardiac magnetic resonance imaging. Magnetic Res-

onance Materials in Physics, Biology and Medicine

2016 29:2, 29:155–195.

Perazzolo Marra, M., Lima, J. A., and Iliceto, S. (2010).

MRI in acute myocardial infarction. European Heart

Journal, 32(3):284–293.

Peterzan, M. A., Rider, O. J., and Anderson, L. J. (2016).

The role of cardiovascular magnetic resonance imag-

ing in heart failure. Cardiac failure review, 2(2):115.

Raisi, Z., Izquierdo, C., Campello, V., Martin-Isla, C.,

Jaggi, A., Harvey, N., Lekadir, K., and Petersen, S.

(2020). Cardiac magnetic resonance radiomics: Basic

principles and clinical perspectives. European Heart

Journal, 21.

Ramos, J., Kockelkorn, T., Ramos, I., Ramos, R., Gin-

neken, B., Viergever, M., Campilho, A., and Grutters,

J. (2014). Content-Based Image Retrieval by Metric

Learning From Radiology Reports: Application to In-

terstitial Lung Diseases. IEEE journal of biomedical

and health informatics, 20.

Rinaldi, A. M. and Russo, C. (2020). A Content Based

Image Retrieval Approach based on Multiple Multi-

media Features Descriptors in E-health Environment.

In 2020 IEEE International Symposium on Medical

Measurements and Applications (MeMeA), pages 1–

Rizvi, D. Q. (2020). Analysis of human brain by mag-

netic resonance imaging using content-based image

retrieval. International journal of health sciences,

14:3–9.

Salerno, M. and Kramer, C. M. (2009). Advances in car-

diovascular MRI for diagnostics: applications in coro-

nary artery disease and cardiomyopathies. Expert

opinion on medical diagnostics, 3(6):673–687.

Siddique, N., Paheding, S., Elkin, C. P., and Devabhaktuni,

V. (2021). U-Net and Its Variants for Medical Image

Segmentation: A Review of Theory and Applications.

IEEE Access, 9:82031–82057.

Situ, Y., Birch, S. C. M., Moreyra, C., and Holloway, C. J.

(2019). Cardiovascular magnetic resonance imaging

for structural heart disease. Cardiovascular Diagnosis

and Therapy, 10(2).

Sivakamasundari, J. and Natarajan, V. (2015). Content

Based retinal Image Retrieval scheme using Harmony

Search Algorithm. 2015 International Conference on

Communication and Signal Processing, ICCSP 2015,

pages 607–611.

Staszewski, P., Jaworski, M., Cao, J., and Rutkowski, L.

(2021). A New Approach to Descriptors Generation

for Image Retrieval by Analyzing Activations of Deep

Neural Network Layers. IEEE Transactions on Neural

Networks and Learning Systems.

Tahir, E., Sinn, M., Bohnen, S., Avanesov, M., S

aring,

D., Stehning, C., Schnackenburg, B., Eulenburg, C.,

Wien, J., Radunski, U. K., Blankenberg, S., Adam,

G., Higgins, C. B., Saeed, M., Muellerleile, K., and

Lund, G. K. (2017). Acute versus Chronic Myocardial

Infarction: Diagnostic Accuracy of Quantitative Na-

tive T1 and T2 Mapping versus Assessment of Edema

on Standard T2-weighted Cardiovascular MR Images

for Differentiation. Radiology, 285(1):83–91. PMID:

28678672.

Tong, Q., Ning, M., Si, W., Liao, X., and Qin, J. (2018). 3D

deeply-supervised U-net based whole heart segmenta-

tion. Lecture Notes in Computer Science (including

subseries Lecture Notes in Artiﬁcial Intelligence and

Lecture Notes in Bioinformatics), 10663 LNCS:224–

232.

Varish, N. and Pal, A. K. (2015). Content based image

retrieval using statistical features of color histogram.

2015 3rd International Conference on Signal Process-

ing, Communication and Networking, ICSCN 2015.

Vesal, S., Maier, A., and Ravikumar, N. (2020). Fully Au-

tomated 3D Cardiac MRI Localisation and Segmenta-

tion Using Deep Neural Networks. Journal of Imag-

ing, 6(7).

Wael, M. and Fahmy, A. S. (2012). Cardiac magnetic res-

onance image classiﬁcation and retrieval based on the

image acquisition technique. In 2012 Cairo Interna-

tional Biomedical Engineering Conference (CIBEC),

pages 126–129.

Wang, Y., Zhang, Y., Xuan, W., Kao, E., Cao, P., Tian, B.,

Ordovas, K., Saloner, D., and Liu, J. (2019). Fully

automatic segmentation of 4D MRI for cardiac func-

tional measurements. Medical Physics, 46:180–189.

Wibowo, A., Triadyaksa, P., Sugiharto, A., Sarwoko,

E. A., Nugroho, F. A., Arai, H., and Kawakubo, M.

(2022). Cardiac Disease Classiﬁcation Using Two-

Dimensional Thickness and Few-Shot Learning Based

on Magnetic Resonance Imaging Image Segmenta-

tion. Journal of Imaging, 8.

Xu, Z., Wu, Z., and Feng, J. (2018). CFUN: combining

faster R-CNN and u-net network for efﬁcient whole

heart segmentation. CoRR, abs/1812.04914.

Zhang, K., Qi, S., Cai, J., Zhao, D., Yu, T., Yue, Y.,

Yao, Y., and Qian, W. (2022). Content-based image

retrieval with a Convolutional Siamese Neural Net-

work: Distinguishing lung cancer and tuberculosis

in CT images. Computers in Biology and Medicine,

140:105096.

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

674